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 %s ",
 189                            lock->l_export->exp_client_uuid.uuid,
 190                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 191                            ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str));
 192
 193                 spin_lock_bh(&expired_lock_thread.elt_lock);
 194                 list_del(&lock->l_pending_chain);
 195                 list_add(&lock->l_pending_chain,
 196                          &expired_lock_thread.elt_expired_locks);
 197                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 198                 wake_up(&expired_lock_thread.elt_waitq);
 199         }
 200
 201         /*
 202          * Make sure the timer will fire again if we have any locks
 203          * left.
 204          */
 205         if (!list_empty(&waiting_locks_list)) {
 206                 unsigned long timeout_rounded;
 207                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 208                                   l_pending_chain);
 209                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 210                 mod_timer(&waiting_locks_timer, timeout_rounded);
 211         }
 212         spin_unlock_bh(&waiting_locks_spinlock);
 213 }
 214
 215 /*
 216  * Indicate that we're waiting for a client to call us back cancelling a given
 217  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 218  * timer to fire appropriately.  (We round up to the next second, to avoid
 219  * floods of timer firings during periods of high lock contention and traffic).
 220  */
 221 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 222 {
 223         unsigned long timeout_rounded;
 224
 225         spin_lock_bh(&waiting_locks_spinlock);
 226         if (!list_empty(&lock->l_pending_chain)) {
 227                 LDLM_DEBUG(lock, "not re-adding to wait list");
 228                 spin_unlock_bh(&waiting_locks_spinlock);
 229                 return 0;
 230         }
 231         LDLM_DEBUG(lock, "adding to wait list");
 232
 233         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 234
 235         timeout_rounded = round_timeout(lock->l_callback_timeout);
 236
 237         if (timeout_rounded < waiting_locks_timer.expires ||
 238             !timer_pending(&waiting_locks_timer)) {
 239                 mod_timer(&waiting_locks_timer, timeout_rounded);
 240         }
 241         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 242         spin_unlock_bh(&waiting_locks_spinlock);
 243         return 1;
 244 }
 245
 246 /*
 247  * Remove a lock from the pending list, likely because it had its cancellation
 248  * callback arrive without incident.  This adjusts the lock-timeout timer if
 249  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 250  */
 251 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 252 {
 253         struct list_head *list_next;
 254
 255         if (lock->l_export == NULL) {
 256                 /* We don't have a "waiting locks list" on clients. */
 257                 LDLM_DEBUG(lock, "client lock: no-op");
 258                 return 0;
 259         }
 260
 261         spin_lock_bh(&waiting_locks_spinlock);
 262
 263         if (list_empty(&lock->l_pending_chain)) {
 264                 spin_unlock_bh(&waiting_locks_spinlock);
 265                 LDLM_DEBUG(lock, "wasn't waiting");
 266                 return 0;
 267         }
 268
 269         list_next = lock->l_pending_chain.next;
 270         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 271                 /* Removing the head of the list, adjust timer. */
 272                 if (list_next == &waiting_locks_list) {
 273                         /* No more, just cancel. */
 274                         del_timer(&waiting_locks_timer);
 275                 } else {
 276                         struct ldlm_lock *next;
 277                         next = list_entry(list_next, struct ldlm_lock,
 278                                           l_pending_chain);
 279                         mod_timer(&waiting_locks_timer,
 280                                   round_timeout(next->l_callback_timeout));
 281                 }
 282         }
 283
 284         spin_lock_bh(&expired_lock_thread.elt_lock);
 285         list_del_init(&lock->l_pending_chain);
 286         spin_unlock_bh(&expired_lock_thread.elt_lock);
 287
 288         spin_unlock_bh(&waiting_locks_spinlock);
 289         LDLM_DEBUG(lock, "removed");
 290         return 1;
 291 }
 292
 293 #else /* !__KERNEL__ */
 294
 295 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 296 {
 297         RETURN(1);
 298 }
 299
 300 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 301 {
 302         RETURN(0);
 303 }
 304
 305 #endif /* __KERNEL__ */
 306
 307 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,const char *ast_type)
 308 {
 309         struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 310         char str[PTL_NALFMT_SIZE];
 311
 312         CERROR("%s AST failed (%d) for res "LPU64"/"LPU64
 313                ", mode %s: evicting client %s@%s NID %s\n",
 314                ast_type, rc,
 315                lock->l_resource->lr_name.name[0],
 316                lock->l_resource->lr_name.name[1],
 317                ldlm_lockname[lock->l_granted_mode],
 318                lock->l_export->exp_client_uuid.uuid,
 319                conn->c_remote_uuid.uuid,
 320                ptlrpc_peernid2str(&conn->c_peer, str));
 321         ptlrpc_fail_export(lock->l_export);
 322 }
 323
 324 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 325                                  struct ptlrpc_request *req, int rc,
 326                                  const char *ast_type)
 327 {
 328         char str[PTL_NALFMT_SIZE];
 329
 330         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 331                 LASSERT(lock->l_export);
 332                 if (lock->l_export->exp_libclient) {
 333                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 334                                    " timeout, just cancelling lock", ast_type,
 335                                    ptlrpc_peernid2str(&req->rq_peer, str));
 336                         ldlm_lock_cancel(lock);
 337                         rc = -ERESTART;
 338                 } else {
 339                         ldlm_del_waiting_lock(lock);
 340                         ldlm_failed_ast(lock, rc, ast_type);
 341                 }
 342         } else if (rc) {
 343                 if (rc == -EINVAL)
 344                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 345                                    " from %s AST - normal race",
 346                                    ptlrpc_peernid2str(&req->rq_peer, str),
 347                                    req->rq_repmsg->status, ast_type);
 348                 else
 349                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 350                                    "from %s AST",
 351                                    ptlrpc_peernid2str(&req->rq_peer, str),
 352                                    (req->rq_repmsg != NULL) ?
 353                                    req->rq_repmsg->status : 0, ast_type);
 354                 ldlm_lock_cancel(lock);
 355                 /* Server-side AST functions are called from ldlm_reprocess_all,
 356                  * which needs to be told to please restart its reprocessing. */
 357                 rc = -ERESTART;
 358         }
 359
 360         return rc;
 361 }
 362
 363 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 364                              struct ldlm_lock_desc *desc,
 365                              void *data, int flag)
 366 {
 367         struct ldlm_request *body;
 368         struct ptlrpc_request *req;
 369         int rc = 0, size = sizeof(*body);
 370         ENTRY;
 371
 372         if (flag == LDLM_CB_CANCELING) {
 373                 /* Don't need to do anything here. */
 374                 RETURN(0);
 375         }
 376
 377         LASSERT(lock);
 378
 379         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 380         if (lock->l_granted_mode != lock->l_req_mode) {
 381                 /* this blocking AST will be communicated as part of the
 382                  * completion AST instead */
 383                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 384                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");                 RETURN(0);
 385         }
 386
 387         if (lock->l_destroyed) {
 388                 /* What's the point? */
 389                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 390                 RETURN(0);
 391         }
 392
 393 #if 0
 394         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 395                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 396                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 397                 RETURN(-ETIMEDOUT);
 398         }
 399 #endif
 400
 401         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 402                               LDLM_BL_CALLBACK, 1, &size, NULL);
 403         if (req == NULL) {
 404                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 405                 RETURN(-ENOMEM);
 406         }
 407
 408         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 409         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 410                sizeof(body->lock_handle1));
 411         memcpy(&body->lock_desc, desc, sizeof(*desc));
 412         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 413
 414         LDLM_DEBUG(lock, "server preparing blocking AST");
 415         req->rq_replen = lustre_msg_size(0, NULL);
 416
 417         if (lock->l_granted_mode == lock->l_req_mode)
 418                 ldlm_add_waiting_lock(lock);
 419         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 420
 421         req->rq_send_state = LUSTRE_IMP_FULL;
 422         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 423         rc = ptlrpc_queue_wait(req);
 424         if (rc != 0)
 425                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 426
 427         ptlrpc_req_finished(req);
 428
 429         RETURN(rc);
 430 }
 431
 432 /* XXX copied from ptlrpc/service.c */
 433 static long timeval_sub(struct timeval *large, struct timeval *small)
 434 {
 435         return (large->tv_sec - small->tv_sec) * 1000000 +
 436                 (large->tv_usec - small->tv_usec);
 437 }
 438
 439 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 440 {
 441         struct ldlm_request *body;
 442         struct ptlrpc_request *req;
 443         struct timeval granted_time;
 444         long total_enqueue_wait;
 445         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 446         ENTRY;
 447
 448         LASSERT(lock != NULL);
 449
 450         do_gettimeofday(&granted_time);
 451         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 452
 453         if (total_enqueue_wait / 1000000 > obd_timeout)
 454                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 455
 456         if (lock->l_resource->lr_lvb_len) {
 457                 buffers = 2;
 458                 size[1] = lock->l_resource->lr_lvb_len;
 459         }
 460
 461         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 462                               LDLM_CP_CALLBACK, buffers, size, NULL);
 463         if (req == NULL)
 464                 RETURN(-ENOMEM);
 465
 466         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 467         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 468                sizeof(body->lock_handle1));
 469         body->lock_flags = flags;
 470         ldlm_lock2desc(lock, &body->lock_desc);
 471
 472         if (buffers == 2) {
 473                 void *lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 474                                            lock->l_resource->lr_lvb_len);
 475                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 476                        lock->l_resource->lr_lvb_len);
 477         }
 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 != 0)
 496                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 497
 498         ptlrpc_req_finished(req);
 499
 500         RETURN(rc);
 501 }
 502
 503 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 504 {
 505         struct ldlm_resource *res = lock->l_resource;
 506         struct ldlm_request *body;
 507         struct ptlrpc_request *req;
 508         int rc = 0, size = sizeof(*body);
 509         ENTRY;
 510
 511         LASSERT(lock != NULL);
 512
 513         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 514                               LDLM_GL_CALLBACK, 1, &size, NULL);
 515         if (req == NULL)
 516                 RETURN(-ENOMEM);
 517
 518         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 519         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 520                sizeof(body->lock_handle1));
 521         ldlm_lock2desc(lock, &body->lock_desc);
 522
 523         size = lock->l_resource->lr_lvb_len;
 524         req->rq_replen = lustre_msg_size(1, &size);
 525
 526         req->rq_send_state = LUSTRE_IMP_FULL;
 527         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 528
 529         rc = ptlrpc_queue_wait(req);
 530         if (rc == -ELDLM_NO_LOCK_DATA)
 531                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 532         else if (rc != 0)
 533                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 534         else
 535                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 536                         (res, req->rq_repmsg, 0, 1);
 537         ptlrpc_req_finished(req);
 538         RETURN(rc);
 539 }
 540
 541 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 542                         ldlm_completion_callback completion_callback,
 543                         ldlm_blocking_callback blocking_callback,
 544                         ldlm_glimpse_callback glimpse_callback)
 545 {
 546         struct obd_device *obddev = req->rq_export->exp_obd;
 547         struct ldlm_reply *dlm_rep;
 548         struct ldlm_request *dlm_req;
 549         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 550         __u32 flags;
 551         ldlm_error_t err = ELDLM_OK;
 552         struct ldlm_lock *lock = NULL;
 553         void *cookie = NULL;
 554         ENTRY;
 555
 556         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 557
 558         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 559                                       lustre_swab_ldlm_request);
 560         if (dlm_req == NULL) {
 561                 CERROR ("Can't unpack dlm_req\n");
 562                 GOTO(out, rc = -EFAULT);
 563         }
 564
 565         flags = dlm_req->lock_flags;
 566
 567         /* The lock's callback data might be set in the policy function */
 568         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 569                                 dlm_req->lock_desc.l_resource.lr_name,
 570                                 dlm_req->lock_desc.l_resource.lr_type,
 571                                 dlm_req->lock_desc.l_req_mode,
 572                                 blocking_callback, completion_callback,
 573                                 glimpse_callback, NULL, 0);
 574         if (!lock)
 575                 GOTO(out, rc = -ENOMEM);
 576
 577         do_gettimeofday(&lock->l_enqueued_time);
 578         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 579                sizeof(lock->l_remote_handle));
 580         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 581
 582         LASSERT(req->rq_export);
 583         lock->l_export = class_export_get(req->rq_export);
 584         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 585         list_add(&lock->l_export_chain,
 586                  &lock->l_export->exp_ldlm_data.led_held_locks);
 587         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 588
 589         if (flags & LDLM_FL_HAS_INTENT) {
 590                 /* In this case, the reply buffer is allocated deep in
 591                  * local_lock_enqueue by the policy function. */
 592                 cookie = req;
 593         } else {
 594                 int buffers = 1;
 595                 if (lock->l_resource->lr_lvb_len) {
 596                         size[1] = lock->l_resource->lr_lvb_len;
 597                         buffers = 2;
 598                 }
 599
 600                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 601                         GOTO(out, rc = -ENOMEM);
 602
 603                 rc = lustre_pack_reply(req, buffers, size, NULL);
 604                 if (rc)
 605                         GOTO(out, rc);
 606         }
 607
 608         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 609                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 610                        sizeof(ldlm_policy_data_t));
 611         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 612                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 613                        sizeof(lock->l_req_extent));
 614
 615         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 616         if (err)
 617                 GOTO(out, err);
 618
 619         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 620         dlm_rep->lock_flags = flags;
 621
 622         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 623         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 624
 625         /* We never send a blocking AST until the lock is granted, but
 626          * we can tell it right now */
 627         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 628         if (lock->l_flags & LDLM_FL_AST_SENT) {
 629                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 630                 if (lock->l_granted_mode == lock->l_req_mode)
 631                         ldlm_add_waiting_lock(lock);
 632         }
 633         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 634
 635         EXIT;
 636  out:
 637         req->rq_status = err;
 638         if (req->rq_reply_state == NULL) {
 639                 err = lustre_pack_reply(req, 0, NULL, NULL);
 640                 if (rc == 0)
 641                         rc = err;
 642                 req->rq_status = rc;
 643         }
 644
 645         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 646          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 647         if (lock) {
 648                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 649                            "(err=%d, rc=%d)", err, rc);
 650
 651                 if (lock->l_resource->lr_lvb_len > 0 && rc == 0) {
 652                         void *lvb = lustre_msg_buf(req->rq_repmsg, 1,
 653                                                   lock->l_resource->lr_lvb_len);
 654                         LASSERT(lvb != NULL);
 655                         memcpy(lvb, lock->l_resource->lr_lvb_data,
 656                                lock->l_resource->lr_lvb_len);
 657                 }
 658
 659                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 660                         ldlm_reprocess_all(lock->l_resource);
 661                 LDLM_LOCK_PUT(lock);
 662         }
 663         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 664                           lock, rc);
 665
 666         return rc;
 667 }
 668
 669 int ldlm_handle_convert(struct ptlrpc_request *req)
 670 {
 671         struct ldlm_request *dlm_req;
 672         struct ldlm_reply *dlm_rep;
 673         struct ldlm_lock *lock;
 674         int rc, size = sizeof(*dlm_rep);
 675         ENTRY;
 676
 677         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 678                                       lustre_swab_ldlm_request);
 679         if (dlm_req == NULL) {
 680                 CERROR ("Can't unpack dlm_req\n");
 681                 RETURN (-EFAULT);
 682         }
 683
 684         rc = lustre_pack_reply(req, 1, &size, NULL);
 685         if (rc) {
 686                 CERROR("out of memory\n");
 687                 RETURN(-ENOMEM);
 688         }
 689         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 690         dlm_rep->lock_flags = dlm_req->lock_flags;
 691
 692         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 693         if (!lock) {
 694                 req->rq_status = EINVAL;
 695         } else {
 696                 LDLM_DEBUG(lock, "server-side convert handler START");
 697                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 698                                   &dlm_rep->lock_flags);
 699                 if (ldlm_del_waiting_lock(lock))
 700                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 701                 req->rq_status = 0;
 702         }
 703
 704         if (lock) {
 705                 ldlm_reprocess_all(lock->l_resource);
 706                 LDLM_DEBUG(lock, "server-side convert handler END");
 707                 LDLM_LOCK_PUT(lock);
 708         } else
 709                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 710
 711         RETURN(0);
 712 }
 713
 714 int ldlm_handle_cancel(struct ptlrpc_request *req)
 715 {
 716         struct ldlm_request *dlm_req;
 717         struct ldlm_lock *lock;
 718         struct ldlm_resource *res;
 719         char str[PTL_NALFMT_SIZE];
 720         int rc;
 721         ENTRY;
 722
 723         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 724                                       lustre_swab_ldlm_request);
 725         if (dlm_req == NULL) {
 726                 CERROR("bad request buffer for cancel\n");
 727                 RETURN(-EFAULT);
 728         }
 729
 730         rc = lustre_pack_reply(req, 0, NULL, NULL);
 731         if (rc) {
 732                 CERROR("out of memory\n");
 733                 RETURN(-ENOMEM);
 734         }
 735
 736         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 737         if (!lock) {
 738                 CERROR("received cancel for unknown lock cookie "LPX64
 739                        " from client %s nid %s\n",
 740                        dlm_req->lock_handle1.cookie,
 741                        req->rq_export->exp_client_uuid.uuid,
 742                        ptlrpc_peernid2str(&req->rq_peer, str));
 743                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 744                                   "(cookie "LPU64")",
 745                                   dlm_req->lock_handle1.cookie);
 746                 req->rq_status = ESTALE;
 747         } else {
 748                 LDLM_DEBUG(lock, "server-side cancel handler START");
 749                 res = lock->l_resource;
 750                 if (res && res->lr_namespace->ns_lvbo &&
 751                     res->lr_namespace->ns_lvbo->lvbo_update) {
 752                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 753                                 (res, NULL, 0, 0);
 754                                 //(res, req->rq_reqmsg, 1);
 755                 }
 756
 757                 ldlm_lock_cancel(lock);
 758                 if (ldlm_del_waiting_lock(lock))
 759                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 760                 req->rq_status = rc;
 761         }
 762
 763         if (ptlrpc_reply(req) != 0)
 764                 LBUG();
 765
 766         if (lock) {
 767                 ldlm_reprocess_all(lock->l_resource);
 768                 LDLM_DEBUG(lock, "server-side cancel handler END");
 769                 LDLM_LOCK_PUT(lock);
 770         }
 771
 772         RETURN(0);
 773 }
 774
 775 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 776                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 777 {
 778         int do_ast;
 779         ENTRY;
 780
 781         l_lock(&ns->ns_lock);
 782         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 783
 784         lock->l_flags |= LDLM_FL_CBPENDING;
 785         do_ast = (!lock->l_readers && !lock->l_writers);
 786
 787         if (do_ast) {
 788                 LDLM_DEBUG(lock, "already unused, calling "
 789                            "callback (%p)", lock->l_blocking_ast);
 790                 if (lock->l_blocking_ast != NULL) {
 791                         l_unlock(&ns->ns_lock);
 792                         l_check_no_ns_lock(ns);
 793                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 794                                              LDLM_CB_BLOCKING);
 795                         l_lock(&ns->ns_lock);
 796                 }
 797         } else {
 798                 LDLM_DEBUG(lock, "Lock still has references, will be"
 799                            " cancelled later");
 800         }
 801
 802         LDLM_DEBUG(lock, "client blocking callback handler END");
 803         l_unlock(&ns->ns_lock);
 804         LDLM_LOCK_PUT(lock);
 805         EXIT;
 806 }
 807
 808 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 809                                     struct ldlm_namespace *ns,
 810                                     struct ldlm_request *dlm_req,
 811                                     struct ldlm_lock *lock)
 812 {
 813         LIST_HEAD(ast_list);
 814         ENTRY;
 815
 816         l_lock(&ns->ns_lock);
 817         LDLM_DEBUG(lock, "client completion callback handler START");
 818
 819         /* If we receive the completion AST before the actual enqueue returned,
 820          * then we might need to switch lock modes, resources, or extents. */
 821         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 822                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 823                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 824         }
 825
 826         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 827                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 828                        sizeof(lock->l_policy_data));
 829                 LDLM_DEBUG(lock, "completion AST, new policy data");
 830         }
 831
 832         ldlm_resource_unlink_lock(lock);
 833         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 834                    &lock->l_resource->lr_name,
 835                    sizeof(lock->l_resource->lr_name)) != 0) {
 836                 ldlm_lock_change_resource(ns, lock,
 837                                          dlm_req->lock_desc.l_resource.lr_name);
 838                 LDLM_DEBUG(lock, "completion AST, new resource");
 839         }
 840
 841         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 842                 lock->l_flags |= LDLM_FL_CBPENDING;
 843                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 844         }
 845
 846         if (lock->l_lvb_len) {
 847                 void *lvb;
 848                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 849                                          lock->l_lvb_swabber);
 850                 if (lvb == NULL) {
 851                         LDLM_ERROR(lock, "completion AST did not contain "
 852                                    "expected LVB!");
 853                 } else {
 854                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 855                 }
 856         }
 857
 858         lock->l_resource->lr_tmp = &ast_list;
 859         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 860         lock->l_resource->lr_tmp = NULL;
 861         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 862         l_unlock(&ns->ns_lock);
 863         LDLM_LOCK_PUT(lock);
 864
 865         ldlm_run_ast_work(ns, &ast_list);
 866
 867         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 868                           lock);
 869         EXIT;
 870 }
 871
 872 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 873                                     struct ldlm_namespace *ns,
 874                                     struct ldlm_request *dlm_req,
 875                                     struct ldlm_lock *lock)
 876 {
 877         int rc = -ENOSYS;
 878         ENTRY;
 879
 880         l_lock(&ns->ns_lock);
 881         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 882
 883         if (lock->l_glimpse_ast != NULL) {
 884                 l_unlock(&ns->ns_lock);
 885                 l_check_no_ns_lock(ns);
 886                 rc = lock->l_glimpse_ast(lock, req);
 887                 l_lock(&ns->ns_lock);
 888         }
 889
 890         if (req->rq_repmsg != NULL) {
 891                 ptlrpc_reply(req);
 892         } else {
 893                 req->rq_status = rc;
 894                 ptlrpc_error(req);
 895         }
 896
 897         if (lock->l_granted_mode == LCK_PW &&
 898             !lock->l_readers && !lock->l_writers &&
 899             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 900                 l_unlock(&ns->ns_lock);
 901                 ldlm_handle_bl_callback(ns, NULL, lock);
 902                 EXIT;
 903                 return;
 904         }
 905
 906         l_unlock(&ns->ns_lock);
 907         LDLM_LOCK_PUT(lock);
 908         EXIT;
 909 }
 910
 911 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 912 {
 913         req->rq_status = rc;
 914         if (req->rq_reply_state == NULL) {
 915                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 916                 if (rc)
 917                         return rc;
 918         }
 919         return ptlrpc_reply(req);
 920 }
 921
 922 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
 923                       struct ldlm_lock *lock)
 924 {
 925 #ifdef __KERNEL__
 926         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
 927         struct ldlm_bl_work_item *blwi;
 928         ENTRY;
 929
 930         OBD_ALLOC(blwi, sizeof(*blwi));
 931         if (blwi == NULL)
 932                 RETURN(-ENOMEM);
 933
 934         blwi->blwi_ns = ns;
 935         if (ld != NULL)
 936                 blwi->blwi_ld = *ld;
 937         blwi->blwi_lock = lock;
 938
 939         spin_lock(&blp->blp_lock);
 940         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
 941         wake_up(&blp->blp_waitq);
 942         spin_unlock(&blp->blp_lock);
 943 #else
 944         LBUG();
 945 #endif
 946
 947         RETURN(0);
 948 }
 949
 950 static int ldlm_callback_handler(struct ptlrpc_request *req)
 951 {
 952         struct ldlm_namespace *ns;
 953         struct ldlm_request *dlm_req;
 954         struct ldlm_lock *lock;
 955         char str[PTL_NALFMT_SIZE];
 956         int rc;
 957         ENTRY;
 958
 959         /* Requests arrive in sender's byte order.  The ptlrpc service
 960          * handler has already checked and, if necessary, byte-swapped the
 961          * incoming request message body, but I am responsible for the
 962          * message buffers. */
 963
 964         if (req->rq_export == NULL) {
 965                 struct ldlm_request *dlm_req;
 966
 967                 CDEBUG(D_RPCTRACE, "operation %d from nid %s with bad "
 968                        "export cookie "LPX64" (ptl req %d/rep %d); this is "
 969                        "normal if this node rebooted with a lock held\n",
 970                        req->rq_reqmsg->opc,
 971                        ptlrpc_peernid2str(&req->rq_peer, str),
 972                        req->rq_reqmsg->handle.cookie,
 973                        req->rq_request_portal, req->rq_reply_portal);
 974
 975                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 976                                              lustre_swab_ldlm_request);
 977                 if (dlm_req != NULL)
 978                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
 979                                dlm_req->lock_handle1.cookie);
 980
 981                 ldlm_callback_reply(req, -ENOTCONN);
 982                 RETURN(0);
 983         }
 984
 985         LASSERT(req->rq_export != NULL);
 986         LASSERT(req->rq_export->exp_obd != NULL);
 987
 988         switch(req->rq_reqmsg->opc) {
 989         case LDLM_BL_CALLBACK:
 990                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 991                 break;
 992         case LDLM_CP_CALLBACK:
 993                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
 994                 break;
 995         case LDLM_GL_CALLBACK:
 996                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
 997                 break;
 998         case OBD_LOG_CANCEL:
 999                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1000                 rc = llog_origin_handle_cancel(req);
1001                 ldlm_callback_reply(req, rc);
1002                 RETURN(0);
1003         case LLOG_ORIGIN_HANDLE_CREATE:
1004                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1005                 rc = llog_origin_handle_create(req);
1006                 ldlm_callback_reply(req, rc);
1007                 RETURN(0);
1008         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1009                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1010                 rc = llog_origin_handle_next_block(req);
1011                 ldlm_callback_reply(req, rc);
1012                 RETURN(0);
1013         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1014                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1015                 rc = llog_origin_handle_read_header(req);
1016                 ldlm_callback_reply(req, rc);
1017                 RETURN(0);
1018         case LLOG_ORIGIN_HANDLE_CLOSE:
1019                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1020                 rc = llog_origin_handle_close(req);
1021                 ldlm_callback_reply(req, rc);
1022                 RETURN(0);
1023         default:
1024                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1025                 ldlm_callback_reply(req, -EPROTO);
1026                 RETURN(0);
1027         }
1028
1029         ns = req->rq_export->exp_obd->obd_namespace;
1030         LASSERT(ns != NULL);
1031
1032         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1033                                       lustre_swab_ldlm_request);
1034         if (dlm_req == NULL) {
1035                 CERROR ("can't unpack dlm_req\n");
1036                 ldlm_callback_reply (req, -EPROTO);
1037                 RETURN (0);
1038         }
1039
1040         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1041         if (!lock) {
1042                 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
1043                        dlm_req->lock_handle1.cookie);
1044                 ldlm_callback_reply(req, -EINVAL);
1045                 RETURN(0);
1046         }
1047
1048         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1049         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1050
1051         /* We want the ost thread to get this reply so that it can respond
1052          * to ost requests (write cache writeback) that might be triggered
1053          * in the callback.
1054          *
1055          * But we'd also like to be able to indicate in the reply that we're
1056          * cancelling right now, because it's unused, or have an intent result
1057          * in the reply, so we might have to push the responsibility for sending
1058          * the reply down into the AST handlers, alas. */
1059
1060         switch (req->rq_reqmsg->opc) {
1061         case LDLM_BL_CALLBACK:
1062                 CDEBUG(D_INODE, "blocking ast\n");
1063 #ifdef __KERNEL__
1064                 rc = ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock);
1065                 ldlm_callback_reply(req, rc);
1066 #else
1067                 rc = 0;
1068                 ldlm_callback_reply(req, rc);
1069                 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1070 #endif
1071                 break;
1072         case LDLM_CP_CALLBACK:
1073                 CDEBUG(D_INODE, "completion ast\n");
1074                 ldlm_callback_reply(req, 0);
1075                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1076                 break;
1077         case LDLM_GL_CALLBACK:
1078                 CDEBUG(D_INODE, "glimpse ast\n");
1079                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1080                 break;
1081         default:
1082                 LBUG();                         /* checked above */
1083         }
1084
1085         RETURN(0);
1086 }
1087
1088 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1089 {
1090         int rc;
1091         ENTRY;
1092
1093         /* Requests arrive in sender's byte order.  The ptlrpc service
1094          * handler has already checked and, if necessary, byte-swapped the
1095          * incoming request message body, but I am responsible for the
1096          * message buffers. */
1097
1098         if (req->rq_export == NULL) {
1099                 struct ldlm_request *dlm_req;
1100                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
1101                        req->rq_reqmsg->opc, req->rq_request_portal,
1102                        req->rq_reply_portal);
1103                 CERROR("--> export cookie: "LPX64"\n",
1104                        req->rq_reqmsg->handle.cookie);
1105                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1106                                              lustre_swab_ldlm_request);
1107                 if (dlm_req != NULL)
1108                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1109                 RETURN(-ENOTCONN);
1110         }
1111
1112         switch (req->rq_reqmsg->opc) {
1113
1114         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1115         case LDLM_CANCEL:
1116                 CDEBUG(D_INODE, "cancel\n");
1117                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1118                 rc = ldlm_handle_cancel(req);
1119                 if (rc)
1120                         break;
1121                 RETURN(0);
1122
1123         default:
1124                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1125                 RETURN(-EINVAL);
1126         }
1127
1128         RETURN(0);
1129 }
1130
1131 #ifdef __KERNEL__
1132 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1133 {
1134         struct ldlm_bl_work_item *blwi = NULL;
1135
1136         spin_lock(&blp->blp_lock);
1137         if (!list_empty(&blp->blp_list)) {
1138                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1139                                   blwi_entry);
1140                 list_del(&blwi->blwi_entry);
1141         }
1142         spin_unlock(&blp->blp_lock);
1143
1144         return blwi;
1145 }
1146
1147 struct ldlm_bl_thread_data {
1148         int                     bltd_num;
1149         struct ldlm_bl_pool     *bltd_blp;
1150 };
1151
1152 static int ldlm_bl_thread_main(void *arg)
1153 {
1154         struct ldlm_bl_thread_data *bltd = arg;
1155         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1156         unsigned long flags;
1157         ENTRY;
1158
1159         /* XXX boiler-plate */
1160         {
1161                 char name[sizeof(current->comm)];
1162                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1163                          bltd->bltd_num);
1164                 kportal_daemonize(name);
1165         }
1166         SIGNAL_MASK_LOCK(current, flags);
1167         sigfillset(&current->blocked);
1168         RECALC_SIGPENDING;
1169         SIGNAL_MASK_UNLOCK(current, flags);
1170
1171         atomic_inc(&blp->blp_num_threads);
1172         complete(&blp->blp_comp);
1173
1174         while(1) {
1175                 struct l_wait_info lwi = { 0 };
1176                 struct ldlm_bl_work_item *blwi = NULL;
1177
1178                 l_wait_event_exclusive(blp->blp_waitq,
1179                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1180                                        &lwi);
1181
1182                 if (blwi->blwi_ns == NULL)
1183                         break;
1184
1185                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1186                                         blwi->blwi_lock);
1187                 OBD_FREE(blwi, sizeof(*blwi));
1188         }
1189
1190         atomic_dec(&blp->blp_num_threads);
1191         complete(&blp->blp_comp);
1192         RETURN(0);
1193 }
1194
1195 #endif
1196
1197 static int ldlm_setup(void);
1198 static int ldlm_cleanup(int force);
1199
1200 int ldlm_get_ref(void)
1201 {
1202         int rc = 0;
1203         down(&ldlm_ref_sem);
1204         if (++ldlm_refcount == 1) {
1205                 rc = ldlm_setup();
1206                 if (rc)
1207                         ldlm_refcount--;
1208         }
1209         up(&ldlm_ref_sem);
1210
1211         RETURN(rc);
1212 }
1213
1214 void ldlm_put_ref(int force)
1215 {
1216         down(&ldlm_ref_sem);
1217         if (ldlm_refcount == 1) {
1218                 int rc = ldlm_cleanup(force);
1219                 if (rc)
1220                         CERROR("ldlm_cleanup failed: %d\n", rc);
1221                 else
1222                         ldlm_refcount--;
1223         } else {
1224                 ldlm_refcount--;
1225         }
1226         up(&ldlm_ref_sem);
1227
1228         EXIT;
1229 }
1230
1231 static int ldlm_setup(void)
1232 {
1233         struct ldlm_bl_pool *blp;
1234         int rc = 0;
1235 #ifdef __KERNEL__
1236         int i;
1237 #endif
1238         ENTRY;
1239
1240         if (ldlm_state != NULL)
1241                 RETURN(-EALREADY);
1242
1243         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1244         if (ldlm_state == NULL)
1245                 RETURN(-ENOMEM);
1246
1247 #ifdef __KERNEL__
1248         rc = ldlm_proc_setup();
1249         if (rc != 0)
1250                 GOTO(out_free, rc);
1251 #endif
1252
1253         ldlm_state->ldlm_cb_service =
1254                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1255                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1256                                 ldlm_callback_handler, "ldlm_cbd",
1257                                 ldlm_svc_proc_dir);
1258
1259         if (!ldlm_state->ldlm_cb_service) {
1260                 CERROR("failed to start service\n");
1261                 GOTO(out_proc, rc = -ENOMEM);
1262         }
1263
1264         ldlm_state->ldlm_cancel_service =
1265                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1266                                 LDLM_CANCEL_REQUEST_PORTAL,
1267                                 LDLM_CANCEL_REPLY_PORTAL,
1268                                 ldlm_cancel_handler, "ldlm_canceld",
1269                                 ldlm_svc_proc_dir);
1270
1271         if (!ldlm_state->ldlm_cancel_service) {
1272                 CERROR("failed to start service\n");
1273                 GOTO(out_proc, rc = -ENOMEM);
1274         }
1275
1276         OBD_ALLOC(blp, sizeof(*blp));
1277         if (blp == NULL)
1278                 GOTO(out_proc, rc = -ENOMEM);
1279         ldlm_state->ldlm_bl_pool = blp;
1280
1281         atomic_set(&blp->blp_num_threads, 0);
1282         init_waitqueue_head(&blp->blp_waitq);
1283         spin_lock_init(&blp->blp_lock);
1284
1285         INIT_LIST_HEAD(&blp->blp_list);
1286
1287 #ifdef __KERNEL__
1288         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1289                 struct ldlm_bl_thread_data bltd = {
1290                         .bltd_num = i,
1291                         .bltd_blp = blp,
1292                 };
1293                 init_completion(&blp->blp_comp);
1294                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1295                 if (rc < 0) {
1296                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1297                         LBUG();
1298                         GOTO(out_thread, rc);
1299                 }
1300                 wait_for_completion(&blp->blp_comp);
1301         }
1302
1303         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1304                                     LDLM_NUM_THREADS, "ldlm_cn");
1305         if (rc) {
1306                 LBUG();
1307                 GOTO(out_thread, rc);
1308         }
1309
1310         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1311                                     LDLM_NUM_THREADS, "ldlm_cb");
1312         if (rc) {
1313                 LBUG();
1314                 GOTO(out_thread, rc);
1315         }
1316
1317         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1318         spin_lock_init(&expired_lock_thread.elt_lock);
1319         expired_lock_thread.elt_state = ELT_STOPPED;
1320         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1321
1322         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1323         if (rc < 0) {
1324                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1325                 GOTO(out_thread, rc);
1326         }
1327
1328         wait_event(expired_lock_thread.elt_waitq,
1329                    expired_lock_thread.elt_state == ELT_READY);
1330
1331         INIT_LIST_HEAD(&waiting_locks_list);
1332         spin_lock_init(&waiting_locks_spinlock);
1333         waiting_locks_timer.function = waiting_locks_callback;
1334         waiting_locks_timer.data = 0;
1335         init_timer(&waiting_locks_timer);
1336 #endif
1337
1338         RETURN(0);
1339
1340 #ifdef __KERNEL__
1341  out_thread:
1342         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1343         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1344 #endif
1345
1346  out_proc:
1347 #ifdef __KERNEL__
1348         ldlm_proc_cleanup();
1349  out_free:
1350 #endif
1351         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1352         ldlm_state = NULL;
1353         return rc;
1354 }
1355
1356 static int ldlm_cleanup(int force)
1357 {
1358 #ifdef __KERNEL__
1359         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1360 #endif
1361         ENTRY;
1362
1363         if (!list_empty(&ldlm_namespace_list)) {
1364                 CERROR("ldlm still has namespaces; clean these up first.\n");
1365                 ldlm_dump_all_namespaces();
1366                 RETURN(-EBUSY);
1367         }
1368
1369 #ifdef __KERNEL__
1370         while (atomic_read(&blp->blp_num_threads) > 0) {
1371                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1372
1373                 init_completion(&blp->blp_comp);
1374
1375                 spin_lock(&blp->blp_lock);
1376                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1377                 wake_up(&blp->blp_waitq);
1378                 spin_unlock(&blp->blp_lock);
1379
1380                 wait_for_completion(&blp->blp_comp);
1381         }
1382         OBD_FREE(blp, sizeof(*blp));
1383
1384         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1385         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1386         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1387         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1388         ldlm_proc_cleanup();
1389
1390         expired_lock_thread.elt_state = ELT_TERMINATE;
1391         wake_up(&expired_lock_thread.elt_waitq);
1392         wait_event(expired_lock_thread.elt_waitq,
1393                    expired_lock_thread.elt_state == ELT_STOPPED);
1394
1395 #endif
1396
1397         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1398         ldlm_state = NULL;
1399
1400         RETURN(0);
1401 }
1402
1403 int __init ldlm_init(void)
1404 {
1405         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1406                                                sizeof(struct ldlm_resource), 0,
1407                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1408         if (ldlm_resource_slab == NULL)
1409                 return -ENOMEM;
1410
1411         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1412                                            sizeof(struct ldlm_lock), 0,
1413                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1414         if (ldlm_lock_slab == NULL) {
1415                 kmem_cache_destroy(ldlm_resource_slab);
1416                 return -ENOMEM;
1417         }
1418
1419         l_lock_init(&ldlm_handle_lock);
1420
1421         return 0;
1422 }
1423
1424 void __exit ldlm_exit(void)
1425 {
1426         if ( ldlm_refcount )
1427                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1428         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
1429                 CERROR("couldn't free ldlm resource slab\n");
1430         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
1431                 CERROR("couldn't free ldlm lock slab\n");
1432 }
1433
1434 /* ldlm_flock.c */
1435 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1436
1437 /* ldlm_extent.c */
1438 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1439
1440 /* ldlm_lock.c */
1441 EXPORT_SYMBOL(ldlm_get_processing_policy);
1442 EXPORT_SYMBOL(ldlm_lock2desc);
1443 EXPORT_SYMBOL(ldlm_register_intent);
1444 EXPORT_SYMBOL(ldlm_lockname);
1445 EXPORT_SYMBOL(ldlm_typename);
1446 EXPORT_SYMBOL(ldlm_lock2handle);
1447 EXPORT_SYMBOL(__ldlm_handle2lock);
1448 EXPORT_SYMBOL(ldlm_lock_get);
1449 EXPORT_SYMBOL(ldlm_lock_put);
1450 EXPORT_SYMBOL(ldlm_lock_match);
1451 EXPORT_SYMBOL(ldlm_lock_cancel);
1452 EXPORT_SYMBOL(ldlm_lock_addref);
1453 EXPORT_SYMBOL(ldlm_lock_decref);
1454 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1455 EXPORT_SYMBOL(ldlm_lock_change_resource);
1456 EXPORT_SYMBOL(ldlm_lock_set_data);
1457 EXPORT_SYMBOL(ldlm_it2str);
1458 EXPORT_SYMBOL(ldlm_lock_dump);
1459 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1460 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1461 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1462 EXPORT_SYMBOL(ldlm_lock_allow_match);
1463
1464 /* ldlm_request.c */
1465 EXPORT_SYMBOL(ldlm_completion_ast);
1466 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1467 EXPORT_SYMBOL(ldlm_cli_convert);
1468 EXPORT_SYMBOL(ldlm_cli_enqueue);
1469 EXPORT_SYMBOL(ldlm_cli_cancel);
1470 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1471 EXPORT_SYMBOL(ldlm_replay_locks);
1472 EXPORT_SYMBOL(ldlm_resource_foreach);
1473 EXPORT_SYMBOL(ldlm_namespace_foreach);
1474 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1475 EXPORT_SYMBOL(ldlm_change_cbdata);
1476
1477 /* ldlm_lockd.c */
1478 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1479 EXPORT_SYMBOL(ldlm_server_completion_ast);
1480 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1481 EXPORT_SYMBOL(ldlm_handle_enqueue);
1482 EXPORT_SYMBOL(ldlm_handle_cancel);
1483 EXPORT_SYMBOL(ldlm_handle_convert);
1484 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1485 EXPORT_SYMBOL(ldlm_get_ref);
1486 EXPORT_SYMBOL(ldlm_put_ref);
1487
1488 #if 0
1489 /* ldlm_test.c */
1490 EXPORT_SYMBOL(ldlm_test);
1491 EXPORT_SYMBOL(ldlm_regression_start);
1492 EXPORT_SYMBOL(ldlm_regression_stop);
1493 #endif
1494
1495 /* ldlm_resource.c */
1496 EXPORT_SYMBOL(ldlm_namespace_new);
1497 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1498 EXPORT_SYMBOL(ldlm_namespace_free);
1499 EXPORT_SYMBOL(ldlm_namespace_dump);
1500 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1501 EXPORT_SYMBOL(ldlm_resource_get);
1502 EXPORT_SYMBOL(ldlm_resource_putref);
1503
1504 /* l_lock.c */
1505 EXPORT_SYMBOL(l_lock);
1506 EXPORT_SYMBOL(l_unlock);
1507
1508 /* ldlm_lib.c */
1509 EXPORT_SYMBOL(client_obd_setup);
1510 EXPORT_SYMBOL(client_obd_cleanup);
1511 EXPORT_SYMBOL(client_connect_import);
1512 EXPORT_SYMBOL(client_disconnect_export);
1513 EXPORT_SYMBOL(target_abort_recovery);
1514 EXPORT_SYMBOL(target_handle_connect);
1515 EXPORT_SYMBOL(target_destroy_export);
1516 EXPORT_SYMBOL(target_cancel_recovery_timer);
1517 EXPORT_SYMBOL(target_send_reply);
1518 EXPORT_SYMBOL(target_queue_recovery_request);
1519 EXPORT_SYMBOL(target_handle_ping);
1520 EXPORT_SYMBOL(target_handle_disconnect);
1521 EXPORT_SYMBOL(target_queue_final_reply);