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 ;
  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) {
 392 #ifdef __KERNEL__
 393                 ldlm_del_waiting_lock(lock);
 394                 ldlm_failed_ast(lock, rc, "blocking");
 395 #else
 396                 /* XXX
 397                  * Here we treat all clients as liblustre. When BLOCKING AST
 398                  * timeout we don't evicting the client and only cancel
 399                  * the lock.
 400                  * restore to orignial implementation later!!!
 401                  * XXX
 402                  */
 403                 CERROR("BLOCKING AST to client (nid "LPU64") timeout, "
 404                        "simply cancel lock 0x%p\n",
 405                        req->rq_peer.peer_nid, lock);
 406                 ldlm_lock_cancel(lock);
 407                 rc = -ERESTART;
 408 #endif
 409         } else if (rc) {
 410                 if (rc == -EINVAL)
 411                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 412                                "from blocking AST for lock %p--normal race\n",
 413                                req->rq_peer.peer_nid,
 414                                req->rq_repmsg->status, lock);
 415                 else if (rc == -ENOTCONN)
 416                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 417                                "from blocking AST for lock %p--this client was "
 418                                "probably rebooted while it held a lock, nothing"
 419                                " serious\n",req->rq_peer.peer_nid,
 420                                req->rq_repmsg->status, lock);
 421                 else
 422                         CDEBUG(D_ERROR, "client (nid "LPU64") returned %d "
 423                                "from blocking AST for lock %p\n",
 424                                req->rq_peer.peer_nid,
 425                                (req->rq_repmsg != NULL)?
 426                                req->rq_repmsg->status : 0,
 427                                lock);
 428                 LDLM_DEBUG(lock, "client sent rc %d rq_status %d from blocking "
 429                            "AST", rc, req->rq_status);
 430                 ldlm_lock_cancel(lock);
 431                 /* Server-side AST functions are called from ldlm_reprocess_all,
 432                  * which needs to be told to please restart its reprocessing. */
 433                 rc = -ERESTART;
 434         }
 435
 436         ptlrpc_req_finished(req);
 437
 438         RETURN(rc);
 439 }
 440
 441 /* XXX copied from ptlrpc/service.c */
 442 static long timeval_sub(struct timeval *large, struct timeval *small)
 443 {
 444         return (large->tv_sec - small->tv_sec) * 1000000 +
 445                 (large->tv_usec - small->tv_usec);
 446 }
 447
 448 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 449 {
 450         struct ldlm_request *body;
 451         struct ptlrpc_request *req;
 452         struct timeval granted_time;
 453         long total_enqueue_wait;
 454         int rc = 0, size = sizeof(*body);
 455         ENTRY;
 456
 457         if (lock == NULL) {
 458                 LBUG();
 459                 RETURN(-EINVAL);
 460         }
 461
 462         do_gettimeofday(&granted_time);
 463         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 464
 465         if (total_enqueue_wait / 1000000 > obd_timeout)
 466                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 467
 468         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 469                               LDLM_CP_CALLBACK, 1, &size, NULL);
 470         if (!req)
 471                 RETURN(-ENOMEM);
 472
 473         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 474         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 475                sizeof(body->lock_handle1));
 476         body->lock_flags = flags;
 477         ldlm_lock2desc(lock, &body->lock_desc);
 478
 479         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 480                    total_enqueue_wait);
 481         req->rq_replen = lustre_msg_size(0, NULL);
 482
 483         req->rq_send_state = LUSTRE_IMP_FULL;
 484         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 485
 486         /* We only send real blocking ASTs after the lock is granted */
 487         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 488         if (lock->l_flags & LDLM_FL_AST_SENT) {
 489                 body->lock_flags |= LDLM_FL_AST_SENT;
 490                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 491         }
 492         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 493
 494         rc = ptlrpc_queue_wait(req);
 495         if (rc == -ETIMEDOUT || rc == -EINTR) {
 496                 ldlm_del_waiting_lock(lock);
 497                 ldlm_failed_ast(lock, rc, "completion");
 498         } else if (rc) {
 499                 LDLM_ERROR(lock, "client sent rc %d rq_status %d from "
 500                            "completion AST\n", rc, req->rq_status);
 501                 ldlm_lock_cancel(lock);
 502                 /* Server-side AST functions are called from ldlm_reprocess_all,
 503                  * which needs to be told to please restart its reprocessing. */
 504                 rc = -ERESTART;
 505         }
 506         ptlrpc_req_finished(req);
 507
 508         RETURN(rc);
 509 }
 510
 511 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 512                         ldlm_completion_callback completion_callback,
 513                         ldlm_blocking_callback blocking_callback)
 514 {
 515         struct obd_device *obddev = req->rq_export->exp_obd;
 516         struct ldlm_reply *dlm_rep;
 517         struct ldlm_request *dlm_req;
 518         int rc, size = sizeof(*dlm_rep), cookielen = 0;
 519         __u32 flags;
 520         ldlm_error_t err;
 521         struct ldlm_lock *lock = NULL;
 522         void *cookie = NULL;
 523         ENTRY;
 524
 525         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 526
 527         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 528                                       lustre_swab_ldlm_request);
 529         if (dlm_req == NULL) {
 530                 CERROR ("Can't unpack dlm_req\n");
 531                 RETURN (-EFAULT);
 532         }
 533
 534         flags = dlm_req->lock_flags;
 535         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN &&
 536             (flags & LDLM_FL_HAS_INTENT)) {
 537                 /* In this case, the reply buffer is allocated deep in
 538                  * local_lock_enqueue by the policy function. */
 539                 cookie = req;
 540                 cookielen = sizeof(*req);
 541         } else {
 542                 rc = lustre_pack_reply(req, 1, &size, NULL);
 543                 if (rc) {
 544                         CERROR("out of memory\n");
 545                         RETURN(-ENOMEM);
 546                 }
 547                 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN) {
 548                         cookie = &dlm_req->lock_desc.l_policy_data;
 549                         cookielen = sizeof(ldlm_policy_data_t);
 550                 }
 551         }
 552
 553         /* The lock's callback data might be set in the policy function */
 554         lock = ldlm_lock_create(obddev->obd_namespace,
 555                                 &dlm_req->lock_handle2,
 556                                 dlm_req->lock_desc.l_resource.lr_name,
 557                                 dlm_req->lock_desc.l_resource.lr_type,
 558                                 dlm_req->lock_desc.l_req_mode,
 559                                 blocking_callback, completion_callback, NULL);
 560         if (!lock)
 561                 GOTO(out, err = -ENOMEM);
 562
 563         do_gettimeofday(&lock->l_enqueued_time);
 564         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 565                sizeof(lock->l_remote_handle));
 566         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 567
 568         LASSERT(req->rq_export);
 569         lock->l_export = class_export_get(req->rq_export);
 570         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 571         list_add(&lock->l_export_chain,
 572                  &lock->l_export->exp_ldlm_data.led_held_locks);
 573         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 574
 575         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, cookielen,
 576                                 &flags);
 577         if (err)
 578                 GOTO(out, err);
 579
 580         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 581         dlm_rep->lock_flags = flags;
 582
 583         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 584         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN) {
 585                 memcpy(&dlm_rep->lock_policy_data, &lock->l_policy_data,
 586                        cookielen);
 587         }
 588         if (dlm_rep->lock_flags & LDLM_FL_LOCK_CHANGED) {
 589                 memcpy(&dlm_rep->lock_resource_name, &lock->l_resource->lr_name,
 590                        sizeof(dlm_rep->lock_resource_name));
 591                 dlm_rep->lock_mode = lock->l_req_mode;
 592         }
 593
 594         /* We never send a blocking AST until the lock is granted, but
 595          * we can tell it right now */
 596         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 597         if (lock->l_flags & LDLM_FL_AST_SENT) {
 598                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 599                 if (lock->l_granted_mode == lock->l_req_mode)
 600                         ldlm_add_waiting_lock(lock);
 601         }
 602         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 603
 604         EXIT;
 605  out:
 606         req->rq_status = err;
 607
 608         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 609          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 610         if (lock) {
 611                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 612                            "(err=%d)", err);
 613                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 614                         ldlm_reprocess_all(lock->l_resource);
 615                 LDLM_LOCK_PUT(lock);
 616         }
 617         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p)", lock);
 618
 619         return 0;
 620 }
 621
 622 int ldlm_handle_convert(struct ptlrpc_request *req)
 623 {
 624         struct ldlm_request *dlm_req;
 625         struct ldlm_reply *dlm_rep;
 626         struct ldlm_lock *lock;
 627         int rc, size = sizeof(*dlm_rep);
 628         ENTRY;
 629
 630         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 631                                       lustre_swab_ldlm_request);
 632         if (dlm_req == NULL) {
 633                 CERROR ("Can't unpack dlm_req\n");
 634                 RETURN (-EFAULT);
 635         }
 636
 637         rc = lustre_pack_reply(req, 1, &size, NULL);
 638         if (rc) {
 639                 CERROR("out of memory\n");
 640                 RETURN(-ENOMEM);
 641         }
 642         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 643         dlm_rep->lock_flags = dlm_req->lock_flags;
 644
 645         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 646         if (!lock) {
 647                 req->rq_status = EINVAL;
 648         } else {
 649                 LDLM_DEBUG(lock, "server-side convert handler START");
 650                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 651                                   &dlm_rep->lock_flags);
 652                 if (ldlm_del_waiting_lock(lock))
 653                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 654                 req->rq_status = 0;
 655         }
 656
 657         if (lock) {
 658                 ldlm_reprocess_all(lock->l_resource);
 659                 LDLM_DEBUG(lock, "server-side convert handler END");
 660                 LDLM_LOCK_PUT(lock);
 661         } else
 662                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 663
 664         RETURN(0);
 665 }
 666
 667 int ldlm_handle_cancel(struct ptlrpc_request *req)
 668 {
 669         struct ldlm_request *dlm_req;
 670         struct ldlm_lock *lock;
 671         char str[PTL_NALFMT_SIZE];
 672         int rc;
 673         ENTRY;
 674
 675         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 676                                       lustre_swab_ldlm_request);
 677         if (dlm_req == NULL) {
 678                 CERROR("bad request buffer for cancel\n");
 679                 RETURN(-EFAULT);
 680         }
 681
 682         rc = lustre_pack_reply(req, 0, NULL, NULL);
 683         if (rc) {
 684                 CERROR("out of memory\n");
 685                 RETURN(-ENOMEM);
 686         }
 687
 688         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 689         if (!lock) {
 690                 CERROR("received cancel for unknown lock cookie "LPX64
 691                        " from client %s nid "LPX64" (%s)\n",
 692                        dlm_req->lock_handle1.cookie,
 693                        req->rq_export->exp_client_uuid.uuid,
 694                        req->rq_peer.peer_nid,
 695                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 696                                        req->rq_peer.peer_nid, str));
 697                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 698                                   "(cookie "LPU64")",
 699                                   dlm_req->lock_handle1.cookie);
 700                 req->rq_status = ESTALE;
 701         } else {
 702                 LDLM_DEBUG(lock, "server-side cancel handler START");
 703                 ldlm_lock_cancel(lock);
 704                 if (ldlm_del_waiting_lock(lock))
 705                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 706                 req->rq_status = 0;
 707         }
 708
 709         if (ptlrpc_reply(req) != 0)
 710                 LBUG();
 711
 712         if (lock) {
 713                 ldlm_reprocess_all(lock->l_resource);
 714                 LDLM_DEBUG(lock, "server-side cancel handler END");
 715                 LDLM_LOCK_PUT(lock);
 716         }
 717
 718         RETURN(0);
 719 }
 720
 721 static void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 722                                     struct ldlm_lock_desc *ld,
 723                                     struct ldlm_lock *lock)
 724 {
 725         int do_ast;
 726         ENTRY;
 727
 728         l_lock(&ns->ns_lock);
 729         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 730
 731         lock->l_flags |= LDLM_FL_CBPENDING;
 732         do_ast = (!lock->l_readers && !lock->l_writers);
 733
 734         if (do_ast) {
 735                 LDLM_DEBUG(lock, "already unused, calling "
 736                            "callback (%p)", lock->l_blocking_ast);
 737                 if (lock->l_blocking_ast != NULL) {
 738                         l_unlock(&ns->ns_lock);
 739                         l_check_no_ns_lock(ns);
 740                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 741                                              LDLM_CB_BLOCKING);
 742                         l_lock(&ns->ns_lock);
 743                 }
 744         } else {
 745                 LDLM_DEBUG(lock, "Lock still has references, will be"
 746                            " cancelled later");
 747         }
 748
 749         LDLM_DEBUG(lock, "client blocking callback handler END");
 750         l_unlock(&ns->ns_lock);
 751         LDLM_LOCK_PUT(lock);
 752         EXIT;
 753 }
 754
 755 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 756                                     struct ldlm_namespace *ns,
 757                                     struct ldlm_request *dlm_req,
 758                                     struct ldlm_lock *lock)
 759 {
 760         LIST_HEAD(ast_list);
 761         ENTRY;
 762
 763         l_lock(&ns->ns_lock);
 764         LDLM_DEBUG(lock, "client completion callback handler START");
 765
 766         /* If we receive the completion AST before the actual enqueue returned,
 767          * then we might need to switch lock modes, resources, or extents. */
 768         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 769                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 770                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 771         }
 772         if (lock->l_resource->lr_type != LDLM_PLAIN)
 773                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 774                        sizeof(lock->l_policy_data));
 775
 776         ldlm_resource_unlink_lock(lock);
 777         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 778                    &lock->l_resource->lr_name,
 779                    sizeof(lock->l_resource->lr_name)) != 0) {
 780                 ldlm_lock_change_resource(ns, lock,
 781                                          dlm_req->lock_desc.l_resource.lr_name);
 782                 LDLM_DEBUG(lock, "completion AST, new resource");
 783         }
 784
 785         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 786                 lock->l_flags |= LDLM_FL_CBPENDING;
 787                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 788         }
 789
 790         lock->l_resource->lr_tmp = &ast_list;
 791         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 792         lock->l_resource->lr_tmp = NULL;
 793         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 794         l_unlock(&ns->ns_lock);
 795         LDLM_LOCK_PUT(lock);
 796
 797         ldlm_run_ast_work(ns, &ast_list);
 798
 799         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 800                           lock);
 801         EXIT;
 802 }
 803
 804 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 805 {
 806         req->rq_status = rc;
 807         rc = lustre_pack_reply(req, 0, NULL, NULL);
 808         if (rc)
 809                 return rc;
 810         return ptlrpc_reply(req);
 811 }
 812
 813 #ifdef __KERNEL__
 814 static int ldlm_bl_to_thread(struct ldlm_state *ldlm, struct ldlm_namespace *ns,
 815                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 816 {
 817         struct ldlm_bl_pool *blp = ldlm->ldlm_bl_pool;
 818         struct ldlm_bl_work_item *blwi;
 819         ENTRY;
 820
 821         OBD_ALLOC(blwi, sizeof(*blwi));
 822         if (blwi == NULL)
 823                 RETURN(-ENOMEM);
 824
 825         blwi->blwi_ns = ns;
 826         blwi->blwi_ld = *ld;
 827         blwi->blwi_lock = lock;
 828
 829         spin_lock(&blp->blp_lock);
 830         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
 831         wake_up(&blp->blp_waitq);
 832         spin_unlock(&blp->blp_lock);
 833
 834         RETURN(0);
 835 }
 836 #endif
 837
 838 static int ldlm_callback_handler(struct ptlrpc_request *req)
 839 {
 840         struct ldlm_namespace *ns;
 841         struct ldlm_request *dlm_req;
 842         struct ldlm_lock *lock;
 843         char str[PTL_NALFMT_SIZE];
 844         int rc;
 845         ENTRY;
 846
 847         /* Requests arrive in sender's byte order.  The ptlrpc service
 848          * handler has already checked and, if necessary, byte-swapped the
 849          * incoming request message body, but I am responsible for the
 850          * message buffers. */
 851
 852         if (req->rq_export == NULL) {
 853                 struct ldlm_request *dlm_req;
 854
 855                 CDEBUG(D_RPCTRACE, "operation %d from nid "LPX64" (%s) with bad "
 856                        "export cookie "LPX64" (ptl req %d/rep %d); this is "
 857                        "normal if this node rebooted with a lock held\n",
 858                        req->rq_reqmsg->opc, req->rq_peer.peer_nid,
 859                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 860                                        req->rq_peer.peer_nid, str),
 861                        req->rq_reqmsg->handle.cookie,
 862                        req->rq_request_portal, req->rq_reply_portal);
 863
 864                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 865                                              lustre_swab_ldlm_request);
 866                 if (dlm_req != NULL)
 867                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
 868                                dlm_req->lock_handle1.cookie);
 869
 870                 ldlm_callback_reply(req, -ENOTCONN);
 871                 RETURN(0);
 872         }
 873
 874         if (req->rq_reqmsg->opc == LDLM_BL_CALLBACK) {
 875                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 876         } else if (req->rq_reqmsg->opc == LDLM_CP_CALLBACK) {
 877                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
 878         } else if (req->rq_reqmsg->opc == OBD_LOG_CANCEL) {
 879                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
 880         } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CREATE) {
 881                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
 882         } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_NEXT_BLOCK) {
 883                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
 884         } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_READ_HEADER) {
 885                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
 886         } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CLOSE) {
 887                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
 888         } else {
 889                 ldlm_callback_reply(req, -EPROTO);
 890                 RETURN(0);
 891         }
 892
 893         LASSERT(req->rq_export != NULL);
 894         LASSERT(req->rq_export->exp_obd != NULL);
 895
 896         /* FIXME - how to send reply */
 897         if (req->rq_reqmsg->opc == OBD_LOG_CANCEL) {
 898                 int rc = llog_origin_handle_cancel(req);
 899                 ldlm_callback_reply(req, rc);
 900                 RETURN(0);
 901         }
 902         if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CREATE) {
 903                 int rc = llog_origin_handle_create(req);
 904                 req->rq_status = rc;
 905                 ptlrpc_reply(req);
 906                 RETURN(0);
 907         }
 908         if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_NEXT_BLOCK) {
 909                 int rc = llog_origin_handle_next_block(req);
 910                 req->rq_status = rc;
 911                 ptlrpc_reply(req);
 912                 RETURN(0);
 913         }
 914         if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_READ_HEADER) {
 915                 int rc = llog_origin_handle_read_header(req);
 916                 req->rq_status = rc;
 917                 ptlrpc_reply(req);
 918                 RETURN(0);
 919         }
 920         if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CLOSE) {
 921                 int rc = llog_origin_handle_close(req);
 922                 ldlm_callback_reply(req, rc);
 923                 RETURN(0);
 924         }
 925
 926         ns = req->rq_export->exp_obd->obd_namespace;
 927         LASSERT(ns != NULL);
 928
 929         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 930                                       lustre_swab_ldlm_request);
 931         if (dlm_req == NULL) {
 932                 CERROR ("can't unpack dlm_req\n");
 933                 ldlm_callback_reply (req, -EPROTO);
 934                 RETURN (0);
 935         }
 936
 937         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
 938         if (!lock) {
 939                 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
 940                        dlm_req->lock_handle1.cookie);
 941                 ldlm_callback_reply(req, -EINVAL);
 942                 RETURN(0);
 943         }
 944
 945         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
 946         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
 947
 948         /* We want the ost thread to get this reply so that it can respond
 949          * to ost requests (write cache writeback) that might be triggered
 950          * in the callback.
 951          *
 952          * But we'd also like to be able to indicate in the reply that we're
 953          * cancelling right now, because it's unused, or have an intent result
 954          * in the reply, so we might have to push the responsibility for sending
 955          * the reply down into the AST handlers, alas. */
 956         if (req->rq_reqmsg->opc != LDLM_BL_CALLBACK)
 957                 ldlm_callback_reply(req, 0);
 958
 959         switch (req->rq_reqmsg->opc) {
 960         case LDLM_BL_CALLBACK:
 961                 CDEBUG(D_INODE, "blocking ast\n");
 962 #ifdef __KERNEL__
 963                 rc = ldlm_bl_to_thread(ldlm, ns, &dlm_req->lock_desc, lock);
 964                 ldlm_callback_reply(req, rc);
 965 #else
 966                 rc = 0;
 967                 ldlm_callback_reply(req, rc);
 968                 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
 969 #endif
 970                 break;
 971         case LDLM_CP_CALLBACK:
 972                 CDEBUG(D_INODE, "completion ast\n");
 973                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
 974                 break;
 975         default:
 976                 LBUG();                         /* checked above */
 977         }
 978
 979         RETURN(0);
 980 }
 981
 982 static int ldlm_cancel_handler(struct ptlrpc_request *req)
 983 {
 984         int rc;
 985         ENTRY;
 986
 987         /* Requests arrive in sender's byte order.  The ptlrpc service
 988          * handler has already checked and, if necessary, byte-swapped the
 989          * incoming request message body, but I am responsible for the
 990          * message buffers. */
 991
 992         if (req->rq_export == NULL) {
 993                 struct ldlm_request *dlm_req;
 994                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
 995                        req->rq_reqmsg->opc, req->rq_request_portal,
 996                        req->rq_reply_portal);
 997                 CERROR("--> export cookie: "LPX64"\n",
 998                        req->rq_reqmsg->handle.cookie);
 999                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1000                                              lustre_swab_ldlm_request);
1001                 if (dlm_req != NULL)
1002                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1003                 RETURN(-ENOTCONN);
1004         }
1005
1006         switch (req->rq_reqmsg->opc) {
1007
1008         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1009         case LDLM_CANCEL:
1010                 CDEBUG(D_INODE, "cancel\n");
1011                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1012                 rc = ldlm_handle_cancel(req);
1013                 if (rc)
1014                         break;
1015                 RETURN(0);
1016
1017         default:
1018                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1019                 RETURN(-EINVAL);
1020         }
1021
1022         RETURN(0);
1023 }
1024
1025 #ifdef __KERNEL__
1026 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1027 {
1028         struct ldlm_bl_work_item *blwi = NULL;
1029
1030         spin_lock(&blp->blp_lock);
1031         if (!list_empty(&blp->blp_list)) {
1032                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1033                                   blwi_entry);
1034                 list_del(&blwi->blwi_entry);
1035         }
1036         spin_unlock(&blp->blp_lock);
1037
1038         return blwi;
1039 }
1040
1041 struct ldlm_bl_thread_data {
1042         int                     bltd_num;
1043         struct ldlm_bl_pool     *bltd_blp;
1044 };
1045
1046 static int ldlm_bl_thread_main(void *arg)
1047 {
1048         struct ldlm_bl_thread_data *bltd = arg;
1049         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1050         unsigned long flags;
1051         ENTRY;
1052
1053         /* XXX boiler-plate */
1054         {
1055                 char name[sizeof(current->comm)];
1056                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1057                          bltd->bltd_num);
1058                 kportal_daemonize(name);
1059         }
1060         SIGNAL_MASK_LOCK(current, flags);
1061         sigfillset(&current->blocked);
1062         RECALC_SIGPENDING;
1063         SIGNAL_MASK_UNLOCK(current, flags);
1064
1065         atomic_inc(&blp->blp_num_threads);
1066         complete(&blp->blp_comp);
1067
1068         while(1) {
1069                 struct l_wait_info lwi = { 0 };
1070                 struct ldlm_bl_work_item *blwi = NULL;
1071
1072                 l_wait_event_exclusive(blp->blp_waitq,
1073                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1074                                        &lwi);
1075
1076                 if (blwi->blwi_ns == NULL)
1077                         break;
1078
1079                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1080                                         blwi->blwi_lock);
1081                 OBD_FREE(blwi, sizeof(*blwi));
1082         }
1083
1084         atomic_dec(&blp->blp_num_threads);
1085         complete(&blp->blp_comp);
1086         RETURN(0);
1087 }
1088
1089 #endif
1090
1091 static int ldlm_setup(void);
1092 static int ldlm_cleanup(int force);
1093
1094 int ldlm_get_ref(void)
1095 {
1096         int rc = 0;
1097         down(&ldlm_ref_sem);
1098         if (++ldlm_refcount == 1) {
1099                 rc = ldlm_setup();
1100                 if (rc)
1101                         ldlm_refcount--;
1102         }
1103         up(&ldlm_ref_sem);
1104
1105         RETURN(rc);
1106 }
1107
1108 void ldlm_put_ref(int force)
1109 {
1110         down(&ldlm_ref_sem);
1111         if (ldlm_refcount == 1) {
1112                 int rc = ldlm_cleanup(force);
1113                 if (rc)
1114                         CERROR("ldlm_cleanup failed: %d\n", rc);
1115                 else
1116                         ldlm_refcount--;
1117         } else {
1118                 ldlm_refcount--;
1119         }
1120         up(&ldlm_ref_sem);
1121
1122         EXIT;
1123 }
1124
1125 static int ldlm_setup(void)
1126 {
1127         struct ldlm_bl_pool *blp;
1128         int rc = 0;
1129 #ifdef __KERNEL__
1130         int i;
1131 #endif
1132         ENTRY;
1133
1134         if (ldlm != NULL)
1135                 RETURN(-EALREADY);
1136
1137         OBD_ALLOC(ldlm, sizeof(*ldlm));
1138         if (ldlm == NULL)
1139                 RETURN(-ENOMEM);
1140
1141 #ifdef __KERNEL__
1142         rc = ldlm_proc_setup();
1143         if (rc != 0)
1144                 GOTO(out_free, rc);
1145 #endif
1146
1147         ldlm->ldlm_cb_service =
1148                 ptlrpc_init_svc(LDLM_NEVENTS, LDLM_NBUFS, LDLM_BUFSIZE,
1149                                 LDLM_MAXREQSIZE, LDLM_CB_REQUEST_PORTAL,
1150                                 LDLM_CB_REPLY_PORTAL,
1151                                 ldlm_callback_handler, "ldlm_cbd",
1152                                 ldlm_svc_proc_dir);
1153
1154         if (!ldlm->ldlm_cb_service) {
1155                 CERROR("failed to start service\n");
1156                 GOTO(out_proc, rc = -ENOMEM);
1157         }
1158
1159         ldlm->ldlm_cancel_service =
1160                 ptlrpc_init_svc(LDLM_NEVENTS, LDLM_NBUFS, LDLM_BUFSIZE,
1161                                 LDLM_MAXREQSIZE, LDLM_CANCEL_REQUEST_PORTAL,
1162                                 LDLM_CANCEL_REPLY_PORTAL,
1163                                 ldlm_cancel_handler, "ldlm_canceld",
1164                                 ldlm_svc_proc_dir);
1165
1166         if (!ldlm->ldlm_cancel_service) {
1167                 CERROR("failed to start service\n");
1168                 GOTO(out_proc, rc = -ENOMEM);
1169         }
1170
1171         OBD_ALLOC(blp, sizeof(*blp));
1172         if (blp == NULL)
1173                 GOTO(out_proc, rc = -ENOMEM);
1174         ldlm->ldlm_bl_pool = blp;
1175
1176         atomic_set(&blp->blp_num_threads, 0);
1177         init_waitqueue_head(&blp->blp_waitq);
1178         spin_lock_init(&blp->blp_lock);
1179
1180         INIT_LIST_HEAD(&blp->blp_list);
1181
1182 #ifdef __KERNEL__
1183         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1184                 struct ldlm_bl_thread_data bltd = {
1185                         .bltd_num = i,
1186                         .bltd_blp = blp,
1187                 };
1188                 init_completion(&blp->blp_comp);
1189                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1190                 if (rc < 0) {
1191                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1192                         LBUG();
1193                         GOTO(out_thread, rc);
1194                 }
1195                 wait_for_completion(&blp->blp_comp);
1196         }
1197
1198         rc = ptlrpc_start_n_threads(NULL, ldlm->ldlm_cancel_service,
1199                                     LDLM_NUM_THREADS, "ldlm_cn");
1200         if (rc) {
1201                 LBUG();
1202                 GOTO(out_thread, rc);
1203         }
1204
1205         rc = ptlrpc_start_n_threads(NULL, ldlm->ldlm_cb_service,
1206                                     LDLM_NUM_THREADS, "ldlm_cb");
1207         if (rc) {
1208                 LBUG();
1209                 GOTO(out_thread, rc);
1210         }
1211
1212         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1213         spin_lock_init(&expired_lock_thread.elt_lock);
1214         expired_lock_thread.elt_state = ELT_STOPPED;
1215         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1216
1217         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1218         if (rc < 0) {
1219                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1220                 GOTO(out_thread, rc);
1221         }
1222
1223         wait_event(expired_lock_thread.elt_waitq,
1224                    expired_lock_thread.elt_state == ELT_READY);
1225
1226         INIT_LIST_HEAD(&waiting_locks_list);
1227         spin_lock_init(&waiting_locks_spinlock);
1228         waiting_locks_timer.function = waiting_locks_callback;
1229         waiting_locks_timer.data = 0;
1230         init_timer(&waiting_locks_timer);
1231 #endif
1232
1233         RETURN(0);
1234
1235 #ifdef __KERNEL__
1236  out_thread:
1237         ptlrpc_unregister_service(ldlm->ldlm_cancel_service);
1238         ptlrpc_unregister_service(ldlm->ldlm_cb_service);
1239 #endif
1240
1241  out_proc:
1242 #ifdef __KERNEL__
1243         ldlm_proc_cleanup();
1244  out_free:
1245 #endif
1246         OBD_FREE(ldlm, sizeof(*ldlm));
1247         ldlm = NULL;
1248         return rc;
1249 }
1250
1251 static int ldlm_cleanup(int force)
1252 {
1253 #ifdef __KERNEL__
1254         struct ldlm_bl_pool *blp = ldlm->ldlm_bl_pool;
1255 #endif
1256         ENTRY;
1257
1258         if (!list_empty(&ldlm_namespace_list)) {
1259                 CERROR("ldlm still has namespaces; clean these up first.\n");
1260                 ldlm_dump_all_namespaces();
1261                 RETURN(-EBUSY);
1262         }
1263
1264 #ifdef __KERNEL__
1265         while (atomic_read(&blp->blp_num_threads) > 0) {
1266                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1267
1268                 init_completion(&blp->blp_comp);
1269
1270                 spin_lock(&blp->blp_lock);
1271                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1272                 wake_up(&blp->blp_waitq);
1273                 spin_unlock(&blp->blp_lock);
1274
1275                 wait_for_completion(&blp->blp_comp);
1276         }
1277         OBD_FREE(blp, sizeof(*blp));
1278
1279         ptlrpc_stop_all_threads(ldlm->ldlm_cb_service);
1280         ptlrpc_unregister_service(ldlm->ldlm_cb_service);
1281         ptlrpc_stop_all_threads(ldlm->ldlm_cancel_service);
1282         ptlrpc_unregister_service(ldlm->ldlm_cancel_service);
1283         ldlm_proc_cleanup();
1284
1285         expired_lock_thread.elt_state = ELT_TERMINATE;
1286         wake_up(&expired_lock_thread.elt_waitq);
1287         wait_event(expired_lock_thread.elt_waitq,
1288                    expired_lock_thread.elt_state == ELT_STOPPED);
1289
1290 #endif
1291
1292         OBD_FREE(ldlm, sizeof(*ldlm));
1293         ldlm = NULL;
1294
1295         RETURN(0);
1296 }
1297
1298 int __init ldlm_init(void)
1299 {
1300         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1301                                                sizeof(struct ldlm_resource), 0,
1302                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1303         if (ldlm_resource_slab == NULL)
1304                 return -ENOMEM;
1305
1306         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1307                                            sizeof(struct ldlm_lock), 0,
1308                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1309         if (ldlm_lock_slab == NULL) {
1310                 kmem_cache_destroy(ldlm_resource_slab);
1311                 return -ENOMEM;
1312         }
1313
1314         l_lock_init(&ldlm_handle_lock);
1315
1316         return 0;
1317 }
1318
1319 void __exit ldlm_exit(void)
1320 {
1321         if ( ldlm_refcount )
1322                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1323         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
1324                 CERROR("couldn't free ldlm resource slab\n");
1325         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
1326                 CERROR("couldn't free ldlm lock slab\n");
1327 }
1328
1329 /* ldlm_flock.c */
1330 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1331
1332 /* ldlm_lock.c */
1333 EXPORT_SYMBOL(ldlm_lock2desc);
1334 EXPORT_SYMBOL(ldlm_register_intent);
1335 EXPORT_SYMBOL(ldlm_unregister_intent);
1336 EXPORT_SYMBOL(ldlm_lockname);
1337 EXPORT_SYMBOL(ldlm_typename);
1338 EXPORT_SYMBOL(ldlm_lock2handle);
1339 EXPORT_SYMBOL(__ldlm_handle2lock);
1340 EXPORT_SYMBOL(ldlm_lock_put);
1341 EXPORT_SYMBOL(ldlm_lock_match);
1342 EXPORT_SYMBOL(ldlm_lock_cancel);
1343 EXPORT_SYMBOL(ldlm_lock_addref);
1344 EXPORT_SYMBOL(ldlm_lock_decref);
1345 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1346 EXPORT_SYMBOL(ldlm_lock_change_resource);
1347 EXPORT_SYMBOL(ldlm_lock_set_data);
1348 EXPORT_SYMBOL(ldlm_it2str);
1349 EXPORT_SYMBOL(ldlm_lock_dump);
1350 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1351 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1352 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1353
1354 /* ldlm_request.c */
1355 EXPORT_SYMBOL(ldlm_completion_ast);
1356 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1357 EXPORT_SYMBOL(ldlm_cli_convert);
1358 EXPORT_SYMBOL(ldlm_cli_enqueue);
1359 EXPORT_SYMBOL(ldlm_cli_cancel);
1360 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1361 EXPORT_SYMBOL(ldlm_replay_locks);
1362 EXPORT_SYMBOL(ldlm_resource_foreach);
1363 EXPORT_SYMBOL(ldlm_namespace_foreach);
1364 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1365 EXPORT_SYMBOL(ldlm_change_cbdata);
1366
1367 /* ldlm_lockd.c */
1368 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1369 EXPORT_SYMBOL(ldlm_server_completion_ast);
1370 EXPORT_SYMBOL(ldlm_handle_enqueue);
1371 EXPORT_SYMBOL(ldlm_handle_cancel);
1372 EXPORT_SYMBOL(ldlm_handle_convert);
1373 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1374 EXPORT_SYMBOL(ldlm_get_ref);
1375 EXPORT_SYMBOL(ldlm_put_ref);
1376
1377 #if 0
1378 /* ldlm_test.c */
1379 EXPORT_SYMBOL(ldlm_test);
1380 EXPORT_SYMBOL(ldlm_regression_start);
1381 EXPORT_SYMBOL(ldlm_regression_stop);
1382 #endif
1383
1384 /* ldlm_resource.c */
1385 EXPORT_SYMBOL(ldlm_namespace_new);
1386 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1387 EXPORT_SYMBOL(ldlm_namespace_free);
1388
1389 /* l_lock.c */
1390 EXPORT_SYMBOL(l_lock);
1391 EXPORT_SYMBOL(l_unlock);
1392
1393 /* ldlm_lib.c */
1394 EXPORT_SYMBOL(client_obd_setup);
1395 EXPORT_SYMBOL(client_obd_cleanup);
1396 EXPORT_SYMBOL(client_connect_import);
1397 EXPORT_SYMBOL(client_disconnect_export);
1398 EXPORT_SYMBOL(target_abort_recovery);
1399 EXPORT_SYMBOL(target_handle_connect);
1400 EXPORT_SYMBOL(target_destroy_export);
1401 EXPORT_SYMBOL(target_cancel_recovery_timer);
1402 EXPORT_SYMBOL(target_send_reply);
1403 EXPORT_SYMBOL(target_queue_recovery_request);
1404 EXPORT_SYMBOL(target_handle_ping);
1405 EXPORT_SYMBOL(target_handle_disconnect);
1406 EXPORT_SYMBOL(target_queue_final_reply);
1407 EXPORT_SYMBOL(ldlm_put_lock_into_req);