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 "ldlm_internal.h"
  41
  42 extern kmem_cache_t *ldlm_resource_slab;
  43 extern kmem_cache_t *ldlm_lock_slab;
  44 extern struct lustre_lock ldlm_handle_lock;
  45 extern struct list_head ldlm_namespace_list;
  46 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
  47 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
  48
  49 static DECLARE_MUTEX(ldlm_ref_sem);
  50 static int ldlm_refcount = 0;
  51
  52 /* LDLM state */
  53
  54 static struct ldlm_state *ldlm_state;
  55
  56 inline unsigned long round_timeout(unsigned long timeout)
  57 {
  58         return ((timeout / HZ) + 1) * HZ;
  59 }
  60
  61 #ifdef __KERNEL__
  62 /* XXX should this be per-ldlm? */
  63 static struct list_head waiting_locks_list;
  64 static spinlock_t waiting_locks_spinlock;
  65 static struct timer_list waiting_locks_timer;
  66
  67 static struct expired_lock_thread {
  68         wait_queue_head_t         elt_waitq;
  69         int                       elt_state;
  70         struct list_head          elt_expired_locks;
  71         spinlock_t                elt_lock;
  72 } expired_lock_thread;
  73 #endif
  74
  75 #define ELT_STOPPED   0
  76 #define ELT_READY     1
  77 #define ELT_TERMINATE 2
  78
  79 struct ldlm_bl_pool {
  80         spinlock_t              blp_lock;
  81         struct list_head        blp_list;
  82         wait_queue_head_t       blp_waitq;
  83         atomic_t                blp_num_threads;
  84         struct completion       blp_comp;
  85 };
  86
  87 struct ldlm_bl_work_item {
  88         struct list_head        blwi_entry;
  89         struct ldlm_namespace   *blwi_ns;
  90         struct ldlm_lock_desc   blwi_ld;
  91         struct ldlm_lock        *blwi_lock;
  92 };
  93
  94 #ifdef __KERNEL__
  95
  96 static inline int have_expired_locks(void)
  97 {
  98         int need_to_run;
  99
 100         spin_lock_bh(&expired_lock_thread.elt_lock);
 101         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 102         spin_unlock_bh(&expired_lock_thread.elt_lock);
 103
 104         RETURN(need_to_run);
 105 }
 106
 107 static int expired_lock_main(void *arg)
 108 {
 109         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 110         struct l_wait_info lwi = { 0 };
 111         unsigned long flags;
 112
 113         ENTRY;
 114         lock_kernel();
 115         kportal_daemonize("ldlm_elt");
 116
 117         SIGNAL_MASK_LOCK(current, flags);
 118         sigfillset(&current->blocked);
 119         RECALC_SIGPENDING;
 120         SIGNAL_MASK_UNLOCK(current, flags);
 121
 122         unlock_kernel();
 123
 124         expired_lock_thread.elt_state = ELT_READY;
 125         wake_up(&expired_lock_thread.elt_waitq);
 126
 127         while (1) {
 128                 struct list_head *tmp, *n, work_list;
 129                 l_wait_event(expired_lock_thread.elt_waitq,
 130                              have_expired_locks() ||
 131                              expired_lock_thread.elt_state == ELT_TERMINATE,
 132                              &lwi);
 133
 134                 spin_lock_bh(&expired_lock_thread.elt_lock);
 135                 while (!list_empty(expired)) {
 136                         struct ldlm_lock *lock;
 137
 138                         list_add(&work_list, expired);
 139                         list_del_init(expired);
 140
 141                         list_for_each_entry(lock, &work_list, l_pending_chain) {
 142                                 LDLM_DEBUG(lock, "moving to work list");
 143                         }
 144
 145                         spin_unlock_bh(&expired_lock_thread.elt_lock);
 146
 147
 148                         list_for_each_safe(tmp, n, &work_list) {
 149                                  lock = list_entry(tmp, struct ldlm_lock,
 150                                                    l_pending_chain);
 151                                  ptlrpc_fail_export(lock->l_export);
 152                         }
 153
 154
 155                         if (!list_empty(&work_list)) {
 156                                 list_for_each_entry(lock, &work_list, l_pending_chain) {
 157                                         LDLM_ERROR(lock, "still on work list!");
 158                                 }
 159                         }
 160                         LASSERTF (list_empty(&work_list),
 161                                   "some exports not failed properly\n");
 162
 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         spin_lock_bh(&waiting_locks_spinlock);
 182         while (!list_empty(&waiting_locks_list)) {
 183                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 184                                   l_pending_chain);
 185
 186                 if (lock->l_callback_timeout > jiffies)
 187                         break;
 188
 189                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 190                            "%s@%s nid "LPX64" (%s) ",
 191                            lock->l_export->exp_client_uuid.uuid,
 192                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 193                            lock->l_export->exp_connection->c_peer.peer_nid,
 194                            portals_nid2str(lock->l_export->exp_connection->c_peer.peer_ni->pni_number,
 195                                            lock->l_export->exp_connection->c_peer.peer_nid,
 196                                            str));
 197
 198                 spin_lock_bh(&expired_lock_thread.elt_lock);
 199                 list_del(&lock->l_pending_chain);
 200                 list_add(&lock->l_pending_chain,
 201                          &expired_lock_thread.elt_expired_locks);
 202                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 203                 wake_up(&expired_lock_thread.elt_waitq);
 204         }
 205
 206         /*
 207          * Make sure the timer will fire again if we have any locks
 208          * left.
 209          */
 210         if (!list_empty(&waiting_locks_list)) {
 211                 unsigned long timeout_rounded;
 212                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 213                                   l_pending_chain);
 214                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 215                 mod_timer(&waiting_locks_timer, timeout_rounded);
 216         }
 217         spin_unlock_bh(&waiting_locks_spinlock);
 218 }
 219
 220 /*
 221  * Indicate that we're waiting for a client to call us back cancelling a given
 222  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 223  * timer to fire appropriately.  (We round up to the next second, to avoid
 224  * floods of timer firings during periods of high lock contention and traffic).
 225  */
 226 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 227 {
 228         unsigned long timeout_rounded;
 229
 230         spin_lock_bh(&waiting_locks_spinlock);
 231         if (!list_empty(&lock->l_pending_chain)) {
 232                 LDLM_DEBUG(lock, "not re-adding to wait list");
 233                 spin_unlock_bh(&waiting_locks_spinlock);
 234                 return 0;
 235         }
 236         LDLM_DEBUG(lock, "adding to wait list");
 237
 238         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 239
 240         timeout_rounded = round_timeout(lock->l_callback_timeout);
 241
 242         if (timeout_rounded < waiting_locks_timer.expires ||
 243             !timer_pending(&waiting_locks_timer)) {
 244                 mod_timer(&waiting_locks_timer, timeout_rounded);
 245         }
 246         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 247         spin_unlock_bh(&waiting_locks_spinlock);
 248         return 1;
 249 }
 250
 251 /*
 252  * Remove a lock from the pending list, likely because it had its cancellation
 253  * callback arrive without incident.  This adjusts the lock-timeout timer if
 254  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 255  */
 256 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 257 {
 258         struct list_head *list_next;
 259
 260         if (lock->l_export == NULL) {
 261                 /* We don't have a "waiting locks list" on clients. */
 262                 LDLM_DEBUG(lock, "client lock: no-op");
 263                 return 0;
 264         }
 265
 266         spin_lock_bh(&waiting_locks_spinlock);
 267
 268         if (list_empty(&lock->l_pending_chain)) {
 269                 spin_unlock_bh(&waiting_locks_spinlock);
 270                 LDLM_DEBUG(lock, "wasn't waiting");
 271                 return 0;
 272         }
 273
 274         list_next = lock->l_pending_chain.next;
 275         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 276                 /* Removing the head of the list, adjust timer. */
 277                 if (list_next == &waiting_locks_list) {
 278                         /* No more, just cancel. */
 279                         del_timer(&waiting_locks_timer);
 280                 } else {
 281                         struct ldlm_lock *next;
 282                         next = list_entry(list_next, struct ldlm_lock,
 283                                           l_pending_chain);
 284                         mod_timer(&waiting_locks_timer,
 285                                   round_timeout(next->l_callback_timeout));
 286                 }
 287         }
 288
 289         spin_lock_bh(&expired_lock_thread.elt_lock);
 290         list_del_init(&lock->l_pending_chain);
 291         spin_unlock_bh(&expired_lock_thread.elt_lock);
 292
 293         spin_unlock_bh(&waiting_locks_spinlock);
 294         LDLM_DEBUG(lock, "removed");
 295         return 1;
 296 }
 297
 298 #else /* !__KERNEL__ */
 299
 300 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 301 {
 302         RETURN(1);
 303 }
 304
 305 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 306 {
 307         RETURN(0);
 308 }
 309
 310 #endif /* __KERNEL__ */
 311
 312 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, char *ast_type)
 313 {
 314         const struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 315         char str[PTL_NALFMT_SIZE];
 316
 317         CERROR("%s AST failed (%d) for res "LPU64"/"LPU64
 318                ", mode %s: evicting client %s@%s NID "LPX64" (%s)\n",
 319                ast_type, rc,
 320                lock->l_resource->lr_name.name[0],
 321                lock->l_resource->lr_name.name[1],
 322                ldlm_lockname[lock->l_granted_mode],
 323                lock->l_export->exp_client_uuid.uuid,
 324                conn->c_remote_uuid.uuid, conn->c_peer.peer_nid,
 325                portals_nid2str(conn->c_peer.peer_ni->pni_number,
 326                                conn->c_peer.peer_nid, str));
 327         ptlrpc_fail_export(lock->l_export);
 328 }
 329
 330 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 331                              struct ldlm_lock_desc *desc,
 332                              void *data, int flag)
 333 {
 334         struct ldlm_request *body;
 335         struct ptlrpc_request *req;
 336         int rc = 0, size = sizeof(*body);
 337         ENTRY;
 338
 339         if (flag == LDLM_CB_CANCELING) {
 340                 /* Don't need to do anything here. */
 341                 RETURN(0);
 342         }
 343
 344         LASSERT(lock);
 345
 346         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 347         if (lock->l_granted_mode != lock->l_req_mode) {
 348                 /* this blocking AST will be communicated as part of the
 349                  * completion AST instead */
 350                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 351                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");                 RETURN(0);
 352         }
 353
 354         if (lock->l_destroyed) {
 355                 /* What's the point? */
 356                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 357                 RETURN(0);
 358         }
 359
 360 #if 0
 361         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 362                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 363                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 364                 RETURN(-ETIMEDOUT);
 365         }
 366 #endif
 367
 368         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 369                               LDLM_BL_CALLBACK, 1, &size, NULL);
 370         if (req == NULL) {
 371                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 372                 RETURN(-ENOMEM);
 373         }
 374
 375         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 376         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 377                sizeof(body->lock_handle1));
 378         memcpy(&body->lock_desc, desc, sizeof(*desc));
 379         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 380
 381         LDLM_DEBUG(lock, "server preparing blocking AST");
 382         req->rq_replen = lustre_msg_size(0, NULL);
 383
 384         if (lock->l_granted_mode == lock->l_req_mode)
 385                 ldlm_add_waiting_lock(lock);
 386         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 387
 388         req->rq_send_state = LUSTRE_IMP_FULL;
 389         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 390         rc = ptlrpc_queue_wait(req);
 391         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 392                 LASSERT(lock->l_export);
 393                 if (lock->l_export->exp_libclient) {
 394                         CDEBUG(D_HA, "BLOCKING AST to liblustre client (nid "
 395                                LPU64") timeout, simply cancel lock 0x%p\n",
 396                                req->rq_peer.peer_nid, lock);
 397                         ldlm_lock_cancel(lock);
 398                         rc = -ERESTART;
 399                 } else {
 400                         ldlm_del_waiting_lock(lock);
 401                         ldlm_failed_ast(lock, rc, "blocking");
 402                 }
 403         } else if (rc) {
 404                 if (rc == -EINVAL)
 405                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 406                                "from blocking AST for lock %p--normal race\n",
 407                                req->rq_peer.peer_nid,
 408                                req->rq_repmsg->status, lock);
 409                 else if (rc == -ENOTCONN)
 410                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 411                                "from blocking AST for lock %p--this client was "
 412                                "probably rebooted while it held a lock, nothing"
 413                                " serious\n",req->rq_peer.peer_nid,
 414                                req->rq_repmsg->status, lock);
 415                 else
 416                         CDEBUG(D_ERROR, "client (nid "LPU64") returned %d "
 417                                "from blocking AST for lock %p\n",
 418                                req->rq_peer.peer_nid,
 419                                (req->rq_repmsg != NULL)?
 420                                req->rq_repmsg->status : 0,
 421                                lock);
 422                 LDLM_DEBUG(lock, "client sent rc %d rq_status %d from blocking "
 423                            "AST", rc, req->rq_status);
 424                 ldlm_lock_cancel(lock);
 425                 /* Server-side AST functions are called from ldlm_reprocess_all,
 426                  * which needs to be told to please restart its reprocessing. */
 427                 rc = -ERESTART;
 428         }
 429
 430         ptlrpc_req_finished(req);
 431
 432         RETURN(rc);
 433 }
 434
 435 /* XXX copied from ptlrpc/service.c */
 436 static long timeval_sub(struct timeval *large, struct timeval *small)
 437 {
 438         return (large->tv_sec - small->tv_sec) * 1000000 +
 439                 (large->tv_usec - small->tv_usec);
 440 }
 441
 442 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 443 {
 444         struct ldlm_request *body;
 445         struct ptlrpc_request *req;
 446         struct timeval granted_time;
 447         long total_enqueue_wait;
 448         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 449         ENTRY;
 450
 451         LASSERT(lock != NULL);
 452
 453         do_gettimeofday(&granted_time);
 454         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 455
 456         if (total_enqueue_wait / 1000000 > obd_timeout)
 457                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 458
 459         if (lock->l_resource->lr_lvb_len) {
 460                 buffers = 2;
 461                 size[1] = lock->l_resource->lr_lvb_len;
 462         }
 463
 464         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 465                               LDLM_CP_CALLBACK, buffers, size, NULL);
 466         if (req == NULL)
 467                 RETURN(-ENOMEM);
 468
 469         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 470         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 471                sizeof(body->lock_handle1));
 472         body->lock_flags = flags;
 473         ldlm_lock2desc(lock, &body->lock_desc);
 474
 475         if (buffers == 2) {
 476                 void *lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 477                                            lock->l_resource->lr_lvb_len);
 478                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 479                        lock->l_resource->lr_lvb_len);
 480         }
 481
 482         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 483                    total_enqueue_wait);
 484         req->rq_replen = lustre_msg_size(0, NULL);
 485
 486         req->rq_send_state = LUSTRE_IMP_FULL;
 487         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 488
 489         /* We only send real blocking ASTs after the lock is granted */
 490         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 491         if (lock->l_flags & LDLM_FL_AST_SENT) {
 492                 body->lock_flags |= LDLM_FL_AST_SENT;
 493                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 494         }
 495         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 496
 497         rc = ptlrpc_queue_wait(req);
 498         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 499                 ldlm_del_waiting_lock(lock);
 500                 ldlm_failed_ast(lock, rc, "completion");
 501         } else if (rc == -EINVAL) {
 502                 LDLM_DEBUG(lock, "lost the race -- client no longer has this "
 503                            "lock");
 504         } else if (rc) {
 505                 LDLM_ERROR(lock, "client sent rc %d rq_status %d from "
 506                            "completion AST", rc, req->rq_status);
 507                 ldlm_lock_cancel(lock);
 508                 /* Server-side AST functions are called from ldlm_reprocess_all,
 509                  * which needs to be told to please restart its reprocessing. */
 510                 rc = -ERESTART;
 511         }
 512         ptlrpc_req_finished(req);
 513
 514         RETURN(rc);
 515 }
 516
 517 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 518 {
 519         struct ldlm_resource *res = lock->l_resource;
 520         struct ldlm_request *body;
 521         struct ptlrpc_request *req;
 522         int rc = 0, size = sizeof(*body);
 523         ENTRY;
 524
 525         LASSERT(lock != NULL);
 526
 527         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 528                               LDLM_GL_CALLBACK, 1, &size, NULL);
 529         if (req == NULL)
 530                 RETURN(-ENOMEM);
 531
 532         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 533         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 534                sizeof(body->lock_handle1));
 535         ldlm_lock2desc(lock, &body->lock_desc);
 536
 537         size = lock->l_resource->lr_lvb_len;
 538         req->rq_replen = lustre_msg_size(1, &size);
 539
 540         req->rq_send_state = LUSTRE_IMP_FULL;
 541         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 542
 543         rc = ptlrpc_queue_wait(req);
 544         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 545                 ldlm_del_waiting_lock(lock);
 546                 ldlm_failed_ast(lock, rc, "glimpse");
 547         } else if (rc == -EINVAL) {
 548                 LDLM_DEBUG(lock, "lost the race -- client no longer has this "
 549                            "lock");
 550         } else if (rc == -ELDLM_NO_LOCK_DATA) {
 551                 LDLM_DEBUG(lock, "lost a race -- client has a lock, but no "
 552                            "inode");
 553         } else if (rc) {
 554                 LDLM_ERROR(lock, "client sent rc %d rq_status %d from "
 555                            "glimpse AST", rc, req->rq_status);
 556         } else {
 557                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 558                         (res, req->rq_repmsg, 0, 1);
 559         }
 560         ptlrpc_req_finished(req);
 561         RETURN(rc);
 562 }
 563
 564 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 565                         ldlm_completion_callback completion_callback,
 566                         ldlm_blocking_callback blocking_callback,
 567                         ldlm_glimpse_callback glimpse_callback)
 568 {
 569         struct obd_device *obddev = req->rq_export->exp_obd;
 570         struct ldlm_reply *dlm_rep;
 571         struct ldlm_request *dlm_req;
 572         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 573         __u32 flags;
 574         ldlm_error_t err = ELDLM_OK;
 575         struct ldlm_lock *lock = NULL;
 576         void *cookie = NULL;
 577         ENTRY;
 578
 579         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 580
 581         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 582                                       lustre_swab_ldlm_request);
 583         if (dlm_req == NULL) {
 584                 CERROR ("Can't unpack dlm_req\n");
 585                 GOTO(out, rc = -EFAULT);
 586         }
 587
 588         flags = dlm_req->lock_flags;
 589
 590         /* The lock's callback data might be set in the policy function */
 591         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 592                                 dlm_req->lock_desc.l_resource.lr_name,
 593                                 dlm_req->lock_desc.l_resource.lr_type,
 594                                 dlm_req->lock_desc.l_req_mode,
 595                                 blocking_callback, completion_callback,
 596                                 glimpse_callback, NULL, 0);
 597         if (!lock)
 598                 GOTO(out, rc = -ENOMEM);
 599
 600         do_gettimeofday(&lock->l_enqueued_time);
 601         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 602                sizeof(lock->l_remote_handle));
 603         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 604
 605         LASSERT(req->rq_export);
 606         lock->l_export = class_export_get(req->rq_export);
 607         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 608         list_add(&lock->l_export_chain,
 609                  &lock->l_export->exp_ldlm_data.led_held_locks);
 610         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 611
 612         if (flags & LDLM_FL_HAS_INTENT) {
 613                 /* In this case, the reply buffer is allocated deep in
 614                  * local_lock_enqueue by the policy function. */
 615                 cookie = req;
 616         } else {
 617                 int buffers = 1;
 618                 if (lock->l_resource->lr_lvb_len) {
 619                         size[1] = lock->l_resource->lr_lvb_len;
 620                         buffers = 2;
 621                 }
 622
 623                 rc = lustre_pack_reply(req, buffers, size, NULL);
 624                 if (rc)
 625                         GOTO(out, rc);
 626         }
 627
 628         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 629                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 630                        sizeof(ldlm_policy_data_t));
 631         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 632                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 633                        sizeof(lock->l_req_extent));
 634
 635         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 636         if (err)
 637                 GOTO(out, err);
 638
 639         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 640         dlm_rep->lock_flags = flags;
 641
 642         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 643         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 644
 645         /* We never send a blocking AST until the lock is granted, but
 646          * we can tell it right now */
 647         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 648         if (lock->l_flags & LDLM_FL_AST_SENT) {
 649                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 650                 if (lock->l_granted_mode == lock->l_req_mode)
 651                         ldlm_add_waiting_lock(lock);
 652         }
 653         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 654
 655         EXIT;
 656  out:
 657         req->rq_status = err;
 658         if (req->rq_reply_state == NULL) {
 659                 err = lustre_pack_reply(req, 0, NULL, NULL);
 660                 if (rc == 0)
 661                         rc = err;
 662         }
 663
 664         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 665          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 666         if (lock) {
 667                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 668                            "(err=%d, rc=%d)", err, rc);
 669
 670                 if (lock->l_resource->lr_lvb_len > 0) {
 671                         void *lvb = lustre_msg_buf(req->rq_repmsg, 1,
 672                                                   lock->l_resource->lr_lvb_len);
 673                         memcpy(lvb, lock->l_resource->lr_lvb_data,
 674                                lock->l_resource->lr_lvb_len);
 675                 }
 676
 677                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 678                         ldlm_reprocess_all(lock->l_resource);
 679                 LDLM_LOCK_PUT(lock);
 680         }
 681         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 682                           lock, rc);
 683
 684         return rc;
 685 }
 686
 687 int ldlm_handle_convert(struct ptlrpc_request *req)
 688 {
 689         struct ldlm_request *dlm_req;
 690         struct ldlm_reply *dlm_rep;
 691         struct ldlm_lock *lock;
 692         int rc, size = sizeof(*dlm_rep);
 693         ENTRY;
 694
 695         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 696                                       lustre_swab_ldlm_request);
 697         if (dlm_req == NULL) {
 698                 CERROR ("Can't unpack dlm_req\n");
 699                 RETURN (-EFAULT);
 700         }
 701
 702         rc = lustre_pack_reply(req, 1, &size, NULL);
 703         if (rc) {
 704                 CERROR("out of memory\n");
 705                 RETURN(-ENOMEM);
 706         }
 707         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 708         dlm_rep->lock_flags = dlm_req->lock_flags;
 709
 710         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 711         if (!lock) {
 712                 req->rq_status = EINVAL;
 713         } else {
 714                 LDLM_DEBUG(lock, "server-side convert handler START");
 715                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 716                                   &dlm_rep->lock_flags);
 717                 if (ldlm_del_waiting_lock(lock))
 718                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 719                 req->rq_status = 0;
 720         }
 721
 722         if (lock) {
 723                 ldlm_reprocess_all(lock->l_resource);
 724                 LDLM_DEBUG(lock, "server-side convert handler END");
 725                 LDLM_LOCK_PUT(lock);
 726         } else
 727                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 728
 729         RETURN(0);
 730 }
 731
 732 int ldlm_handle_cancel(struct ptlrpc_request *req)
 733 {
 734         struct ldlm_request *dlm_req;
 735         struct ldlm_lock *lock;
 736         struct ldlm_resource *res;
 737         char str[PTL_NALFMT_SIZE];
 738         int rc;
 739         ENTRY;
 740
 741         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 742                                       lustre_swab_ldlm_request);
 743         if (dlm_req == NULL) {
 744                 CERROR("bad request buffer for cancel\n");
 745                 RETURN(-EFAULT);
 746         }
 747
 748         rc = lustre_pack_reply(req, 0, NULL, NULL);
 749         if (rc) {
 750                 CERROR("out of memory\n");
 751                 RETURN(-ENOMEM);
 752         }
 753
 754         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 755         if (!lock) {
 756                 CERROR("received cancel for unknown lock cookie "LPX64
 757                        " from client %s nid "LPX64" (%s)\n",
 758                        dlm_req->lock_handle1.cookie,
 759                        req->rq_export->exp_client_uuid.uuid,
 760                        req->rq_peer.peer_nid,
 761                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 762                                        req->rq_peer.peer_nid, str));
 763                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 764                                   "(cookie "LPU64")",
 765                                   dlm_req->lock_handle1.cookie);
 766                 req->rq_status = ESTALE;
 767         } else {
 768                 LDLM_DEBUG(lock, "server-side cancel handler START");
 769                 res = lock->l_resource;
 770                 if (res && res->lr_namespace->ns_lvbo &&
 771                     res->lr_namespace->ns_lvbo->lvbo_update) {
 772                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 773                                 (res, NULL, 0, 0);
 774                                 //(res, req->rq_reqmsg, 1);
 775                 }
 776
 777                 ldlm_lock_cancel(lock);
 778                 if (ldlm_del_waiting_lock(lock))
 779                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 780                 req->rq_status = rc;
 781         }
 782
 783         if (ptlrpc_reply(req) != 0)
 784                 LBUG();
 785
 786         if (lock) {
 787                 ldlm_reprocess_all(lock->l_resource);
 788                 LDLM_DEBUG(lock, "server-side cancel handler END");
 789                 LDLM_LOCK_PUT(lock);
 790         }
 791
 792         RETURN(0);
 793 }
 794
 795 static void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 796                                     struct ldlm_lock_desc *ld,
 797                                     struct ldlm_lock *lock)
 798 {
 799         int do_ast;
 800         ENTRY;
 801
 802         l_lock(&ns->ns_lock);
 803         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 804
 805         lock->l_flags |= LDLM_FL_CBPENDING;
 806         do_ast = (!lock->l_readers && !lock->l_writers);
 807
 808         if (do_ast) {
 809                 LDLM_DEBUG(lock, "already unused, calling "
 810                            "callback (%p)", lock->l_blocking_ast);
 811                 if (lock->l_blocking_ast != NULL) {
 812                         l_unlock(&ns->ns_lock);
 813                         l_check_no_ns_lock(ns);
 814                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 815                                              LDLM_CB_BLOCKING);
 816                         l_lock(&ns->ns_lock);
 817                 }
 818         } else {
 819                 LDLM_DEBUG(lock, "Lock still has references, will be"
 820                            " cancelled later");
 821         }
 822
 823         LDLM_DEBUG(lock, "client blocking callback handler END");
 824         l_unlock(&ns->ns_lock);
 825         LDLM_LOCK_PUT(lock);
 826         EXIT;
 827 }
 828
 829 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 830                                     struct ldlm_namespace *ns,
 831                                     struct ldlm_request *dlm_req,
 832                                     struct ldlm_lock *lock)
 833 {
 834         LIST_HEAD(ast_list);
 835         ENTRY;
 836
 837         l_lock(&ns->ns_lock);
 838         LDLM_DEBUG(lock, "client completion callback handler START");
 839
 840         /* If we receive the completion AST before the actual enqueue returned,
 841          * then we might need to switch lock modes, resources, or extents. */
 842         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 843                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 844                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 845         }
 846         if (lock->l_resource->lr_type != LDLM_PLAIN)
 847                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 848                        sizeof(lock->l_policy_data));
 849
 850         ldlm_resource_unlink_lock(lock);
 851         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 852                    &lock->l_resource->lr_name,
 853                    sizeof(lock->l_resource->lr_name)) != 0) {
 854                 ldlm_lock_change_resource(ns, lock,
 855                                          dlm_req->lock_desc.l_resource.lr_name);
 856                 LDLM_DEBUG(lock, "completion AST, new resource");
 857         }
 858
 859         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 860                 lock->l_flags |= LDLM_FL_CBPENDING;
 861                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 862         }
 863
 864         if (lock->l_lvb_len) {
 865                 void *lvb;
 866                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 867                                          lock->l_lvb_swabber);
 868                 if (lvb == NULL) {
 869                         LDLM_ERROR(lock, "completion AST did not contain "
 870                                    "expected LVB!");
 871                 } else {
 872                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 873                 }
 874         }
 875
 876         lock->l_resource->lr_tmp = &ast_list;
 877         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 878         lock->l_resource->lr_tmp = NULL;
 879         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 880         l_unlock(&ns->ns_lock);
 881         LDLM_LOCK_PUT(lock);
 882
 883         ldlm_run_ast_work(ns, &ast_list);
 884
 885         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 886                           lock);
 887         EXIT;
 888 }
 889
 890 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 891                                     struct ldlm_namespace *ns,
 892                                     struct ldlm_request *dlm_req,
 893                                     struct ldlm_lock *lock)
 894 {
 895         int rc = -ENOSYS;
 896         ENTRY;
 897
 898         l_lock(&ns->ns_lock);
 899         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 900
 901         if (lock->l_glimpse_ast != NULL) {
 902                 l_unlock(&ns->ns_lock);
 903                 l_check_no_ns_lock(ns);
 904                 rc = lock->l_glimpse_ast(lock, req);
 905                 l_lock(&ns->ns_lock);
 906         }
 907
 908         if (req->rq_repmsg != NULL) {
 909                 ptlrpc_reply(req);
 910         } else {
 911                 req->rq_status = rc;
 912                 ptlrpc_error(req);
 913         }
 914
 915         if (lock->l_granted_mode == LCK_PW &&
 916             !lock->l_readers && !lock->l_writers &&
 917             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 918                 l_unlock(&ns->ns_lock);
 919                 ldlm_handle_bl_callback(ns, NULL, lock);
 920                 EXIT;
 921                 return;
 922         }
 923
 924         l_unlock(&ns->ns_lock);
 925         LDLM_LOCK_PUT(lock);
 926         EXIT;
 927 }
 928
 929 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 930 {
 931         req->rq_status = rc;
 932         if (req->rq_reply_state == NULL) {
 933                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 934                 if (rc)
 935                         return rc;
 936         }
 937         return ptlrpc_reply(req);
 938 }
 939
 940 #ifdef __KERNEL__
 941 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
 942                       struct ldlm_lock *lock)
 943 {
 944         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
 945         struct ldlm_bl_work_item *blwi;
 946         ENTRY;
 947
 948         OBD_ALLOC(blwi, sizeof(*blwi));
 949         if (blwi == NULL)
 950                 RETURN(-ENOMEM);
 951
 952         blwi->blwi_ns = ns;
 953         if (ld != NULL)
 954                 blwi->blwi_ld = *ld;
 955         blwi->blwi_lock = lock;
 956
 957         spin_lock(&blp->blp_lock);
 958         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
 959         wake_up(&blp->blp_waitq);
 960         spin_unlock(&blp->blp_lock);
 961
 962         RETURN(0);
 963 }
 964 #endif
 965
 966 static int ldlm_callback_handler(struct ptlrpc_request *req)
 967 {
 968         struct ldlm_namespace *ns;
 969         struct ldlm_request *dlm_req;
 970         struct ldlm_lock *lock;
 971         char str[PTL_NALFMT_SIZE];
 972         int rc;
 973         ENTRY;
 974
 975         /* Requests arrive in sender's byte order.  The ptlrpc service
 976          * handler has already checked and, if necessary, byte-swapped the
 977          * incoming request message body, but I am responsible for the
 978          * message buffers. */
 979
 980         if (req->rq_export == NULL) {
 981                 struct ldlm_request *dlm_req;
 982
 983                 CDEBUG(D_RPCTRACE, "operation %d from nid "LPX64" (%s) with bad "
 984                        "export cookie "LPX64" (ptl req %d/rep %d); this is "
 985                        "normal if this node rebooted with a lock held\n",
 986                        req->rq_reqmsg->opc, req->rq_peer.peer_nid,
 987                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 988                                        req->rq_peer.peer_nid, str),
 989                        req->rq_reqmsg->handle.cookie,
 990                        req->rq_request_portal, req->rq_reply_portal);
 991
 992                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 993                                              lustre_swab_ldlm_request);
 994                 if (dlm_req != NULL)
 995                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
 996                                dlm_req->lock_handle1.cookie);
 997
 998                 ldlm_callback_reply(req, -ENOTCONN);
 999                 RETURN(0);
1000         }
1001
1002         LASSERT(req->rq_export != NULL);
1003         LASSERT(req->rq_export->exp_obd != NULL);
1004
1005         switch(req->rq_reqmsg->opc) {
1006         case LDLM_BL_CALLBACK:
1007                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1008                 break;
1009         case LDLM_CP_CALLBACK:
1010                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1011                 break;
1012         case LDLM_GL_CALLBACK:
1013                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1014                 break;
1015         case OBD_LOG_CANCEL:
1016                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1017                 rc = llog_origin_handle_cancel(req);
1018                 ldlm_callback_reply(req, rc);
1019                 RETURN(0);
1020         case LLOG_ORIGIN_HANDLE_CREATE:
1021                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1022                 rc = llog_origin_handle_create(req);
1023                 ldlm_callback_reply(req, rc);
1024                 RETURN(0);
1025         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1026                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1027                 rc = llog_origin_handle_next_block(req);
1028                 ldlm_callback_reply(req, rc);
1029                 RETURN(0);
1030         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1031                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1032                 rc = llog_origin_handle_read_header(req);
1033                 ldlm_callback_reply(req, rc);
1034                 RETURN(0);
1035         case LLOG_ORIGIN_HANDLE_CLOSE:
1036                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1037                 rc = llog_origin_handle_close(req);
1038                 ldlm_callback_reply(req, rc);
1039                 RETURN(0);
1040         default:
1041                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1042                 ldlm_callback_reply(req, -EPROTO);
1043                 RETURN(0);
1044         }
1045
1046         ns = req->rq_export->exp_obd->obd_namespace;
1047         LASSERT(ns != NULL);
1048
1049         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1050                                       lustre_swab_ldlm_request);
1051         if (dlm_req == NULL) {
1052                 CERROR ("can't unpack dlm_req\n");
1053                 ldlm_callback_reply (req, -EPROTO);
1054                 RETURN (0);
1055         }
1056
1057         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1058         if (!lock) {
1059                 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
1060                        dlm_req->lock_handle1.cookie);
1061                 ldlm_callback_reply(req, -EINVAL);
1062                 RETURN(0);
1063         }
1064
1065         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1066         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1067
1068         /* We want the ost thread to get this reply so that it can respond
1069          * to ost requests (write cache writeback) that might be triggered
1070          * in the callback.
1071          *
1072          * But we'd also like to be able to indicate in the reply that we're
1073          * cancelling right now, because it's unused, or have an intent result
1074          * in the reply, so we might have to push the responsibility for sending
1075          * the reply down into the AST handlers, alas. */
1076
1077         switch (req->rq_reqmsg->opc) {
1078         case LDLM_BL_CALLBACK:
1079                 CDEBUG(D_INODE, "blocking ast\n");
1080 #ifdef __KERNEL__
1081                 rc = ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock);
1082                 ldlm_callback_reply(req, rc);
1083 #else
1084                 rc = 0;
1085                 ldlm_callback_reply(req, rc);
1086                 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1087 #endif
1088                 break;
1089         case LDLM_CP_CALLBACK:
1090                 CDEBUG(D_INODE, "completion ast\n");
1091                 ldlm_callback_reply(req, 0);
1092                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1093                 break;
1094         case LDLM_GL_CALLBACK:
1095                 CDEBUG(D_INODE, "glimpse ast\n");
1096                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1097                 break;
1098         default:
1099                 LBUG();                         /* checked above */
1100         }
1101
1102         RETURN(0);
1103 }
1104
1105 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1106 {
1107         int rc;
1108         ENTRY;
1109
1110         /* Requests arrive in sender's byte order.  The ptlrpc service
1111          * handler has already checked and, if necessary, byte-swapped the
1112          * incoming request message body, but I am responsible for the
1113          * message buffers. */
1114
1115         if (req->rq_export == NULL) {
1116                 struct ldlm_request *dlm_req;
1117                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
1118                        req->rq_reqmsg->opc, req->rq_request_portal,
1119                        req->rq_reply_portal);
1120                 CERROR("--> export cookie: "LPX64"\n",
1121                        req->rq_reqmsg->handle.cookie);
1122                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1123                                              lustre_swab_ldlm_request);
1124                 if (dlm_req != NULL)
1125                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1126                 RETURN(-ENOTCONN);
1127         }
1128
1129         switch (req->rq_reqmsg->opc) {
1130
1131         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1132         case LDLM_CANCEL:
1133                 CDEBUG(D_INODE, "cancel\n");
1134                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1135                 rc = ldlm_handle_cancel(req);
1136                 if (rc)
1137                         break;
1138                 RETURN(0);
1139
1140         default:
1141                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1142                 RETURN(-EINVAL);
1143         }
1144
1145         RETURN(0);
1146 }
1147
1148 #ifdef __KERNEL__
1149 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1150 {
1151         struct ldlm_bl_work_item *blwi = NULL;
1152
1153         spin_lock(&blp->blp_lock);
1154         if (!list_empty(&blp->blp_list)) {
1155                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1156                                   blwi_entry);
1157                 list_del(&blwi->blwi_entry);
1158         }
1159         spin_unlock(&blp->blp_lock);
1160
1161         return blwi;
1162 }
1163
1164 struct ldlm_bl_thread_data {
1165         int                     bltd_num;
1166         struct ldlm_bl_pool     *bltd_blp;
1167 };
1168
1169 static int ldlm_bl_thread_main(void *arg)
1170 {
1171         struct ldlm_bl_thread_data *bltd = arg;
1172         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1173         unsigned long flags;
1174         ENTRY;
1175
1176         /* XXX boiler-plate */
1177         {
1178                 char name[sizeof(current->comm)];
1179                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1180                          bltd->bltd_num);
1181                 kportal_daemonize(name);
1182         }
1183         SIGNAL_MASK_LOCK(current, flags);
1184         sigfillset(&current->blocked);
1185         RECALC_SIGPENDING;
1186         SIGNAL_MASK_UNLOCK(current, flags);
1187
1188         atomic_inc(&blp->blp_num_threads);
1189         complete(&blp->blp_comp);
1190
1191         while(1) {
1192                 struct l_wait_info lwi = { 0 };
1193                 struct ldlm_bl_work_item *blwi = NULL;
1194
1195                 l_wait_event_exclusive(blp->blp_waitq,
1196                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1197                                        &lwi);
1198
1199                 if (blwi->blwi_ns == NULL)
1200                         break;
1201
1202                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1203                                         blwi->blwi_lock);
1204                 OBD_FREE(blwi, sizeof(*blwi));
1205         }
1206
1207         atomic_dec(&blp->blp_num_threads);
1208         complete(&blp->blp_comp);
1209         RETURN(0);
1210 }
1211
1212 #endif
1213
1214 static int ldlm_setup(void);
1215 static int ldlm_cleanup(int force);
1216
1217 int ldlm_get_ref(void)
1218 {
1219         int rc = 0;
1220         down(&ldlm_ref_sem);
1221         if (++ldlm_refcount == 1) {
1222                 rc = ldlm_setup();
1223                 if (rc)
1224                         ldlm_refcount--;
1225         }
1226         up(&ldlm_ref_sem);
1227
1228         RETURN(rc);
1229 }
1230
1231 void ldlm_put_ref(int force)
1232 {
1233         down(&ldlm_ref_sem);
1234         if (ldlm_refcount == 1) {
1235                 int rc = ldlm_cleanup(force);
1236                 if (rc)
1237                         CERROR("ldlm_cleanup failed: %d\n", rc);
1238                 else
1239                         ldlm_refcount--;
1240         } else {
1241                 ldlm_refcount--;
1242         }
1243         up(&ldlm_ref_sem);
1244
1245         EXIT;
1246 }
1247
1248 static int ldlm_setup(void)
1249 {
1250         struct ldlm_bl_pool *blp;
1251         int rc = 0;
1252 #ifdef __KERNEL__
1253         int i;
1254 #endif
1255         ENTRY;
1256
1257         if (ldlm_state != NULL)
1258                 RETURN(-EALREADY);
1259
1260         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1261         if (ldlm_state == NULL)
1262                 RETURN(-ENOMEM);
1263
1264 #ifdef __KERNEL__
1265         rc = ldlm_proc_setup();
1266         if (rc != 0)
1267                 GOTO(out_free, rc);
1268 #endif
1269
1270         ldlm_state->ldlm_cb_service =
1271                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1272                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1273                                 ldlm_callback_handler, "ldlm_cbd",
1274                                 ldlm_svc_proc_dir);
1275
1276         if (!ldlm_state->ldlm_cb_service) {
1277                 CERROR("failed to start service\n");
1278                 GOTO(out_proc, rc = -ENOMEM);
1279         }
1280
1281         ldlm_state->ldlm_cancel_service =
1282                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1283                                 LDLM_CANCEL_REQUEST_PORTAL,
1284                                 LDLM_CANCEL_REPLY_PORTAL,
1285                                 ldlm_cancel_handler, "ldlm_canceld",
1286                                 ldlm_svc_proc_dir);
1287
1288         if (!ldlm_state->ldlm_cancel_service) {
1289                 CERROR("failed to start service\n");
1290                 GOTO(out_proc, rc = -ENOMEM);
1291         }
1292
1293         OBD_ALLOC(blp, sizeof(*blp));
1294         if (blp == NULL)
1295                 GOTO(out_proc, rc = -ENOMEM);
1296         ldlm_state->ldlm_bl_pool = blp;
1297
1298         atomic_set(&blp->blp_num_threads, 0);
1299         init_waitqueue_head(&blp->blp_waitq);
1300         spin_lock_init(&blp->blp_lock);
1301
1302         INIT_LIST_HEAD(&blp->blp_list);
1303
1304 #ifdef __KERNEL__
1305         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1306                 struct ldlm_bl_thread_data bltd = {
1307                         .bltd_num = i,
1308                         .bltd_blp = blp,
1309                 };
1310                 init_completion(&blp->blp_comp);
1311                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1312                 if (rc < 0) {
1313                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1314                         LBUG();
1315                         GOTO(out_thread, rc);
1316                 }
1317                 wait_for_completion(&blp->blp_comp);
1318         }
1319
1320         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1321                                     LDLM_NUM_THREADS, "ldlm_cn");
1322         if (rc) {
1323                 LBUG();
1324                 GOTO(out_thread, rc);
1325         }
1326
1327         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1328                                     LDLM_NUM_THREADS, "ldlm_cb");
1329         if (rc) {
1330                 LBUG();
1331                 GOTO(out_thread, rc);
1332         }
1333
1334         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1335         spin_lock_init(&expired_lock_thread.elt_lock);
1336         expired_lock_thread.elt_state = ELT_STOPPED;
1337         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1338
1339         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1340         if (rc < 0) {
1341                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1342                 GOTO(out_thread, rc);
1343         }
1344
1345         wait_event(expired_lock_thread.elt_waitq,
1346                    expired_lock_thread.elt_state == ELT_READY);
1347
1348         INIT_LIST_HEAD(&waiting_locks_list);
1349         spin_lock_init(&waiting_locks_spinlock);
1350         waiting_locks_timer.function = waiting_locks_callback;
1351         waiting_locks_timer.data = 0;
1352         init_timer(&waiting_locks_timer);
1353 #endif
1354
1355         RETURN(0);
1356
1357 #ifdef __KERNEL__
1358  out_thread:
1359         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1360         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1361 #endif
1362
1363  out_proc:
1364 #ifdef __KERNEL__
1365         ldlm_proc_cleanup();
1366  out_free:
1367 #endif
1368         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1369         ldlm_state = NULL;
1370         return rc;
1371 }
1372
1373 static int ldlm_cleanup(int force)
1374 {
1375 #ifdef __KERNEL__
1376         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1377 #endif
1378         ENTRY;
1379
1380         if (!list_empty(&ldlm_namespace_list)) {
1381                 CERROR("ldlm still has namespaces; clean these up first.\n");
1382                 ldlm_dump_all_namespaces();
1383                 RETURN(-EBUSY);
1384         }
1385
1386 #ifdef __KERNEL__
1387         while (atomic_read(&blp->blp_num_threads) > 0) {
1388                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1389
1390                 init_completion(&blp->blp_comp);
1391
1392                 spin_lock(&blp->blp_lock);
1393                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1394                 wake_up(&blp->blp_waitq);
1395                 spin_unlock(&blp->blp_lock);
1396
1397                 wait_for_completion(&blp->blp_comp);
1398         }
1399         OBD_FREE(blp, sizeof(*blp));
1400
1401         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1402         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1403         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1404         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1405         ldlm_proc_cleanup();
1406
1407         expired_lock_thread.elt_state = ELT_TERMINATE;
1408         wake_up(&expired_lock_thread.elt_waitq);
1409         wait_event(expired_lock_thread.elt_waitq,
1410                    expired_lock_thread.elt_state == ELT_STOPPED);
1411
1412 #endif
1413
1414         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1415         ldlm_state = NULL;
1416
1417         RETURN(0);
1418 }
1419
1420 int __init ldlm_init(void)
1421 {
1422         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1423                                                sizeof(struct ldlm_resource), 0,
1424                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1425         if (ldlm_resource_slab == NULL)
1426                 return -ENOMEM;
1427
1428         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1429                                            sizeof(struct ldlm_lock), 0,
1430                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1431         if (ldlm_lock_slab == NULL) {
1432                 kmem_cache_destroy(ldlm_resource_slab);
1433                 return -ENOMEM;
1434         }
1435
1436         l_lock_init(&ldlm_handle_lock);
1437
1438         return 0;
1439 }
1440
1441 void __exit ldlm_exit(void)
1442 {
1443         if ( ldlm_refcount )
1444                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1445         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
1446                 CERROR("couldn't free ldlm resource slab\n");
1447         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
1448                 CERROR("couldn't free ldlm lock slab\n");
1449 }
1450
1451 /* ldlm_flock.c */
1452 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1453
1454 /* ldlm_extent.c */
1455 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1456
1457 /* ldlm_lock.c */
1458 EXPORT_SYMBOL(ldlm_get_processing_policy);
1459 EXPORT_SYMBOL(ldlm_lock2desc);
1460 EXPORT_SYMBOL(ldlm_register_intent);
1461 EXPORT_SYMBOL(ldlm_lockname);
1462 EXPORT_SYMBOL(ldlm_typename);
1463 EXPORT_SYMBOL(ldlm_lock2handle);
1464 EXPORT_SYMBOL(__ldlm_handle2lock);
1465 EXPORT_SYMBOL(ldlm_lock_get);
1466 EXPORT_SYMBOL(ldlm_lock_put);
1467 EXPORT_SYMBOL(ldlm_lock_match);
1468 EXPORT_SYMBOL(ldlm_lock_cancel);
1469 EXPORT_SYMBOL(ldlm_lock_addref);
1470 EXPORT_SYMBOL(ldlm_lock_decref);
1471 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1472 EXPORT_SYMBOL(ldlm_lock_change_resource);
1473 EXPORT_SYMBOL(ldlm_lock_set_data);
1474 EXPORT_SYMBOL(ldlm_it2str);
1475 EXPORT_SYMBOL(ldlm_lock_dump);
1476 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1477 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1478 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1479 EXPORT_SYMBOL(ldlm_lock_allow_match);
1480
1481 /* ldlm_request.c */
1482 EXPORT_SYMBOL(ldlm_completion_ast);
1483 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1484 EXPORT_SYMBOL(ldlm_cli_convert);
1485 EXPORT_SYMBOL(ldlm_cli_enqueue);
1486 EXPORT_SYMBOL(ldlm_cli_cancel);
1487 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1488 EXPORT_SYMBOL(ldlm_replay_locks);
1489 EXPORT_SYMBOL(ldlm_resource_foreach);
1490 EXPORT_SYMBOL(ldlm_namespace_foreach);
1491 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1492 EXPORT_SYMBOL(ldlm_change_cbdata);
1493
1494 /* ldlm_lockd.c */
1495 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1496 EXPORT_SYMBOL(ldlm_server_completion_ast);
1497 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1498 EXPORT_SYMBOL(ldlm_handle_enqueue);
1499 EXPORT_SYMBOL(ldlm_handle_cancel);
1500 EXPORT_SYMBOL(ldlm_handle_convert);
1501 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1502 EXPORT_SYMBOL(ldlm_get_ref);
1503 EXPORT_SYMBOL(ldlm_put_ref);
1504
1505 #if 0
1506 /* ldlm_test.c */
1507 EXPORT_SYMBOL(ldlm_test);
1508 EXPORT_SYMBOL(ldlm_regression_start);
1509 EXPORT_SYMBOL(ldlm_regression_stop);
1510 #endif
1511
1512 /* ldlm_resource.c */
1513 EXPORT_SYMBOL(ldlm_namespace_new);
1514 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1515 EXPORT_SYMBOL(ldlm_namespace_free);
1516 EXPORT_SYMBOL(ldlm_namespace_dump);
1517 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1518 EXPORT_SYMBOL(ldlm_resource_get);
1519 EXPORT_SYMBOL(ldlm_resource_putref);
1520
1521 /* l_lock.c */
1522 EXPORT_SYMBOL(l_lock);
1523 EXPORT_SYMBOL(l_unlock);
1524
1525 /* ldlm_lib.c */
1526 EXPORT_SYMBOL(client_obd_setup);
1527 EXPORT_SYMBOL(client_obd_cleanup);
1528 EXPORT_SYMBOL(client_connect_import);
1529 EXPORT_SYMBOL(client_disconnect_export);
1530 EXPORT_SYMBOL(target_abort_recovery);
1531 EXPORT_SYMBOL(target_handle_connect);
1532 EXPORT_SYMBOL(target_destroy_export);
1533 EXPORT_SYMBOL(target_cancel_recovery_timer);
1534 EXPORT_SYMBOL(target_send_reply);
1535 EXPORT_SYMBOL(target_queue_recovery_request);
1536 EXPORT_SYMBOL(target_handle_ping);
1537 EXPORT_SYMBOL(target_handle_disconnect);
1538 EXPORT_SYMBOL(target_queue_final_reply);