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 <portals/list.h>
  41 #include "ldlm_internal.h"
  42
  43 extern kmem_cache_t *ldlm_resource_slab;
  44 extern kmem_cache_t *ldlm_lock_slab;
  45 extern struct lustre_lock ldlm_handle_lock;
  46 extern struct list_head ldlm_namespace_list;
  47 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
  48 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
  49
  50 static DECLARE_MUTEX(ldlm_ref_sem);
  51 static int ldlm_refcount = 0;
  52
  53 /* LDLM state */
  54
  55 static struct ldlm_state *ldlm_state;
  56
  57 inline unsigned long round_timeout(unsigned long timeout)
  58 {
  59         return ((timeout / HZ) + 1) * HZ;
  60 }
  61
  62 #ifdef __KERNEL__
  63 /* XXX should this be per-ldlm? */
  64 static struct list_head waiting_locks_list;
  65 static spinlock_t waiting_locks_spinlock;
  66 static struct timer_list waiting_locks_timer;
  67
  68 static struct expired_lock_thread {
  69         wait_queue_head_t         elt_waitq;
  70         int                       elt_state;
  71         struct list_head          elt_expired_locks;
  72         spinlock_t                elt_lock;
  73 } expired_lock_thread;
  74 #endif
  75
  76 #define ELT_STOPPED   0
  77 #define ELT_READY     1
  78 #define ELT_TERMINATE 2
  79
  80 struct ldlm_bl_pool {
  81         spinlock_t              blp_lock;
  82         struct list_head        blp_list;
  83         wait_queue_head_t       blp_waitq;
  84         atomic_t                blp_num_threads;
  85         struct completion       blp_comp;
  86 };
  87
  88 struct ldlm_bl_work_item {
  89         struct list_head        blwi_entry;
  90         struct ldlm_namespace   *blwi_ns;
  91         struct ldlm_lock_desc   blwi_ld;
  92         struct ldlm_lock        *blwi_lock;
  93 };
  94
  95 #ifdef __KERNEL__
  96
  97 static inline int have_expired_locks(void)
  98 {
  99         int need_to_run;
 100
 101         spin_lock_bh(&expired_lock_thread.elt_lock);
 102         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 103         spin_unlock_bh(&expired_lock_thread.elt_lock);
 104
 105         RETURN(need_to_run);
 106 }
 107
 108 static int expired_lock_main(void *arg)
 109 {
 110         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 111         struct l_wait_info lwi = { 0 };
 112         unsigned long flags;
 113
 114         ENTRY;
 115         lock_kernel();
 116         kportal_daemonize("ldlm_elt");
 117
 118         SIGNAL_MASK_LOCK(current, flags);
 119         sigfillset(&current->blocked);
 120         RECALC_SIGPENDING;
 121         SIGNAL_MASK_UNLOCK(current, flags);
 122
 123         unlock_kernel();
 124
 125         expired_lock_thread.elt_state = ELT_READY;
 126         wake_up(&expired_lock_thread.elt_waitq);
 127
 128         while (1) {
 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 obd_export *export;
 137                         struct ldlm_lock *lock;
 138
 139                         lock = list_entry(expired->next, struct ldlm_lock,
 140                                           l_pending_chain);
 141                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
 142                             (void *)lock >= LP_POISON) {
 143                                 CERROR("free lock on elt list %p\n", lock);
 144                                 LBUG();
 145                         }
 146                         list_del_init(&lock->l_pending_chain);
 147                         if ((void *)lock->l_export < LP_POISON + PAGE_SIZE &&
 148                             (void *)lock->l_export >= LP_POISON + PAGE_SIZE) {
 149                                 CERROR("lock with free export on elt list %p\n",
 150                                        export);
 151                                 lock->l_export = NULL;
 152                                 LDLM_ERROR(lock, "free export\n");
 153                                 continue;
 154                         }
 155                         export = class_export_get(lock->l_export);
 156                         spin_unlock_bh(&expired_lock_thread.elt_lock);
 157
 158                         ptlrpc_fail_export(export);
 159                         class_export_put(export);
 160                         spin_lock_bh(&expired_lock_thread.elt_lock);
 161                 }
 162                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 163
 164                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 165                         break;
 166         }
 167
 168         expired_lock_thread.elt_state = ELT_STOPPED;
 169         wake_up(&expired_lock_thread.elt_waitq);
 170         RETURN(0);
 171 }
 172
 173 static void waiting_locks_callback(unsigned long unused)
 174 {
 175         struct ldlm_lock *lock;
 176         char str[PTL_NALFMT_SIZE];
 177
 178         spin_lock_bh(&waiting_locks_spinlock);
 179         while (!list_empty(&waiting_locks_list)) {
 180                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 181                                   l_pending_chain);
 182
 183                 if ((lock->l_callback_timeout > jiffies) ||
 184                     (lock->l_req_mode == LCK_GROUP))
 185                         break;
 186
 187                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 188                            "%s@%s nid "LPX64" (%s) ",
 189                            lock->l_export->exp_client_uuid.uuid,
 190                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 191                            lock->l_export->exp_connection->c_peer.peer_nid,
 192                            portals_nid2str(lock->l_export->exp_connection->c_peer.peer_ni->pni_number,
 193                                            lock->l_export->exp_connection->c_peer.peer_nid,
 194                                            str));
 195
 196                 spin_lock_bh(&expired_lock_thread.elt_lock);
 197                 list_del(&lock->l_pending_chain);
 198                 list_add(&lock->l_pending_chain,
 199                          &expired_lock_thread.elt_expired_locks);
 200                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 201                 wake_up(&expired_lock_thread.elt_waitq);
 202         }
 203
 204         /*
 205          * Make sure the timer will fire again if we have any locks
 206          * left.
 207          */
 208         if (!list_empty(&waiting_locks_list)) {
 209                 unsigned long timeout_rounded;
 210                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 211                                   l_pending_chain);
 212                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 213                 mod_timer(&waiting_locks_timer, timeout_rounded);
 214         }
 215         spin_unlock_bh(&waiting_locks_spinlock);
 216 }
 217
 218 /*
 219  * Indicate that we're waiting for a client to call us back cancelling a given
 220  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 221  * timer to fire appropriately.  (We round up to the next second, to avoid
 222  * floods of timer firings during periods of high lock contention and traffic).
 223  */
 224 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 225 {
 226         unsigned long timeout_rounded;
 227
 228         spin_lock_bh(&waiting_locks_spinlock);
 229         if (!list_empty(&lock->l_pending_chain)) {
 230                 LDLM_DEBUG(lock, "not re-adding to wait list");
 231                 spin_unlock_bh(&waiting_locks_spinlock);
 232                 return 0;
 233         }
 234         LDLM_DEBUG(lock, "adding to wait list");
 235
 236         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 237
 238         timeout_rounded = round_timeout(lock->l_callback_timeout);
 239
 240         if (timeout_rounded < waiting_locks_timer.expires ||
 241             !timer_pending(&waiting_locks_timer)) {
 242                 mod_timer(&waiting_locks_timer, timeout_rounded);
 243         }
 244         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 245         spin_unlock_bh(&waiting_locks_spinlock);
 246         return 1;
 247 }
 248
 249 /*
 250  * Remove a lock from the pending list, likely because it had its cancellation
 251  * callback arrive without incident.  This adjusts the lock-timeout timer if
 252  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 253  */
 254 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 255 {
 256         struct list_head *list_next;
 257
 258         if (lock->l_export == NULL) {
 259                 /* We don't have a "waiting locks list" on clients. */
 260                 LDLM_DEBUG(lock, "client lock: no-op");
 261                 return 0;
 262         }
 263
 264         spin_lock_bh(&waiting_locks_spinlock);
 265
 266         if (list_empty(&lock->l_pending_chain)) {
 267                 spin_unlock_bh(&waiting_locks_spinlock);
 268                 LDLM_DEBUG(lock, "wasn't waiting");
 269                 return 0;
 270         }
 271
 272         list_next = lock->l_pending_chain.next;
 273         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 274                 /* Removing the head of the list, adjust timer. */
 275                 if (list_next == &waiting_locks_list) {
 276                         /* No more, just cancel. */
 277                         del_timer(&waiting_locks_timer);
 278                 } else {
 279                         struct ldlm_lock *next;
 280                         next = list_entry(list_next, struct ldlm_lock,
 281                                           l_pending_chain);
 282                         mod_timer(&waiting_locks_timer,
 283                                   round_timeout(next->l_callback_timeout));
 284                 }
 285         }
 286
 287         spin_lock_bh(&expired_lock_thread.elt_lock);
 288         list_del_init(&lock->l_pending_chain);
 289         spin_unlock_bh(&expired_lock_thread.elt_lock);
 290
 291         spin_unlock_bh(&waiting_locks_spinlock);
 292         LDLM_DEBUG(lock, "removed");
 293         return 1;
 294 }
 295
 296 #else /* !__KERNEL__ */
 297
 298 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 299 {
 300         RETURN(1);
 301 }
 302
 303 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 304 {
 305         RETURN(0);
 306 }
 307
 308 #endif /* __KERNEL__ */
 309
 310 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,const char *ast_type)
 311 {
 312         const struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 313         char str[PTL_NALFMT_SIZE];
 314
 315         CERROR("%s AST failed (%d) for res "LPU64"/"LPU64
 316                ", mode %s: evicting client %s@%s NID "LPX64" (%s)\n",
 317                ast_type, rc,
 318                lock->l_resource->lr_name.name[0],
 319                lock->l_resource->lr_name.name[1],
 320                ldlm_lockname[lock->l_granted_mode],
 321                lock->l_export->exp_client_uuid.uuid,
 322                conn->c_remote_uuid.uuid, conn->c_peer.peer_nid,
 323                portals_nid2str(conn->c_peer.peer_ni->pni_number,
 324                                conn->c_peer.peer_nid, str));
 325         ptlrpc_fail_export(lock->l_export);
 326 }
 327
 328 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 329                                  struct ptlrpc_request *req, int rc,
 330                                  const char *ast_type)
 331 {
 332         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 333                 LASSERT(lock->l_export);
 334                 if (lock->l_export->exp_libclient) {
 335                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid "
 336                                    LPU64") timeout, just cancelling lock",
 337                                    ast_type, req->rq_peer.peer_nid);
 338                         ldlm_lock_cancel(lock);
 339                         rc = -ERESTART;
 340                 } else {
 341                         ldlm_del_waiting_lock(lock);
 342                         ldlm_failed_ast(lock, rc, ast_type);
 343                 }
 344         } else if (rc) {
 345                 if (rc == -EINVAL)
 346                         LDLM_DEBUG(lock, "client (nid "LPU64") returned %d"
 347                                    " from %s AST - normal race",
 348                                    req->rq_peer.peer_nid,
 349                                    req->rq_repmsg->status, ast_type);
 350                 else
 351                         LDLM_ERROR(lock, "client (nid "LPU64") returned %d "
 352                                    "from %s AST", req->rq_peer.peer_nid,
 353                                    (req->rq_repmsg != NULL) ?
 354                                    req->rq_repmsg->status : 0, ast_type);
 355                 ldlm_lock_cancel(lock);
 356                 /* Server-side AST functions are called from ldlm_reprocess_all,
 357                  * which needs to be told to please restart its reprocessing. */
 358                 rc = -ERESTART;
 359         }
 360
 361         return rc;
 362 }
 363
 364 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 365                              struct ldlm_lock_desc *desc,
 366                              void *data, int flag)
 367 {
 368         struct ldlm_request *body;
 369         struct ptlrpc_request *req;
 370         int rc = 0, size = sizeof(*body);
 371         ENTRY;
 372
 373         if (flag == LDLM_CB_CANCELING) {
 374                 /* Don't need to do anything here. */
 375                 RETURN(0);
 376         }
 377
 378         LASSERT(lock);
 379
 380         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 381         if (lock->l_granted_mode != lock->l_req_mode) {
 382                 /* this blocking AST will be communicated as part of the
 383                  * completion AST instead */
 384                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 385                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");                 RETURN(0);
 386         }
 387
 388         if (lock->l_destroyed) {
 389                 /* What's the point? */
 390                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 391                 RETURN(0);
 392         }
 393
 394 #if 0
 395         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 396                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 397                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 398                 RETURN(-ETIMEDOUT);
 399         }
 400 #endif
 401
 402         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 403                               LDLM_BL_CALLBACK, 1, &size, NULL);
 404         if (req == NULL) {
 405                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 406                 RETURN(-ENOMEM);
 407         }
 408
 409         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 410         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 411                sizeof(body->lock_handle1));
 412         memcpy(&body->lock_desc, desc, sizeof(*desc));
 413         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 414
 415         LDLM_DEBUG(lock, "server preparing blocking AST");
 416         req->rq_replen = lustre_msg_size(0, NULL);
 417
 418         if (lock->l_granted_mode == lock->l_req_mode)
 419                 ldlm_add_waiting_lock(lock);
 420         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 421
 422         req->rq_send_state = LUSTRE_IMP_FULL;
 423         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 424         rc = ptlrpc_queue_wait(req);
 425         if (rc != 0)
 426                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 427
 428         ptlrpc_req_finished(req);
 429
 430         RETURN(rc);
 431 }
 432
 433 /* XXX copied from ptlrpc/service.c */
 434 static long timeval_sub(struct timeval *large, struct timeval *small)
 435 {
 436         return (large->tv_sec - small->tv_sec) * 1000000 +
 437                 (large->tv_usec - small->tv_usec);
 438 }
 439
 440 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 441 {
 442         struct ldlm_request *body;
 443         struct ptlrpc_request *req;
 444         struct timeval granted_time;
 445         long total_enqueue_wait;
 446         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 447         ENTRY;
 448
 449         LASSERT(lock != NULL);
 450
 451         do_gettimeofday(&granted_time);
 452         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 453
 454         if (total_enqueue_wait / 1000000 > obd_timeout)
 455                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 456
 457         if (lock->l_resource->lr_lvb_len) {
 458                 buffers = 2;
 459                 size[1] = lock->l_resource->lr_lvb_len;
 460         }
 461
 462         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 463                               LDLM_CP_CALLBACK, buffers, size, NULL);
 464         if (req == NULL)
 465                 RETURN(-ENOMEM);
 466
 467         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 468         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 469                sizeof(body->lock_handle1));
 470         body->lock_flags = flags;
 471         ldlm_lock2desc(lock, &body->lock_desc);
 472
 473         if (buffers == 2) {
 474                 void *lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 475                                            lock->l_resource->lr_lvb_len);
 476                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 477                        lock->l_resource->lr_lvb_len);
 478         }
 479
 480         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 481                    total_enqueue_wait);
 482         req->rq_replen = lustre_msg_size(0, NULL);
 483
 484         req->rq_send_state = LUSTRE_IMP_FULL;
 485         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 486
 487         /* We only send real blocking ASTs after the lock is granted */
 488         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 489         if (lock->l_flags & LDLM_FL_AST_SENT) {
 490                 body->lock_flags |= LDLM_FL_AST_SENT;
 491                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 492         }
 493         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 494
 495         rc = ptlrpc_queue_wait(req);
 496         if (rc != 0)
 497                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 498
 499         ptlrpc_req_finished(req);
 500
 501         RETURN(rc);
 502 }
 503
 504 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 505 {
 506         struct ldlm_resource *res = lock->l_resource;
 507         struct ldlm_request *body;
 508         struct ptlrpc_request *req;
 509         int rc = 0, size = sizeof(*body);
 510         ENTRY;
 511
 512         LASSERT(lock != NULL);
 513
 514         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 515                               LDLM_GL_CALLBACK, 1, &size, NULL);
 516         if (req == NULL)
 517                 RETURN(-ENOMEM);
 518
 519         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 520         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 521                sizeof(body->lock_handle1));
 522         ldlm_lock2desc(lock, &body->lock_desc);
 523
 524         size = lock->l_resource->lr_lvb_len;
 525         req->rq_replen = lustre_msg_size(1, &size);
 526
 527         req->rq_send_state = LUSTRE_IMP_FULL;
 528         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 529
 530         rc = ptlrpc_queue_wait(req);
 531         if (rc == -ELDLM_NO_LOCK_DATA)
 532                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 533         else if (rc != 0)
 534                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 535         else
 536                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 537                         (res, req->rq_repmsg, 0, 1);
 538         ptlrpc_req_finished(req);
 539         RETURN(rc);
 540 }
 541
 542 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 543                         ldlm_completion_callback completion_callback,
 544                         ldlm_blocking_callback blocking_callback,
 545                         ldlm_glimpse_callback glimpse_callback)
 546 {
 547         struct obd_device *obddev = req->rq_export->exp_obd;
 548         struct ldlm_reply *dlm_rep;
 549         struct ldlm_request *dlm_req;
 550         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 551         __u32 flags;
 552         ldlm_error_t err = ELDLM_OK;
 553         struct ldlm_lock *lock = NULL;
 554         void *cookie = NULL;
 555         ENTRY;
 556
 557         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 558
 559         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 560                                       lustre_swab_ldlm_request);
 561         if (dlm_req == NULL) {
 562                 CERROR ("Can't unpack dlm_req\n");
 563                 GOTO(out, rc = -EFAULT);
 564         }
 565
 566         flags = dlm_req->lock_flags;
 567
 568         /* The lock's callback data might be set in the policy function */
 569         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 570                                 dlm_req->lock_desc.l_resource.lr_name,
 571                                 dlm_req->lock_desc.l_resource.lr_type,
 572                                 dlm_req->lock_desc.l_req_mode,
 573                                 blocking_callback, completion_callback,
 574                                 glimpse_callback, NULL, 0);
 575         if (!lock)
 576                 GOTO(out, rc = -ENOMEM);
 577
 578         do_gettimeofday(&lock->l_enqueued_time);
 579         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 580                sizeof(lock->l_remote_handle));
 581         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 582
 583         LASSERT(req->rq_export);
 584         lock->l_export = class_export_get(req->rq_export);
 585         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 586         list_add(&lock->l_export_chain,
 587                  &lock->l_export->exp_ldlm_data.led_held_locks);
 588         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 589
 590         if (flags & LDLM_FL_HAS_INTENT) {
 591                 /* In this case, the reply buffer is allocated deep in
 592                  * local_lock_enqueue by the policy function. */
 593                 cookie = req;
 594         } else {
 595                 int buffers = 1;
 596                 if (lock->l_resource->lr_lvb_len) {
 597                         size[1] = lock->l_resource->lr_lvb_len;
 598                         buffers = 2;
 599                 }
 600
 601                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 602                         GOTO(out, rc = -ENOMEM);
 603
 604                 rc = lustre_pack_reply(req, buffers, size, NULL);
 605                 if (rc)
 606                         GOTO(out, rc);
 607         }
 608
 609         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 610                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 611                        sizeof(ldlm_policy_data_t));
 612         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 613                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 614                        sizeof(lock->l_req_extent));
 615
 616         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 617         if (err)
 618                 GOTO(out, err);
 619
 620         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 621         dlm_rep->lock_flags = flags;
 622
 623         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 624         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 625
 626         /* We never send a blocking AST until the lock is granted, but
 627          * we can tell it right now */
 628         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 629         if (lock->l_flags & LDLM_FL_AST_SENT) {
 630                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 631                 if (lock->l_granted_mode == lock->l_req_mode)
 632                         ldlm_add_waiting_lock(lock);
 633         }
 634         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 635
 636         EXIT;
 637  out:
 638         req->rq_status = err;
 639         if (req->rq_reply_state == NULL) {
 640                 err = lustre_pack_reply(req, 0, NULL, NULL);
 641                 if (rc == 0)
 642                         rc = err;
 643                 req->rq_status = rc;
 644         }
 645
 646         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 647          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 648         if (lock) {
 649                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 650                            "(err=%d, rc=%d)", err, rc);
 651
 652                 if (lock->l_resource->lr_lvb_len > 0 && rc == 0) {
 653                         void *lvb = lustre_msg_buf(req->rq_repmsg, 1,
 654                                                   lock->l_resource->lr_lvb_len);
 655                         LASSERT(lvb != NULL);
 656                         memcpy(lvb, lock->l_resource->lr_lvb_data,
 657                                lock->l_resource->lr_lvb_len);
 658                 }
 659
 660                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 661                         ldlm_reprocess_all(lock->l_resource);
 662                 LDLM_LOCK_PUT(lock);
 663         }
 664         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 665                           lock, rc);
 666
 667         return rc;
 668 }
 669
 670 int ldlm_handle_convert(struct ptlrpc_request *req)
 671 {
 672         struct ldlm_request *dlm_req;
 673         struct ldlm_reply *dlm_rep;
 674         struct ldlm_lock *lock;
 675         int rc, size = sizeof(*dlm_rep);
 676         ENTRY;
 677
 678         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 679                                       lustre_swab_ldlm_request);
 680         if (dlm_req == NULL) {
 681                 CERROR ("Can't unpack dlm_req\n");
 682                 RETURN (-EFAULT);
 683         }
 684
 685         rc = lustre_pack_reply(req, 1, &size, NULL);
 686         if (rc) {
 687                 CERROR("out of memory\n");
 688                 RETURN(-ENOMEM);
 689         }
 690         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 691         dlm_rep->lock_flags = dlm_req->lock_flags;
 692
 693         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 694         if (!lock) {
 695                 req->rq_status = EINVAL;
 696         } else {
 697                 LDLM_DEBUG(lock, "server-side convert handler START");
 698                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 699                                   &dlm_rep->lock_flags);
 700                 if (ldlm_del_waiting_lock(lock))
 701                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 702                 req->rq_status = 0;
 703         }
 704
 705         if (lock) {
 706                 ldlm_reprocess_all(lock->l_resource);
 707                 LDLM_DEBUG(lock, "server-side convert handler END");
 708                 LDLM_LOCK_PUT(lock);
 709         } else
 710                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 711
 712         RETURN(0);
 713 }
 714
 715 int ldlm_handle_cancel(struct ptlrpc_request *req)
 716 {
 717         struct ldlm_request *dlm_req;
 718         struct ldlm_lock *lock;
 719         struct ldlm_resource *res;
 720         char str[PTL_NALFMT_SIZE];
 721         int rc;
 722         ENTRY;
 723
 724         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 725                                       lustre_swab_ldlm_request);
 726         if (dlm_req == NULL) {
 727                 CERROR("bad request buffer for cancel\n");
 728                 RETURN(-EFAULT);
 729         }
 730
 731         rc = lustre_pack_reply(req, 0, NULL, NULL);
 732         if (rc) {
 733                 CERROR("out of memory\n");
 734                 RETURN(-ENOMEM);
 735         }
 736
 737         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 738         if (!lock) {
 739                 CERROR("received cancel for unknown lock cookie "LPX64
 740                        " from client %s nid "LPX64" (%s)\n",
 741                        dlm_req->lock_handle1.cookie,
 742                        req->rq_export->exp_client_uuid.uuid,
 743                        req->rq_peer.peer_nid,
 744                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 745                                        req->rq_peer.peer_nid, str));
 746                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 747                                   "(cookie "LPU64")",
 748                                   dlm_req->lock_handle1.cookie);
 749                 req->rq_status = ESTALE;
 750         } else {
 751                 LDLM_DEBUG(lock, "server-side cancel handler START");
 752                 res = lock->l_resource;
 753                 if (res && res->lr_namespace->ns_lvbo &&
 754                     res->lr_namespace->ns_lvbo->lvbo_update) {
 755                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 756                                 (res, NULL, 0, 0);
 757                                 //(res, req->rq_reqmsg, 1);
 758                 }
 759
 760                 ldlm_lock_cancel(lock);
 761                 if (ldlm_del_waiting_lock(lock))
 762                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 763                 req->rq_status = rc;
 764         }
 765
 766         if (ptlrpc_reply(req) != 0)
 767                 LBUG();
 768
 769         if (lock) {
 770                 ldlm_reprocess_all(lock->l_resource);
 771                 LDLM_DEBUG(lock, "server-side cancel handler END");
 772                 LDLM_LOCK_PUT(lock);
 773         }
 774
 775         RETURN(0);
 776 }
 777
 778 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 779                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 780 {
 781         int do_ast;
 782         ENTRY;
 783
 784         l_lock(&ns->ns_lock);
 785         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 786
 787         lock->l_flags |= LDLM_FL_CBPENDING;
 788         do_ast = (!lock->l_readers && !lock->l_writers);
 789
 790         if (do_ast) {
 791                 LDLM_DEBUG(lock, "already unused, calling "
 792                            "callback (%p)", lock->l_blocking_ast);
 793                 if (lock->l_blocking_ast != NULL) {
 794                         l_unlock(&ns->ns_lock);
 795                         l_check_no_ns_lock(ns);
 796                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 797                                              LDLM_CB_BLOCKING);
 798                         l_lock(&ns->ns_lock);
 799                 }
 800         } else {
 801                 LDLM_DEBUG(lock, "Lock still has references, will be"
 802                            " cancelled later");
 803         }
 804
 805         LDLM_DEBUG(lock, "client blocking callback handler END");
 806         l_unlock(&ns->ns_lock);
 807         LDLM_LOCK_PUT(lock);
 808         EXIT;
 809 }
 810
 811 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 812                                     struct ldlm_namespace *ns,
 813                                     struct ldlm_request *dlm_req,
 814                                     struct ldlm_lock *lock)
 815 {
 816         LIST_HEAD(ast_list);
 817         ENTRY;
 818
 819         l_lock(&ns->ns_lock);
 820         LDLM_DEBUG(lock, "client completion callback handler START");
 821
 822         /* If we receive the completion AST before the actual enqueue returned,
 823          * then we might need to switch lock modes, resources, or extents. */
 824         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 825                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 826                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 827         }
 828
 829         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 830                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 831                        sizeof(lock->l_policy_data));
 832                 LDLM_DEBUG(lock, "completion AST, new policy data");
 833         }
 834
 835         ldlm_resource_unlink_lock(lock);
 836         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 837                    &lock->l_resource->lr_name,
 838                    sizeof(lock->l_resource->lr_name)) != 0) {
 839                 ldlm_lock_change_resource(ns, lock,
 840                                          dlm_req->lock_desc.l_resource.lr_name);
 841                 LDLM_DEBUG(lock, "completion AST, new resource");
 842         }
 843
 844         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 845                 lock->l_flags |= LDLM_FL_CBPENDING;
 846                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 847         }
 848
 849         if (lock->l_lvb_len) {
 850                 void *lvb;
 851                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 852                                          lock->l_lvb_swabber);
 853                 if (lvb == NULL) {
 854                         LDLM_ERROR(lock, "completion AST did not contain "
 855                                    "expected LVB!");
 856                 } else {
 857                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 858                 }
 859         }
 860
 861         lock->l_resource->lr_tmp = &ast_list;
 862         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 863         lock->l_resource->lr_tmp = NULL;
 864         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 865         l_unlock(&ns->ns_lock);
 866         LDLM_LOCK_PUT(lock);
 867
 868         ldlm_run_ast_work(ns, &ast_list);
 869
 870         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 871                           lock);
 872         EXIT;
 873 }
 874
 875 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 876                                     struct ldlm_namespace *ns,
 877                                     struct ldlm_request *dlm_req,
 878                                     struct ldlm_lock *lock)
 879 {
 880         int rc = -ENOSYS;
 881         ENTRY;
 882
 883         l_lock(&ns->ns_lock);
 884         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 885
 886         if (lock->l_glimpse_ast != NULL) {
 887                 l_unlock(&ns->ns_lock);
 888                 l_check_no_ns_lock(ns);
 889                 rc = lock->l_glimpse_ast(lock, req);
 890                 l_lock(&ns->ns_lock);
 891         }
 892
 893         if (req->rq_repmsg != NULL) {
 894                 ptlrpc_reply(req);
 895         } else {
 896                 req->rq_status = rc;
 897                 ptlrpc_error(req);
 898         }
 899
 900         if (lock->l_granted_mode == LCK_PW &&
 901             !lock->l_readers && !lock->l_writers &&
 902             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 903                 l_unlock(&ns->ns_lock);
 904                 ldlm_handle_bl_callback(ns, NULL, lock);
 905                 EXIT;
 906                 return;
 907         }
 908
 909         l_unlock(&ns->ns_lock);
 910         LDLM_LOCK_PUT(lock);
 911         EXIT;
 912 }
 913
 914 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 915 {
 916         req->rq_status = rc;
 917         if (req->rq_reply_state == NULL) {
 918                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 919                 if (rc)
 920                         return rc;
 921         }
 922         return ptlrpc_reply(req);
 923 }
 924
 925 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
 926                       struct ldlm_lock *lock)
 927 {
 928 #ifdef __KERNEL__
 929         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
 930         struct ldlm_bl_work_item *blwi;
 931         ENTRY;
 932
 933         OBD_ALLOC(blwi, sizeof(*blwi));
 934         if (blwi == NULL)
 935                 RETURN(-ENOMEM);
 936
 937         blwi->blwi_ns = ns;
 938         if (ld != NULL)
 939                 blwi->blwi_ld = *ld;
 940         blwi->blwi_lock = lock;
 941
 942         spin_lock(&blp->blp_lock);
 943         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
 944         wake_up(&blp->blp_waitq);
 945         spin_unlock(&blp->blp_lock);
 946 #else
 947         LBUG();
 948 #endif
 949
 950         RETURN(0);
 951 }
 952
 953 static int ldlm_callback_handler(struct ptlrpc_request *req)
 954 {
 955         struct ldlm_namespace *ns;
 956         struct ldlm_request *dlm_req;
 957         struct ldlm_lock *lock;
 958         char str[PTL_NALFMT_SIZE];
 959         int rc;
 960         ENTRY;
 961
 962         /* Requests arrive in sender's byte order.  The ptlrpc service
 963          * handler has already checked and, if necessary, byte-swapped the
 964          * incoming request message body, but I am responsible for the
 965          * message buffers. */
 966
 967         if (req->rq_export == NULL) {
 968                 struct ldlm_request *dlm_req;
 969
 970                 CDEBUG(D_RPCTRACE, "operation %d from nid "LPX64" (%s) with bad "
 971                        "export cookie "LPX64" (ptl req %d/rep %d); this is "
 972                        "normal if this node rebooted with a lock held\n",
 973                        req->rq_reqmsg->opc, req->rq_peer.peer_nid,
 974                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 975                                        req->rq_peer.peer_nid, str),
 976                        req->rq_reqmsg->handle.cookie,
 977                        req->rq_request_portal, req->rq_reply_portal);
 978
 979                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 980                                              lustre_swab_ldlm_request);
 981                 if (dlm_req != NULL)
 982                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
 983                                dlm_req->lock_handle1.cookie);
 984
 985                 ldlm_callback_reply(req, -ENOTCONN);
 986                 RETURN(0);
 987         }
 988
 989         LASSERT(req->rq_export != NULL);
 990         LASSERT(req->rq_export->exp_obd != NULL);
 991
 992         switch(req->rq_reqmsg->opc) {
 993         case LDLM_BL_CALLBACK:
 994                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 995                 break;
 996         case LDLM_CP_CALLBACK:
 997                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
 998                 break;
 999         case LDLM_GL_CALLBACK:
1000                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1001                 break;
1002         case OBD_LOG_CANCEL:
1003                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1004                 rc = llog_origin_handle_cancel(req);
1005                 ldlm_callback_reply(req, rc);
1006                 RETURN(0);
1007         case LLOG_ORIGIN_HANDLE_CREATE:
1008                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1009                 rc = llog_origin_handle_create(req);
1010                 ldlm_callback_reply(req, rc);
1011                 RETURN(0);
1012         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1013                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1014                 rc = llog_origin_handle_next_block(req);
1015                 ldlm_callback_reply(req, rc);
1016                 RETURN(0);
1017         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1018                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1019                 rc = llog_origin_handle_read_header(req);
1020                 ldlm_callback_reply(req, rc);
1021                 RETURN(0);
1022         case LLOG_ORIGIN_HANDLE_CLOSE:
1023                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1024                 rc = llog_origin_handle_close(req);
1025                 ldlm_callback_reply(req, rc);
1026                 RETURN(0);
1027         default:
1028                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1029                 ldlm_callback_reply(req, -EPROTO);
1030                 RETURN(0);
1031         }
1032
1033         ns = req->rq_export->exp_obd->obd_namespace;
1034         LASSERT(ns != NULL);
1035
1036         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1037                                       lustre_swab_ldlm_request);
1038         if (dlm_req == NULL) {
1039                 CERROR ("can't unpack dlm_req\n");
1040                 ldlm_callback_reply (req, -EPROTO);
1041                 RETURN (0);
1042         }
1043
1044         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1045         if (!lock) {
1046                 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
1047                        dlm_req->lock_handle1.cookie);
1048                 ldlm_callback_reply(req, -EINVAL);
1049                 RETURN(0);
1050         }
1051
1052         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1053         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1054
1055         /* We want the ost thread to get this reply so that it can respond
1056          * to ost requests (write cache writeback) that might be triggered
1057          * in the callback.
1058          *
1059          * But we'd also like to be able to indicate in the reply that we're
1060          * cancelling right now, because it's unused, or have an intent result
1061          * in the reply, so we might have to push the responsibility for sending
1062          * the reply down into the AST handlers, alas. */
1063
1064         switch (req->rq_reqmsg->opc) {
1065         case LDLM_BL_CALLBACK:
1066                 CDEBUG(D_INODE, "blocking ast\n");
1067 #ifdef __KERNEL__
1068                 rc = ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock);
1069                 ldlm_callback_reply(req, rc);
1070 #else
1071                 rc = 0;
1072                 ldlm_callback_reply(req, rc);
1073                 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1074 #endif
1075                 break;
1076         case LDLM_CP_CALLBACK:
1077                 CDEBUG(D_INODE, "completion ast\n");
1078                 ldlm_callback_reply(req, 0);
1079                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1080                 break;
1081         case LDLM_GL_CALLBACK:
1082                 CDEBUG(D_INODE, "glimpse ast\n");
1083                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1084                 break;
1085         default:
1086                 LBUG();                         /* checked above */
1087         }
1088
1089         RETURN(0);
1090 }
1091
1092 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1093 {
1094         int rc;
1095         ENTRY;
1096
1097         /* Requests arrive in sender's byte order.  The ptlrpc service
1098          * handler has already checked and, if necessary, byte-swapped the
1099          * incoming request message body, but I am responsible for the
1100          * message buffers. */
1101
1102         if (req->rq_export == NULL) {
1103                 struct ldlm_request *dlm_req;
1104                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
1105                        req->rq_reqmsg->opc, req->rq_request_portal,
1106                        req->rq_reply_portal);
1107                 CERROR("--> export cookie: "LPX64"\n",
1108                        req->rq_reqmsg->handle.cookie);
1109                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1110                                              lustre_swab_ldlm_request);
1111                 if (dlm_req != NULL)
1112                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1113                 RETURN(-ENOTCONN);
1114         }
1115
1116         switch (req->rq_reqmsg->opc) {
1117
1118         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1119         case LDLM_CANCEL:
1120                 CDEBUG(D_INODE, "cancel\n");
1121                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1122                 rc = ldlm_handle_cancel(req);
1123                 if (rc)
1124                         break;
1125                 RETURN(0);
1126
1127         default:
1128                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1129                 RETURN(-EINVAL);
1130         }
1131
1132         RETURN(0);
1133 }
1134
1135 #ifdef __KERNEL__
1136 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1137 {
1138         struct ldlm_bl_work_item *blwi = NULL;
1139
1140         spin_lock(&blp->blp_lock);
1141         if (!list_empty(&blp->blp_list)) {
1142                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1143                                   blwi_entry);
1144                 list_del(&blwi->blwi_entry);
1145         }
1146         spin_unlock(&blp->blp_lock);
1147
1148         return blwi;
1149 }
1150
1151 struct ldlm_bl_thread_data {
1152         int                     bltd_num;
1153         struct ldlm_bl_pool     *bltd_blp;
1154 };
1155
1156 static int ldlm_bl_thread_main(void *arg)
1157 {
1158         struct ldlm_bl_thread_data *bltd = arg;
1159         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1160         unsigned long flags;
1161         ENTRY;
1162
1163         /* XXX boiler-plate */
1164         {
1165                 char name[sizeof(current->comm)];
1166                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1167                          bltd->bltd_num);
1168                 kportal_daemonize(name);
1169         }
1170         SIGNAL_MASK_LOCK(current, flags);
1171         sigfillset(&current->blocked);
1172         RECALC_SIGPENDING;
1173         SIGNAL_MASK_UNLOCK(current, flags);
1174
1175         atomic_inc(&blp->blp_num_threads);
1176         complete(&blp->blp_comp);
1177
1178         while(1) {
1179                 struct l_wait_info lwi = { 0 };
1180                 struct ldlm_bl_work_item *blwi = NULL;
1181
1182                 l_wait_event_exclusive(blp->blp_waitq,
1183                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1184                                        &lwi);
1185
1186                 if (blwi->blwi_ns == NULL)
1187                         break;
1188
1189                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1190                                         blwi->blwi_lock);
1191                 OBD_FREE(blwi, sizeof(*blwi));
1192         }
1193
1194         atomic_dec(&blp->blp_num_threads);
1195         complete(&blp->blp_comp);
1196         RETURN(0);
1197 }
1198
1199 #endif
1200
1201 static int ldlm_setup(void);
1202 static int ldlm_cleanup(int force);
1203
1204 int ldlm_get_ref(void)
1205 {
1206         int rc = 0;
1207         down(&ldlm_ref_sem);
1208         if (++ldlm_refcount == 1) {
1209                 rc = ldlm_setup();
1210                 if (rc)
1211                         ldlm_refcount--;
1212         }
1213         up(&ldlm_ref_sem);
1214
1215         RETURN(rc);
1216 }
1217
1218 void ldlm_put_ref(int force)
1219 {
1220         down(&ldlm_ref_sem);
1221         if (ldlm_refcount == 1) {
1222                 int rc = ldlm_cleanup(force);
1223                 if (rc)
1224                         CERROR("ldlm_cleanup failed: %d\n", rc);
1225                 else
1226                         ldlm_refcount--;
1227         } else {
1228                 ldlm_refcount--;
1229         }
1230         up(&ldlm_ref_sem);
1231
1232         EXIT;
1233 }
1234
1235 static int ldlm_setup(void)
1236 {
1237         struct ldlm_bl_pool *blp;
1238         int rc = 0;
1239 #ifdef __KERNEL__
1240         int i;
1241 #endif
1242         ENTRY;
1243
1244         if (ldlm_state != NULL)
1245                 RETURN(-EALREADY);
1246
1247         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1248         if (ldlm_state == NULL)
1249                 RETURN(-ENOMEM);
1250
1251 #ifdef __KERNEL__
1252         rc = ldlm_proc_setup();
1253         if (rc != 0)
1254                 GOTO(out_free, rc);
1255 #endif
1256
1257         ldlm_state->ldlm_cb_service =
1258                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1259                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1260                                 ldlm_callback_handler, "ldlm_cbd",
1261                                 ldlm_svc_proc_dir);
1262
1263         if (!ldlm_state->ldlm_cb_service) {
1264                 CERROR("failed to start service\n");
1265                 GOTO(out_proc, rc = -ENOMEM);
1266         }
1267
1268         ldlm_state->ldlm_cancel_service =
1269                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1270                                 LDLM_CANCEL_REQUEST_PORTAL,
1271                                 LDLM_CANCEL_REPLY_PORTAL,
1272                                 ldlm_cancel_handler, "ldlm_canceld",
1273                                 ldlm_svc_proc_dir);
1274
1275         if (!ldlm_state->ldlm_cancel_service) {
1276                 CERROR("failed to start service\n");
1277                 GOTO(out_proc, rc = -ENOMEM);
1278         }
1279
1280         OBD_ALLOC(blp, sizeof(*blp));
1281         if (blp == NULL)
1282                 GOTO(out_proc, rc = -ENOMEM);
1283         ldlm_state->ldlm_bl_pool = blp;
1284
1285         atomic_set(&blp->blp_num_threads, 0);
1286         init_waitqueue_head(&blp->blp_waitq);
1287         spin_lock_init(&blp->blp_lock);
1288
1289         INIT_LIST_HEAD(&blp->blp_list);
1290
1291 #ifdef __KERNEL__
1292         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1293                 struct ldlm_bl_thread_data bltd = {
1294                         .bltd_num = i,
1295                         .bltd_blp = blp,
1296                 };
1297                 init_completion(&blp->blp_comp);
1298                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1299                 if (rc < 0) {
1300                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1301                         LBUG();
1302                         GOTO(out_thread, rc);
1303                 }
1304                 wait_for_completion(&blp->blp_comp);
1305         }
1306
1307         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1308                                     LDLM_NUM_THREADS, "ldlm_cn");
1309         if (rc) {
1310                 LBUG();
1311                 GOTO(out_thread, rc);
1312         }
1313
1314         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1315                                     LDLM_NUM_THREADS, "ldlm_cb");
1316         if (rc) {
1317                 LBUG();
1318                 GOTO(out_thread, rc);
1319         }
1320
1321         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1322         spin_lock_init(&expired_lock_thread.elt_lock);
1323         expired_lock_thread.elt_state = ELT_STOPPED;
1324         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1325
1326         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1327         if (rc < 0) {
1328                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1329                 GOTO(out_thread, rc);
1330         }
1331
1332         wait_event(expired_lock_thread.elt_waitq,
1333                    expired_lock_thread.elt_state == ELT_READY);
1334
1335         INIT_LIST_HEAD(&waiting_locks_list);
1336         spin_lock_init(&waiting_locks_spinlock);
1337         waiting_locks_timer.function = waiting_locks_callback;
1338         waiting_locks_timer.data = 0;
1339         init_timer(&waiting_locks_timer);
1340 #endif
1341
1342         RETURN(0);
1343
1344 #ifdef __KERNEL__
1345  out_thread:
1346         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1347         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1348 #endif
1349
1350  out_proc:
1351 #ifdef __KERNEL__
1352         ldlm_proc_cleanup();
1353  out_free:
1354 #endif
1355         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1356         ldlm_state = NULL;
1357         return rc;
1358 }
1359
1360 static int ldlm_cleanup(int force)
1361 {
1362 #ifdef __KERNEL__
1363         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1364 #endif
1365         ENTRY;
1366
1367         if (!list_empty(&ldlm_namespace_list)) {
1368                 CERROR("ldlm still has namespaces; clean these up first.\n");
1369                 ldlm_dump_all_namespaces();
1370                 RETURN(-EBUSY);
1371         }
1372
1373 #ifdef __KERNEL__
1374         while (atomic_read(&blp->blp_num_threads) > 0) {
1375                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1376
1377                 init_completion(&blp->blp_comp);
1378
1379                 spin_lock(&blp->blp_lock);
1380                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1381                 wake_up(&blp->blp_waitq);
1382                 spin_unlock(&blp->blp_lock);
1383
1384                 wait_for_completion(&blp->blp_comp);
1385         }
1386         OBD_FREE(blp, sizeof(*blp));
1387
1388         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1389         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1390         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1391         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1392         ldlm_proc_cleanup();
1393
1394         expired_lock_thread.elt_state = ELT_TERMINATE;
1395         wake_up(&expired_lock_thread.elt_waitq);
1396         wait_event(expired_lock_thread.elt_waitq,
1397                    expired_lock_thread.elt_state == ELT_STOPPED);
1398
1399 #endif
1400
1401         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1402         ldlm_state = NULL;
1403
1404         RETURN(0);
1405 }
1406
1407 int __init ldlm_init(void)
1408 {
1409         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1410                                                sizeof(struct ldlm_resource), 0,
1411                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1412         if (ldlm_resource_slab == NULL)
1413                 return -ENOMEM;
1414
1415         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1416                                            sizeof(struct ldlm_lock), 0,
1417                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1418         if (ldlm_lock_slab == NULL) {
1419                 kmem_cache_destroy(ldlm_resource_slab);
1420                 return -ENOMEM;
1421         }
1422
1423         l_lock_init(&ldlm_handle_lock);
1424
1425         return 0;
1426 }
1427
1428 void __exit ldlm_exit(void)
1429 {
1430         if ( ldlm_refcount )
1431                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1432         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
1433                 CERROR("couldn't free ldlm resource slab\n");
1434         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
1435                 CERROR("couldn't free ldlm lock slab\n");
1436 }
1437
1438 /* ldlm_flock.c */
1439 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1440
1441 /* ldlm_extent.c */
1442 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1443
1444 /* ldlm_lock.c */
1445 EXPORT_SYMBOL(ldlm_get_processing_policy);
1446 EXPORT_SYMBOL(ldlm_lock2desc);
1447 EXPORT_SYMBOL(ldlm_register_intent);
1448 EXPORT_SYMBOL(ldlm_lockname);
1449 EXPORT_SYMBOL(ldlm_typename);
1450 EXPORT_SYMBOL(ldlm_lock2handle);
1451 EXPORT_SYMBOL(__ldlm_handle2lock);
1452 EXPORT_SYMBOL(ldlm_lock_get);
1453 EXPORT_SYMBOL(ldlm_lock_put);
1454 EXPORT_SYMBOL(ldlm_lock_match);
1455 EXPORT_SYMBOL(ldlm_lock_cancel);
1456 EXPORT_SYMBOL(ldlm_lock_addref);
1457 EXPORT_SYMBOL(ldlm_lock_decref);
1458 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1459 EXPORT_SYMBOL(ldlm_lock_change_resource);
1460 EXPORT_SYMBOL(ldlm_lock_set_data);
1461 EXPORT_SYMBOL(ldlm_it2str);
1462 EXPORT_SYMBOL(ldlm_lock_dump);
1463 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1464 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1465 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1466 EXPORT_SYMBOL(ldlm_lock_allow_match);
1467
1468 /* ldlm_request.c */
1469 EXPORT_SYMBOL(ldlm_completion_ast);
1470 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1471 EXPORT_SYMBOL(ldlm_cli_convert);
1472 EXPORT_SYMBOL(ldlm_cli_enqueue);
1473 EXPORT_SYMBOL(ldlm_cli_cancel);
1474 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1475 EXPORT_SYMBOL(ldlm_replay_locks);
1476 EXPORT_SYMBOL(ldlm_resource_foreach);
1477 EXPORT_SYMBOL(ldlm_namespace_foreach);
1478 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1479 EXPORT_SYMBOL(ldlm_change_cbdata);
1480
1481 /* ldlm_lockd.c */
1482 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1483 EXPORT_SYMBOL(ldlm_server_completion_ast);
1484 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1485 EXPORT_SYMBOL(ldlm_handle_enqueue);
1486 EXPORT_SYMBOL(ldlm_handle_cancel);
1487 EXPORT_SYMBOL(ldlm_handle_convert);
1488 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1489 EXPORT_SYMBOL(ldlm_get_ref);
1490 EXPORT_SYMBOL(ldlm_put_ref);
1491
1492 #if 0
1493 /* ldlm_test.c */
1494 EXPORT_SYMBOL(ldlm_test);
1495 EXPORT_SYMBOL(ldlm_regression_start);
1496 EXPORT_SYMBOL(ldlm_regression_stop);
1497 #endif
1498
1499 /* ldlm_resource.c */
1500 EXPORT_SYMBOL(ldlm_namespace_new);
1501 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1502 EXPORT_SYMBOL(ldlm_namespace_free);
1503 EXPORT_SYMBOL(ldlm_namespace_dump);
1504 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1505 EXPORT_SYMBOL(ldlm_resource_get);
1506 EXPORT_SYMBOL(ldlm_resource_putref);
1507
1508 /* l_lock.c */
1509 EXPORT_SYMBOL(l_lock);
1510 EXPORT_SYMBOL(l_unlock);
1511
1512 /* ldlm_lib.c */
1513 EXPORT_SYMBOL(client_obd_setup);
1514 EXPORT_SYMBOL(client_obd_cleanup);
1515 EXPORT_SYMBOL(client_connect_import);
1516 EXPORT_SYMBOL(client_disconnect_export);
1517 EXPORT_SYMBOL(target_abort_recovery);
1518 EXPORT_SYMBOL(target_handle_connect);
1519 EXPORT_SYMBOL(target_destroy_export);
1520 EXPORT_SYMBOL(target_cancel_recovery_timer);
1521 EXPORT_SYMBOL(target_send_reply);
1522 EXPORT_SYMBOL(target_queue_recovery_request);
1523 EXPORT_SYMBOL(target_handle_ping);
1524 EXPORT_SYMBOL(target_handle_disconnect);
1525 EXPORT_SYMBOL(target_queue_final_reply);