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
  48 static DECLARE_MUTEX(ldlm_ref_sem);
  49 static int ldlm_refcount;
  50 /* LDLM state */
  51
  52 static struct ldlm_state *ldlm_state;
  53
  54 inline unsigned long round_timeout(unsigned long timeout)
  55 {
  56         return ((timeout / HZ) + 1) * HZ;
  57 }
  58
  59 #ifdef __KERNEL__
  60 /* XXX should this be per-ldlm? */
  61 static struct list_head waiting_locks_list;
  62 static spinlock_t waiting_locks_spinlock;
  63 static struct timer_list waiting_locks_timer;
  64
  65 static struct expired_lock_thread {
  66         wait_queue_head_t         elt_waitq;
  67         int                       elt_state;
  68         struct list_head          elt_expired_locks;
  69         spinlock_t                elt_lock;
  70 } expired_lock_thread;
  71 #endif
  72
  73 #if !defined(ENOTSUPP)
  74 #  define ENOTSUPP 524
  75 #endif
  76
  77 #define ELT_STOPPED   0
  78 #define ELT_READY     1
  79 #define ELT_TERMINATE 2
  80
  81 struct ldlm_bl_pool {
  82         spinlock_t              blp_lock;
  83         struct list_head        blp_list;
  84         wait_queue_head_t       blp_waitq;
  85         atomic_t                blp_num_threads;
  86         struct completion       blp_comp;
  87 };
  88
  89 struct ldlm_bl_work_item {
  90         struct list_head        blwi_entry;
  91         struct ldlm_namespace   *blwi_ns;
  92         struct ldlm_lock_desc   blwi_ld;
  93         struct ldlm_lock        *blwi_lock;
  94 };
  95
  96 #ifdef __KERNEL__
  97
  98 static inline int have_expired_locks(void)
  99 {
 100         int need_to_run;
 101
 102         spin_lock_bh(&expired_lock_thread.elt_lock);
 103         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 104         spin_unlock_bh(&expired_lock_thread.elt_lock);
 105
 106         RETURN(need_to_run);
 107 }
 108
 109 static int expired_lock_main(void *arg)
 110 {
 111         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 112         struct l_wait_info lwi = { 0 };
 113         unsigned long flags;
 114
 115         ENTRY;
 116         lock_kernel();
 117         kportal_daemonize("ldlm_elt");
 118
 119         SIGNAL_MASK_LOCK(current, flags);
 120         sigfillset(&current->blocked);
 121         RECALC_SIGPENDING;
 122         SIGNAL_MASK_UNLOCK(current, flags);
 123
 124         unlock_kernel();
 125
 126         expired_lock_thread.elt_state = ELT_READY;
 127         wake_up(&expired_lock_thread.elt_waitq);
 128
 129         while (1) {
 130                 l_wait_event(expired_lock_thread.elt_waitq,
 131                              have_expired_locks() ||
 132                              expired_lock_thread.elt_state == ELT_TERMINATE,
 133                              &lwi);
 134
 135                 spin_lock_bh(&expired_lock_thread.elt_lock);
 136                 while (!list_empty(expired)) {
 137                         struct obd_export *export;
 138                         struct ldlm_lock *lock;
 139
 140                         lock = list_entry(expired->next, struct ldlm_lock,
 141                                           l_pending_chain);
 142                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
 143                             (void *)lock >= LP_POISON) {
 144                                 CERROR("free lock on elt list %p\n", lock);
 145                                 LBUG();
 146                         }
 147                         list_del_init(&lock->l_pending_chain);
 148                         if ((void *)lock->l_export < LP_POISON + PAGE_SIZE &&
 149                             (void *)lock->l_export >= LP_POISON + PAGE_SIZE) {
 150                                 CERROR("lock with free export on elt list %p\n",
 151                                        export);
 152                                 lock->l_export = NULL;
 153                                 LDLM_ERROR(lock, "free export\n");
 154                                 continue;
 155                         }
 156                         export = class_export_get(lock->l_export);
 157                         spin_unlock_bh(&expired_lock_thread.elt_lock);
 158
 159                         ptlrpc_fail_export(export);
 160                         class_export_put(export);
 161                         spin_lock_bh(&expired_lock_thread.elt_lock);
 162                 }
 163                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 164
 165                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 166                         break;
 167         }
 168
 169         expired_lock_thread.elt_state = ELT_STOPPED;
 170         wake_up(&expired_lock_thread.elt_waitq);
 171         RETURN(0);
 172 }
 173
 174 static void waiting_locks_callback(unsigned long unused)
 175 {
 176         struct ldlm_lock *lock;
 177         char str[PTL_NALFMT_SIZE];
 178
 179         spin_lock_bh(&waiting_locks_spinlock);
 180         while (!list_empty(&waiting_locks_list)) {
 181                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 182                                   l_pending_chain);
 183
 184                 if ((lock->l_callback_timeout > jiffies) ||
 185                     (lock->l_req_mode == LCK_GROUP))
 186                         break;
 187
 188                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 189                            "%s@%s nid %s ",
 190                            lock->l_export->exp_client_uuid.uuid,
 191                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 192                            ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str));
 193
 194                 spin_lock_bh(&expired_lock_thread.elt_lock);
 195                 list_del(&lock->l_pending_chain);
 196                 list_add(&lock->l_pending_chain,
 197                          &expired_lock_thread.elt_expired_locks);
 198                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 199                 wake_up(&expired_lock_thread.elt_waitq);
 200         }
 201
 202         /*
 203          * Make sure the timer will fire again if we have any locks
 204          * left.
 205          */
 206         if (!list_empty(&waiting_locks_list)) {
 207                 unsigned long timeout_rounded;
 208                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 209                                   l_pending_chain);
 210                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 211                 mod_timer(&waiting_locks_timer, timeout_rounded);
 212         }
 213         spin_unlock_bh(&waiting_locks_spinlock);
 214 }
 215
 216 /*
 217  * Indicate that we're waiting for a client to call us back cancelling a given
 218  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 219  * timer to fire appropriately.  (We round up to the next second, to avoid
 220  * floods of timer firings during periods of high lock contention and traffic).
 221  */
 222 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 223 {
 224         unsigned long timeout_rounded;
 225
 226         spin_lock_bh(&waiting_locks_spinlock);
 227         if (!list_empty(&lock->l_pending_chain)) {
 228                 LDLM_DEBUG(lock, "not re-adding to wait list");
 229                 spin_unlock_bh(&waiting_locks_spinlock);
 230                 return 0;
 231         }
 232         LDLM_DEBUG(lock, "adding to wait list");
 233
 234         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 235
 236         timeout_rounded = round_timeout(lock->l_callback_timeout);
 237
 238         if (timeout_rounded < waiting_locks_timer.expires ||
 239             !timer_pending(&waiting_locks_timer)) {
 240                 mod_timer(&waiting_locks_timer, timeout_rounded);
 241         }
 242         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 243         spin_unlock_bh(&waiting_locks_spinlock);
 244         return 1;
 245 }
 246
 247 /*
 248  * Remove a lock from the pending list, likely because it had its cancellation
 249  * callback arrive without incident.  This adjusts the lock-timeout timer if
 250  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 251  */
 252 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 253 {
 254         struct list_head *list_next;
 255
 256         if (lock->l_export == NULL) {
 257                 /* We don't have a "waiting locks list" on clients. */
 258                 LDLM_DEBUG(lock, "client lock: no-op");
 259                 return 0;
 260         }
 261
 262         spin_lock_bh(&waiting_locks_spinlock);
 263
 264         if (list_empty(&lock->l_pending_chain)) {
 265                 spin_unlock_bh(&waiting_locks_spinlock);
 266                 LDLM_DEBUG(lock, "wasn't waiting");
 267                 return 0;
 268         }
 269
 270         list_next = lock->l_pending_chain.next;
 271         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 272                 /* Removing the head of the list, adjust timer. */
 273                 if (list_next == &waiting_locks_list) {
 274                         /* No more, just cancel. */
 275                         del_timer(&waiting_locks_timer);
 276                 } else {
 277                         struct ldlm_lock *next;
 278                         next = list_entry(list_next, struct ldlm_lock,
 279                                           l_pending_chain);
 280                         mod_timer(&waiting_locks_timer,
 281                                   round_timeout(next->l_callback_timeout));
 282                 }
 283         }
 284
 285         spin_lock_bh(&expired_lock_thread.elt_lock);
 286         list_del_init(&lock->l_pending_chain);
 287         spin_unlock_bh(&expired_lock_thread.elt_lock);
 288
 289         spin_unlock_bh(&waiting_locks_spinlock);
 290         LDLM_DEBUG(lock, "removed");
 291         return 1;
 292 }
 293
 294 #else /* !__KERNEL__ */
 295
 296 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 297 {
 298         RETURN(1);
 299 }
 300
 301 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 302 {
 303         RETURN(0);
 304 }
 305
 306 #endif /* __KERNEL__ */
 307
 308 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,const char *ast_type)
 309 {
 310         struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 311         char str[PTL_NALFMT_SIZE];
 312
 313         LDLM_ERROR(lock, "%s AST failed (%d): evicting client %s@%s NID "LPX64
 314                    " (%s)", ast_type, rc, lock->l_export->exp_client_uuid.uuid,
 315                    conn->c_remote_uuid.uuid, conn->c_peer.peer_id.nid,
 316                    ptlrpc_peernid2str(&conn->c_peer, str));
 317         ptlrpc_fail_export(lock->l_export);
 318 }
 319
 320 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 321                                  struct ptlrpc_request *req, int rc,
 322                                  const char *ast_type)
 323 {
 324         struct ptlrpc_peer *peer = &req->rq_import->imp_connection->c_peer;
 325         char str[PTL_NALFMT_SIZE];
 326
 327         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 328                 LASSERT(lock->l_export);
 329                 if (lock->l_export->exp_libclient) {
 330                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 331                                    " timeout, just cancelling lock", ast_type,
 332                                    ptlrpc_peernid2str(peer, str));
 333                         ldlm_lock_cancel(lock);
 334                         rc = -ERESTART;
 335                 } else {
 336                         ldlm_del_waiting_lock(lock);
 337                         ldlm_failed_ast(lock, rc, ast_type);
 338                 }
 339         } else if (rc) {
 340                 if (rc == -EINVAL)
 341                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 342                                    " from %s AST - normal race",
 343                                    ptlrpc_peernid2str(peer, str),
 344                                    req->rq_repmsg->status, ast_type);
 345                 else
 346                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 347                                    "from %s AST", ptlrpc_peernid2str(peer, str),
 348                                    (req->rq_repmsg != NULL) ?
 349                                    req->rq_repmsg->status : 0, ast_type);
 350                 ldlm_lock_cancel(lock);
 351                 /* Server-side AST functions are called from ldlm_reprocess_all,
 352                  * which needs to be told to please restart its reprocessing. */
 353                 rc = -ERESTART;
 354         }
 355
 356         return rc;
 357 }
 358
 359 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 360                              struct ldlm_lock_desc *desc,
 361                              void *data, int flag)
 362 {
 363         struct ldlm_request *body;
 364         struct ptlrpc_request *req;
 365         int rc = 0, size = sizeof(*body);
 366         ENTRY;
 367
 368         if (flag == LDLM_CB_CANCELING) {
 369                 /* Don't need to do anything here. */
 370                 RETURN(0);
 371         }
 372
 373         LASSERT(lock);
 374
 375         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 376         if (lock->l_granted_mode != lock->l_req_mode) {
 377                 /* this blocking AST will be communicated as part of the
 378                  * completion AST instead */
 379                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 380                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 381                 RETURN(0);
 382         }
 383
 384         if (lock->l_destroyed) {
 385                 /* What's the point? */
 386                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 387                 RETURN(0);
 388         }
 389
 390 #if 0
 391         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 392                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 393                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 394                 RETURN(-ETIMEDOUT);
 395         }
 396 #endif
 397
 398         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 399                               LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK,
 400                               1, &size, NULL);
 401         if (req == NULL) {
 402                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 403                 RETURN(-ENOMEM);
 404         }
 405
 406         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 407         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 408                sizeof(body->lock_handle1));
 409         memcpy(&body->lock_desc, desc, sizeof(*desc));
 410         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 411
 412         LDLM_DEBUG(lock, "server preparing blocking AST");
 413         req->rq_replen = lustre_msg_size(0, NULL);
 414
 415         if (lock->l_granted_mode == lock->l_req_mode)
 416                 ldlm_add_waiting_lock(lock);
 417         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 418
 419         req->rq_send_state = LUSTRE_IMP_FULL;
 420         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 421         rc = ptlrpc_queue_wait(req);
 422         if (rc != 0)
 423                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 424
 425         ptlrpc_req_finished(req);
 426
 427         RETURN(rc);
 428 }
 429
 430 /* XXX copied from ptlrpc/service.c */
 431 static long timeval_sub(struct timeval *large, struct timeval *small)
 432 {
 433         return (large->tv_sec - small->tv_sec) * 1000000 +
 434                 (large->tv_usec - small->tv_usec);
 435 }
 436
 437 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 438 {
 439         struct ldlm_request *body;
 440         struct ptlrpc_request *req;
 441         struct timeval granted_time;
 442         long total_enqueue_wait;
 443         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 444         ENTRY;
 445
 446         LASSERT(lock != NULL);
 447
 448         do_gettimeofday(&granted_time);
 449         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 450
 451         if (total_enqueue_wait / 1000000 > obd_timeout)
 452                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 453
 454         down(&lock->l_resource->lr_lvb_sem);
 455         if (lock->l_resource->lr_lvb_len) {
 456                 buffers = 2;
 457                 size[1] = lock->l_resource->lr_lvb_len;
 458         }
 459         up(&lock->l_resource->lr_lvb_sem);
 460
 461         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 462                               LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK,
 463                               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;
 475
 476                 down(&lock->l_resource->lr_lvb_sem);
 477                 lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 478                                      lock->l_resource->lr_lvb_len);
 479
 480                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 481                        lock->l_resource->lr_lvb_len);
 482                 up(&lock->l_resource->lr_lvb_sem);
 483         }
 484
 485         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 486                    total_enqueue_wait);
 487         req->rq_replen = lustre_msg_size(0, NULL);
 488
 489         req->rq_send_state = LUSTRE_IMP_FULL;
 490         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 491
 492         /* We only send real blocking ASTs after the lock is granted */
 493         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 494         if (lock->l_flags & LDLM_FL_AST_SENT) {
 495                 body->lock_flags |= LDLM_FL_AST_SENT;
 496                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 497         }
 498         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 499
 500         rc = ptlrpc_queue_wait(req);
 501         if (rc != 0)
 502                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 503
 504         ptlrpc_req_finished(req);
 505
 506         RETURN(rc);
 507 }
 508
 509 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 510 {
 511         struct ldlm_resource *res = lock->l_resource;
 512         struct ldlm_request *body;
 513         struct ptlrpc_request *req;
 514         int rc = 0, size = sizeof(*body);
 515         ENTRY;
 516
 517         LASSERT(lock != NULL);
 518
 519         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 520                               LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK,
 521                               1, &size, NULL);
 522         if (req == NULL)
 523                 RETURN(-ENOMEM);
 524
 525         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 526         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 527                sizeof(body->lock_handle1));
 528         ldlm_lock2desc(lock, &body->lock_desc);
 529
 530         size = lock->l_resource->lr_lvb_len;
 531         req->rq_replen = lustre_msg_size(1, &size);
 532
 533         req->rq_send_state = LUSTRE_IMP_FULL;
 534         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 535
 536         rc = ptlrpc_queue_wait(req);
 537         if (rc == -ELDLM_NO_LOCK_DATA)
 538                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 539         else if (rc != 0)
 540                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 541         else
 542                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 543                         (res, req->rq_repmsg, 0, 1);
 544         ptlrpc_req_finished(req);
 545         RETURN(rc);
 546 }
 547
 548 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 549                         ldlm_completion_callback completion_callback,
 550                         ldlm_blocking_callback blocking_callback,
 551                         ldlm_glimpse_callback glimpse_callback)
 552 {
 553         struct obd_device *obddev = req->rq_export->exp_obd;
 554         struct ldlm_reply *dlm_rep;
 555         struct ldlm_request *dlm_req;
 556         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 557         __u32 flags;
 558         ldlm_error_t err = ELDLM_OK;
 559         struct ldlm_lock *lock = NULL;
 560         void *cookie = NULL;
 561         ENTRY;
 562
 563         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 564
 565         dlm_req = lustre_swab_reqbuf (req, MDS_REQ_INTENT_LOCKREQ_OFF,
 566                                       sizeof (*dlm_req),
 567                                       lustre_swab_ldlm_request);
 568         if (dlm_req == NULL) {
 569                 CERROR ("Can't unpack dlm_req\n");
 570                 GOTO(out, rc = -EFAULT);
 571         }
 572
 573         flags = dlm_req->lock_flags;
 574
 575         /* The lock's callback data might be set in the policy function */
 576         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 577                                 dlm_req->lock_desc.l_resource.lr_name,
 578                                 dlm_req->lock_desc.l_resource.lr_type,
 579                                 dlm_req->lock_desc.l_req_mode,
 580                                 blocking_callback, completion_callback,
 581                                 glimpse_callback, NULL, 0);
 582         if (!lock)
 583                 GOTO(out, rc = -ENOMEM);
 584
 585         do_gettimeofday(&lock->l_enqueued_time);
 586         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 587                sizeof(lock->l_remote_handle));
 588         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 589
 590         LASSERT(req->rq_export);
 591         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
 592         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 593         if (req->rq_export->exp_failed) {
 594                 LDLM_ERROR(lock,"lock on destroyed export %p\n",req->rq_export);
 595                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 596                 GOTO(out, err = -ENOTCONN);
 597         }
 598         lock->l_export = class_export_get(req->rq_export);
 599
 600         list_add(&lock->l_export_chain,
 601                  &lock->l_export->exp_ldlm_data.led_held_locks);
 602         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 603
 604         if (flags & LDLM_FL_HAS_INTENT) {
 605                 /* In this case, the reply buffer is allocated deep in
 606                  * local_lock_enqueue by the policy function. */
 607                 cookie = req;
 608         } else {
 609                 int buffers = 1;
 610                 down(&lock->l_resource->lr_lvb_sem);
 611                 if (lock->l_resource->lr_lvb_len) {
 612                         size[1] = lock->l_resource->lr_lvb_len;
 613                         buffers = 2;
 614                 }
 615                 up(&lock->l_resource->lr_lvb_sem);
 616                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 617                         GOTO(out, rc = -ENOMEM);
 618
 619                 rc = lustre_pack_reply(req, buffers, size, NULL);
 620                 if (rc)
 621                         GOTO(out, rc);
 622         }
 623
 624         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 625                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 626                        sizeof(ldlm_policy_data_t));
 627         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 628                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 629                        sizeof(lock->l_req_extent));
 630
 631         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 632         if (err)
 633                 GOTO(out, err);
 634
 635         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 636         dlm_rep->lock_flags = flags;
 637
 638         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 639         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 640
 641         /* We never send a blocking AST until the lock is granted, but
 642          * we can tell it right now */
 643         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 644         if (lock->l_flags & LDLM_FL_AST_SENT) {
 645                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 646                 if (lock->l_granted_mode == lock->l_req_mode)
 647                         ldlm_add_waiting_lock(lock);
 648         }
 649         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 650
 651         EXIT;
 652  out:
 653         req->rq_status = err;
 654         if (req->rq_reply_state == NULL) {
 655                 err = lustre_pack_reply(req, 0, NULL, NULL);
 656                 if (rc == 0)
 657                         rc = err;
 658                 req->rq_status = rc;
 659         }
 660
 661         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 662          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 663         if (lock) {
 664                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 665                            "(err=%d, rc=%d)", err, rc);
 666
 667                 if (rc == 0) {
 668                         down(&lock->l_resource->lr_lvb_sem);
 669                         size[1] = lock->l_resource->lr_lvb_len;
 670                         if (size[1] > 0) {
 671                                 void *lvb = lustre_msg_buf(req->rq_repmsg,
 672                                                            1, size[1]);
 673                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
 674                                          req, lock);
 675
 676                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 677                                        size[1]);
 678                         }
 679                         up(&lock->l_resource->lr_lvb_sem);
 680                 } else {
 681                         ldlm_lock_destroy(lock);
 682
 683                 }
 684
 685                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 686                         ldlm_reprocess_all(lock->l_resource);
 687                 LDLM_LOCK_PUT(lock);
 688         }
 689         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 690                           lock, rc);
 691
 692         return rc;
 693 }
 694
 695 int ldlm_handle_convert(struct ptlrpc_request *req)
 696 {
 697         struct ldlm_request *dlm_req;
 698         struct ldlm_reply *dlm_rep;
 699         struct ldlm_lock *lock;
 700         int rc, size = sizeof(*dlm_rep);
 701         ENTRY;
 702
 703         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 704                                       lustre_swab_ldlm_request);
 705         if (dlm_req == NULL) {
 706                 CERROR ("Can't unpack dlm_req\n");
 707                 RETURN (-EFAULT);
 708         }
 709
 710         rc = lustre_pack_reply(req, 1, &size, NULL);
 711         if (rc) {
 712                 CERROR("out of memory\n");
 713                 RETURN(-ENOMEM);
 714         }
 715         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 716         dlm_rep->lock_flags = dlm_req->lock_flags;
 717
 718         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 719         if (!lock) {
 720                 req->rq_status = EINVAL;
 721         } else {
 722                 LDLM_DEBUG(lock, "server-side convert handler START");
 723                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 724                                   &dlm_rep->lock_flags);
 725                 if (ldlm_del_waiting_lock(lock))
 726                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 727                 req->rq_status = 0;
 728         }
 729
 730         if (lock) {
 731                 ldlm_reprocess_all(lock->l_resource);
 732                 LDLM_DEBUG(lock, "server-side convert handler END");
 733                 LDLM_LOCK_PUT(lock);
 734         } else
 735                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 736
 737         RETURN(0);
 738 }
 739
 740 int ldlm_handle_cancel(struct ptlrpc_request *req)
 741 {
 742         struct ldlm_request *dlm_req;
 743         struct ldlm_lock *lock;
 744         struct ldlm_resource *res;
 745         char str[PTL_NALFMT_SIZE];
 746         int rc;
 747         ENTRY;
 748
 749         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 750                                       lustre_swab_ldlm_request);
 751         if (dlm_req == NULL) {
 752                 CERROR("bad request buffer for cancel\n");
 753                 RETURN(-EFAULT);
 754         }
 755
 756         rc = lustre_pack_reply(req, 0, NULL, NULL);
 757         if (rc) {
 758                 CERROR("out of memory\n");
 759                 RETURN(-ENOMEM);
 760         }
 761
 762         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 763         if (!lock) {
 764                 CERROR("received cancel for unknown lock cookie "LPX64
 765                        " from client %s nid %s\n",
 766                        dlm_req->lock_handle1.cookie,
 767                        req->rq_export->exp_client_uuid.uuid,
 768                        ptlrpc_peernid2str(&req->rq_peer, str));
 769                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 770                                   "(cookie "LPU64")",
 771                                   dlm_req->lock_handle1.cookie);
 772                 req->rq_status = ESTALE;
 773         } else {
 774                 LDLM_DEBUG(lock, "server-side cancel handler START");
 775                 res = lock->l_resource;
 776                 if (res && res->lr_namespace->ns_lvbo &&
 777                     res->lr_namespace->ns_lvbo->lvbo_update) {
 778                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 779                                 (res, NULL, 0, 0);
 780                                 //(res, req->rq_reqmsg, 1);
 781                 }
 782
 783                 ldlm_lock_cancel(lock);
 784                 if (ldlm_del_waiting_lock(lock))
 785                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 786                 req->rq_status = rc;
 787         }
 788
 789         if (ptlrpc_reply(req) != 0)
 790                 LBUG();
 791
 792         if (lock) {
 793                 ldlm_reprocess_all(lock->l_resource);
 794                 LDLM_DEBUG(lock, "server-side cancel handler END");
 795                 LDLM_LOCK_PUT(lock);
 796         }
 797
 798         RETURN(0);
 799 }
 800
 801 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 802                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 803 {
 804         int do_ast;
 805         ENTRY;
 806
 807         l_lock(&ns->ns_lock);
 808         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 809
 810         lock->l_flags |= LDLM_FL_CBPENDING;
 811         do_ast = (!lock->l_readers && !lock->l_writers);
 812
 813         if (do_ast) {
 814                 LDLM_DEBUG(lock, "already unused, calling "
 815                            "callback (%p)", lock->l_blocking_ast);
 816                 if (lock->l_blocking_ast != NULL) {
 817                         l_unlock(&ns->ns_lock);
 818                         l_check_no_ns_lock(ns);
 819                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 820                                              LDLM_CB_BLOCKING);
 821                         l_lock(&ns->ns_lock);
 822                 }
 823         } else {
 824                 LDLM_DEBUG(lock, "Lock still has references, will be"
 825                            " cancelled later");
 826         }
 827
 828         LDLM_DEBUG(lock, "client blocking callback handler END");
 829         l_unlock(&ns->ns_lock);
 830         LDLM_LOCK_PUT(lock);
 831         EXIT;
 832 }
 833
 834 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 835                                     struct ldlm_namespace *ns,
 836                                     struct ldlm_request *dlm_req,
 837                                     struct ldlm_lock *lock)
 838 {
 839         LIST_HEAD(ast_list);
 840         ENTRY;
 841
 842         l_lock(&ns->ns_lock);
 843         LDLM_DEBUG(lock, "client completion callback handler START");
 844
 845         /* If we receive the completion AST before the actual enqueue returned,
 846          * then we might need to switch lock modes, resources, or extents. */
 847         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 848                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 849                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 850         }
 851
 852         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 853                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 854                        sizeof(lock->l_policy_data));
 855                 LDLM_DEBUG(lock, "completion AST, new policy data");
 856         }
 857
 858         ldlm_resource_unlink_lock(lock);
 859         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 860                    &lock->l_resource->lr_name,
 861                    sizeof(lock->l_resource->lr_name)) != 0) {
 862                 ldlm_lock_change_resource(ns, lock,
 863                                          dlm_req->lock_desc.l_resource.lr_name);
 864                 LDLM_DEBUG(lock, "completion AST, new resource");
 865         }
 866
 867         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 868                 lock->l_flags |= LDLM_FL_CBPENDING;
 869                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 870         }
 871
 872         if (lock->l_lvb_len) {
 873                 void *lvb;
 874                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 875                                          lock->l_lvb_swabber);
 876                 if (lvb == NULL) {
 877                         LDLM_ERROR(lock, "completion AST did not contain "
 878                                    "expected LVB!");
 879                 } else {
 880                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 881                 }
 882         }
 883
 884         lock->l_resource->lr_tmp = &ast_list;
 885         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 886         lock->l_resource->lr_tmp = NULL;
 887         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 888         l_unlock(&ns->ns_lock);
 889         LDLM_LOCK_PUT(lock);
 890
 891         ldlm_run_ast_work(ns, &ast_list);
 892
 893         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 894                           lock);
 895         EXIT;
 896 }
 897
 898 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 899                                     struct ldlm_namespace *ns,
 900                                     struct ldlm_request *dlm_req,
 901                                     struct ldlm_lock *lock)
 902 {
 903         int rc = -ENOSYS;
 904         ENTRY;
 905
 906         l_lock(&ns->ns_lock);
 907         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 908
 909         if (lock->l_glimpse_ast != NULL) {
 910                 l_unlock(&ns->ns_lock);
 911                 l_check_no_ns_lock(ns);
 912                 rc = lock->l_glimpse_ast(lock, req);
 913                 l_lock(&ns->ns_lock);
 914         }
 915
 916         if (req->rq_repmsg != NULL) {
 917                 ptlrpc_reply(req);
 918         } else {
 919                 req->rq_status = rc;
 920                 ptlrpc_error(req);
 921         }
 922
 923         l_unlock(&ns->ns_lock);
 924         if (lock->l_granted_mode == LCK_PW &&
 925             !lock->l_readers && !lock->l_writers &&
 926             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 927                 if (ldlm_bl_to_thread(ns, NULL, lock))
 928                         ldlm_handle_bl_callback(ns, NULL, lock);
 929
 930                 EXIT;
 931                 return;
 932         }
 933         LDLM_LOCK_PUT(lock);
 934         EXIT;
 935 }
 936
 937 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 938 {
 939         req->rq_status = rc;
 940         if (req->rq_reply_state == NULL) {
 941                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 942                 if (rc)
 943                         return rc;
 944         }
 945         return ptlrpc_reply(req);
 946 }
 947
 948 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
 949                       struct ldlm_lock *lock)
 950 {
 951 #ifdef __KERNEL__
 952         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
 953         struct ldlm_bl_work_item *blwi;
 954         ENTRY;
 955
 956         OBD_ALLOC(blwi, sizeof(*blwi));
 957         if (blwi == NULL)
 958                 RETURN(-ENOMEM);
 959
 960         blwi->blwi_ns = ns;
 961         if (ld != NULL)
 962                 blwi->blwi_ld = *ld;
 963         blwi->blwi_lock = lock;
 964
 965         spin_lock(&blp->blp_lock);
 966         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
 967         wake_up(&blp->blp_waitq);
 968         spin_unlock(&blp->blp_lock);
 969
 970         RETURN(0);
 971 #else
 972         RETURN(-ENOSYS);
 973 #endif
 974
 975 }
 976
 977 static int ldlm_msg_check_version(struct lustre_msg *msg)
 978 {
 979         int rc;
 980
 981         switch (msg->opc) {
 982         case LDLM_ENQUEUE:
 983         case LDLM_CONVERT:
 984         case LDLM_CANCEL:
 985         case LDLM_BL_CALLBACK:
 986         case LDLM_CP_CALLBACK:
 987         case LDLM_GL_CALLBACK:
 988                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
 989                 if (rc)
 990                         CERROR("bad opc %u version %08x, expecting %08x\n",
 991                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
 992                 break;
 993         case OBD_LOG_CANCEL:
 994         case LLOG_ORIGIN_HANDLE_OPEN:
 995         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
 996         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
 997         case LLOG_ORIGIN_HANDLE_READ_HEADER:
 998         case LLOG_ORIGIN_HANDLE_CLOSE:
 999         case LLOG_CATINFO:
1000                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1001                 if (rc)
1002                         CERROR("bad opc %u version %08x, expecting %08x\n",
1003                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
1004                 break;
1005         default:
1006                 CERROR("LDLM unknown opcode %d\n", msg->opc);
1007                 rc = -ENOTSUPP;
1008                 break;
1009         }
1010
1011         return rc;
1012 }
1013
1014 static int ldlm_callback_handler(struct ptlrpc_request *req)
1015 {
1016         struct ldlm_namespace *ns;
1017         struct ldlm_request *dlm_req;
1018         struct ldlm_lock *lock;
1019         char str[PTL_NALFMT_SIZE];
1020         int rc;
1021         ENTRY;
1022
1023         rc = ldlm_msg_check_version(req->rq_reqmsg);
1024         if (rc) {
1025                 CERROR("LDLM_CB drop mal-formed request\n");
1026                 RETURN(rc);
1027         }
1028
1029         /* Requests arrive in sender's byte order.  The ptlrpc service
1030          * handler has already checked and, if necessary, byte-swapped the
1031          * incoming request message body, but I am responsible for the
1032          * message buffers. */
1033
1034         if (req->rq_export == NULL) {
1035                 struct ldlm_request *dlm_req;
1036
1037                 CDEBUG(D_RPCTRACE, "operation %d from nid %s with bad "
1038                        "export cookie "LPX64"; this is "
1039                        "normal if this node rebooted with a lock held\n",
1040                        req->rq_reqmsg->opc,
1041                        ptlrpc_peernid2str(&req->rq_peer, str),
1042                        req->rq_reqmsg->handle.cookie);
1043                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1044                                              lustre_swab_ldlm_request);
1045                 if (dlm_req != NULL)
1046                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1047                                dlm_req->lock_handle1.cookie);
1048
1049                 ldlm_callback_reply(req, -ENOTCONN);
1050                 RETURN(0);
1051         }
1052
1053         LASSERT(req->rq_export != NULL);
1054         LASSERT(req->rq_export->exp_obd != NULL);
1055
1056         switch(req->rq_reqmsg->opc) {
1057         case LDLM_BL_CALLBACK:
1058                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1059                 break;
1060         case LDLM_CP_CALLBACK:
1061                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1062                 break;
1063         case LDLM_GL_CALLBACK:
1064                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1065                 break;
1066         case OBD_LOG_CANCEL:
1067                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1068                 rc = llog_origin_handle_cancel(req);
1069                 ldlm_callback_reply(req, rc);
1070                 RETURN(0);
1071         case LLOG_ORIGIN_HANDLE_OPEN:
1072                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1073                 rc = llog_origin_handle_open(req);
1074                 ldlm_callback_reply(req, rc);
1075                 RETURN(0);
1076         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1077                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1078                 rc = llog_origin_handle_next_block(req);
1079                 ldlm_callback_reply(req, rc);
1080                 RETURN(0);
1081         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1082                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1083                 rc = llog_origin_handle_prev_block(req);
1084                 ldlm_callback_reply(req, rc);
1085                 RETURN(0);
1086         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1087                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1088                 rc = llog_origin_handle_read_header(req);
1089                 ldlm_callback_reply(req, rc);
1090                 RETURN(0);
1091         case LLOG_ORIGIN_HANDLE_CLOSE:
1092                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1093                 rc = llog_origin_handle_close(req);
1094                 ldlm_callback_reply(req, rc);
1095                 RETURN(0);
1096         default:
1097                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1098                 ldlm_callback_reply(req, -EPROTO);
1099                 RETURN(0);
1100         }
1101
1102         ns = req->rq_export->exp_obd->obd_namespace;
1103         LASSERT(ns != NULL);
1104
1105         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1106                                       lustre_swab_ldlm_request);
1107         if (dlm_req == NULL) {
1108                 CERROR ("can't unpack dlm_req\n");
1109                 ldlm_callback_reply (req, -EPROTO);
1110                 RETURN (0);
1111         }
1112
1113         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1114         if (!lock) {
1115                 LDLM_DEBUG_NOLOCK("callback on lock "LPX64" - lock "
1116                                   "disappeared\n",dlm_req->lock_handle1.cookie);
1117                 ldlm_callback_reply(req, -EINVAL);
1118                 RETURN(0);
1119         }
1120
1121         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1122         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1123
1124         /* We want the ost thread to get this reply so that it can respond
1125          * to ost requests (write cache writeback) that might be triggered
1126          * in the callback.
1127          *
1128          * But we'd also like to be able to indicate in the reply that we're
1129          * cancelling right now, because it's unused, or have an intent result
1130          * in the reply, so we might have to push the responsibility for sending
1131          * the reply down into the AST handlers, alas. */
1132
1133         switch (req->rq_reqmsg->opc) {
1134         case LDLM_BL_CALLBACK:
1135                 CDEBUG(D_INODE, "blocking ast\n");
1136                 ldlm_callback_reply(req, 0);
1137                 if (ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock))
1138                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1139
1140                 break;
1141         case LDLM_CP_CALLBACK:
1142                 CDEBUG(D_INODE, "completion ast\n");
1143                 ldlm_callback_reply(req, 0);
1144                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1145                 break;
1146         case LDLM_GL_CALLBACK:
1147                 CDEBUG(D_INODE, "glimpse ast\n");
1148                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1149                 break;
1150         default:
1151                 LBUG();                         /* checked above */
1152         }
1153
1154         RETURN(0);
1155 }
1156
1157 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1158 {
1159         int rc = 0;
1160         ENTRY;
1161
1162         rc = ldlm_msg_check_version(req->rq_reqmsg);
1163         if (rc) {
1164                 CERROR("LDLM_CL drop mal-formed request\n");
1165                 RETURN(rc);
1166         }
1167
1168         /* Requests arrive in sender's byte order.  The ptlrpc service
1169          * handler has already checked and, if necessary, byte-swapped the
1170          * incoming request message body, but I am responsible for the
1171          * message buffers. */
1172
1173         if (req->rq_export == NULL) {
1174                 struct ldlm_request *dlm_req;
1175                 char str[PTL_NALFMT_SIZE];
1176                 CERROR("operation %d with bad export from NID %s\n",
1177                        req->rq_reqmsg->opc,
1178                        ptlrpc_peernid2str(&req->rq_peer, str));
1179
1180                 CERROR("--> export cookie: "LPX64"\n",
1181                        req->rq_reqmsg->handle.cookie);
1182                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1183                                              lustre_swab_ldlm_request);
1184                 if (dlm_req != NULL)
1185                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1186                 ldlm_callback_reply(req, -ENOTCONN);
1187                 RETURN(0);
1188         }
1189
1190         switch (req->rq_reqmsg->opc) {
1191
1192         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1193         case LDLM_CANCEL:
1194                 CDEBUG(D_INODE, "cancel\n");
1195                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1196                 rc = ldlm_handle_cancel(req);
1197                 break;
1198         default:
1199                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1200                 ldlm_callback_reply(req, -EINVAL);
1201         }
1202
1203         RETURN(rc);
1204 }
1205
1206 #ifdef __KERNEL__
1207 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1208 {
1209         struct ldlm_bl_work_item *blwi = NULL;
1210
1211         spin_lock(&blp->blp_lock);
1212         if (!list_empty(&blp->blp_list)) {
1213                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1214                                   blwi_entry);
1215                 list_del(&blwi->blwi_entry);
1216         }
1217         spin_unlock(&blp->blp_lock);
1218
1219         return blwi;
1220 }
1221
1222 struct ldlm_bl_thread_data {
1223         int                     bltd_num;
1224         struct ldlm_bl_pool     *bltd_blp;
1225 };
1226
1227 static int ldlm_bl_thread_main(void *arg)
1228 {
1229         struct ldlm_bl_thread_data *bltd = arg;
1230         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1231         unsigned long flags;
1232         ENTRY;
1233
1234         /* XXX boiler-plate */
1235         {
1236                 char name[sizeof(current->comm)];
1237                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1238                          bltd->bltd_num);
1239                 kportal_daemonize(name);
1240         }
1241         SIGNAL_MASK_LOCK(current, flags);
1242         sigfillset(&current->blocked);
1243         RECALC_SIGPENDING;
1244         SIGNAL_MASK_UNLOCK(current, flags);
1245
1246         atomic_inc(&blp->blp_num_threads);
1247         complete(&blp->blp_comp);
1248
1249         while(1) {
1250                 struct l_wait_info lwi = { 0 };
1251                 struct ldlm_bl_work_item *blwi = NULL;
1252
1253                 l_wait_event_exclusive(blp->blp_waitq,
1254                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1255                                        &lwi);
1256
1257                 if (blwi->blwi_ns == NULL)
1258                         break;
1259
1260                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1261                                         blwi->blwi_lock);
1262                 OBD_FREE(blwi, sizeof(*blwi));
1263         }
1264
1265         atomic_dec(&blp->blp_num_threads);
1266         complete(&blp->blp_comp);
1267         RETURN(0);
1268 }
1269
1270 #endif
1271
1272 static int ldlm_setup(void);
1273 static int ldlm_cleanup(int force);
1274
1275 int ldlm_get_ref(void)
1276 {
1277         int rc = 0;
1278         down(&ldlm_ref_sem);
1279         if (++ldlm_refcount == 1) {
1280                 rc = ldlm_setup();
1281                 if (rc)
1282                         ldlm_refcount--;
1283         }
1284         up(&ldlm_ref_sem);
1285
1286         RETURN(rc);
1287 }
1288
1289 void ldlm_put_ref(int force)
1290 {
1291         down(&ldlm_ref_sem);
1292         if (ldlm_refcount == 1) {
1293                 int rc = ldlm_cleanup(force);
1294                 if (rc)
1295                         CERROR("ldlm_cleanup failed: %d\n", rc);
1296                 else
1297                         ldlm_refcount--;
1298         } else {
1299                 ldlm_refcount--;
1300         }
1301         up(&ldlm_ref_sem);
1302
1303         EXIT;
1304 }
1305
1306 static int ldlm_setup(void)
1307 {
1308         struct ldlm_bl_pool *blp;
1309         int rc = 0;
1310 #ifdef __KERNEL__
1311         int i;
1312 #endif
1313         ENTRY;
1314
1315         if (ldlm_state != NULL)
1316                 RETURN(-EALREADY);
1317
1318         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1319         if (ldlm_state == NULL)
1320                 RETURN(-ENOMEM);
1321
1322 #ifdef __KERNEL__
1323         rc = ldlm_proc_setup();
1324         if (rc != 0)
1325                 GOTO(out_free, rc);
1326 #endif
1327
1328         ldlm_state->ldlm_cb_service =
1329                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1330                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1331                                 ldlm_callback_handler, "ldlm_cbd",
1332                                 ldlm_svc_proc_dir);
1333
1334         if (!ldlm_state->ldlm_cb_service) {
1335                 CERROR("failed to start service\n");
1336                 GOTO(out_proc, rc = -ENOMEM);
1337         }
1338
1339         ldlm_state->ldlm_cancel_service =
1340                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1341                                 LDLM_CANCEL_REQUEST_PORTAL,
1342                                 LDLM_CANCEL_REPLY_PORTAL,
1343                                 ldlm_cancel_handler, "ldlm_canceld",
1344                                 ldlm_svc_proc_dir);
1345
1346         if (!ldlm_state->ldlm_cancel_service) {
1347                 CERROR("failed to start service\n");
1348                 GOTO(out_proc, rc = -ENOMEM);
1349         }
1350
1351         OBD_ALLOC(blp, sizeof(*blp));
1352         if (blp == NULL)
1353                 GOTO(out_proc, rc = -ENOMEM);
1354         ldlm_state->ldlm_bl_pool = blp;
1355
1356         atomic_set(&blp->blp_num_threads, 0);
1357         init_waitqueue_head(&blp->blp_waitq);
1358         spin_lock_init(&blp->blp_lock);
1359
1360         INIT_LIST_HEAD(&blp->blp_list);
1361
1362 #ifdef __KERNEL__
1363         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1364                 struct ldlm_bl_thread_data bltd = {
1365                         .bltd_num = i,
1366                         .bltd_blp = blp,
1367                 };
1368                 init_completion(&blp->blp_comp);
1369                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1370                 if (rc < 0) {
1371                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1372                         GOTO(out_thread, rc);
1373                 }
1374                 wait_for_completion(&blp->blp_comp);
1375         }
1376
1377         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1378                                     LDLM_NUM_THREADS, "ldlm_cn");
1379         if (rc)
1380                 GOTO(out_thread, rc);
1381
1382         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1383                                     LDLM_NUM_THREADS, "ldlm_cb");
1384         if (rc)
1385                 GOTO(out_thread, rc);
1386
1387         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1388         spin_lock_init(&expired_lock_thread.elt_lock);
1389         expired_lock_thread.elt_state = ELT_STOPPED;
1390         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1391
1392         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1393         if (rc < 0) {
1394                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1395                 GOTO(out_thread, rc);
1396         }
1397
1398         wait_event(expired_lock_thread.elt_waitq,
1399                    expired_lock_thread.elt_state == ELT_READY);
1400
1401         INIT_LIST_HEAD(&waiting_locks_list);
1402         spin_lock_init(&waiting_locks_spinlock);
1403         waiting_locks_timer.function = waiting_locks_callback;
1404         waiting_locks_timer.data = 0;
1405         init_timer(&waiting_locks_timer);
1406 #endif
1407
1408         RETURN(0);
1409
1410 #ifdef __KERNEL__
1411  out_thread:
1412         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1413         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1414 #endif
1415
1416  out_proc:
1417 #ifdef __KERNEL__
1418         ldlm_proc_cleanup();
1419  out_free:
1420 #endif
1421         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1422         ldlm_state = NULL;
1423         return rc;
1424 }
1425
1426 static int ldlm_cleanup(int force)
1427 {
1428 #ifdef __KERNEL__
1429         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1430 #endif
1431         ENTRY;
1432
1433         if (!list_empty(&ldlm_namespace_list)) {
1434                 CERROR("ldlm still has namespaces; clean these up first.\n");
1435                 ldlm_dump_all_namespaces();
1436                 RETURN(-EBUSY);
1437         }
1438
1439 #ifdef __KERNEL__
1440         while (atomic_read(&blp->blp_num_threads) > 0) {
1441                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1442
1443                 init_completion(&blp->blp_comp);
1444
1445                 spin_lock(&blp->blp_lock);
1446                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1447                 wake_up(&blp->blp_waitq);
1448                 spin_unlock(&blp->blp_lock);
1449
1450                 wait_for_completion(&blp->blp_comp);
1451         }
1452         OBD_FREE(blp, sizeof(*blp));
1453
1454         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1455         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1456         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1457         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1458         ldlm_proc_cleanup();
1459
1460         expired_lock_thread.elt_state = ELT_TERMINATE;
1461         wake_up(&expired_lock_thread.elt_waitq);
1462         wait_event(expired_lock_thread.elt_waitq,
1463                    expired_lock_thread.elt_state == ELT_STOPPED);
1464 #else
1465         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1466         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1467 #endif
1468
1469         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1470         ldlm_state = NULL;
1471
1472         RETURN(0);
1473 }
1474
1475 int __init ldlm_init(void)
1476 {
1477         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1478                                                sizeof(struct ldlm_resource), 0,
1479                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1480         if (ldlm_resource_slab == NULL)
1481                 return -ENOMEM;
1482
1483         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1484                                            sizeof(struct ldlm_lock), 0,
1485                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1486         if (ldlm_lock_slab == NULL) {
1487                 kmem_cache_destroy(ldlm_resource_slab);
1488                 return -ENOMEM;
1489         }
1490
1491         l_lock_init(&ldlm_handle_lock);
1492
1493         return 0;
1494 }
1495
1496 void __exit ldlm_exit(void)
1497 {
1498         if ( ldlm_refcount )
1499                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1500         LASSERTF(kmem_cache_destroy(ldlm_resource_slab) == 0,
1501                  "couldn't free ldlm resource slab\n");
1502         LASSERTF(kmem_cache_destroy(ldlm_lock_slab) == 0,
1503                  "couldn't free ldlm lock slab\n");
1504
1505 }
1506
1507 /* ldlm_flock.c */
1508 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1509
1510 /* ldlm_extent.c */
1511 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1512
1513 /* ldlm_lock.c */
1514 EXPORT_SYMBOL(ldlm_get_processing_policy);
1515 EXPORT_SYMBOL(ldlm_lock2desc);
1516 EXPORT_SYMBOL(ldlm_register_intent);
1517 EXPORT_SYMBOL(ldlm_lockname);
1518 EXPORT_SYMBOL(ldlm_typename);
1519 EXPORT_SYMBOL(ldlm_lock2handle);
1520 EXPORT_SYMBOL(__ldlm_handle2lock);
1521 EXPORT_SYMBOL(ldlm_lock_get);
1522 EXPORT_SYMBOL(ldlm_lock_put);
1523 EXPORT_SYMBOL(ldlm_lock_match);
1524 EXPORT_SYMBOL(ldlm_lock_cancel);
1525 EXPORT_SYMBOL(ldlm_lock_addref);
1526 EXPORT_SYMBOL(ldlm_lock_decref);
1527 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1528 EXPORT_SYMBOL(ldlm_lock_change_resource);
1529 EXPORT_SYMBOL(ldlm_lock_set_data);
1530 EXPORT_SYMBOL(ldlm_it2str);
1531 EXPORT_SYMBOL(ldlm_lock_dump);
1532 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1533 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1534 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1535 EXPORT_SYMBOL(ldlm_lock_allow_match);
1536
1537 /* ldlm_request.c */
1538 EXPORT_SYMBOL(ldlm_completion_ast);
1539 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1540 EXPORT_SYMBOL(ldlm_cli_convert);
1541 EXPORT_SYMBOL(ldlm_cli_enqueue);
1542 EXPORT_SYMBOL(ldlm_cli_cancel);
1543 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1544 EXPORT_SYMBOL(ldlm_replay_locks);
1545 EXPORT_SYMBOL(ldlm_resource_foreach);
1546 EXPORT_SYMBOL(ldlm_namespace_foreach);
1547 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1548 EXPORT_SYMBOL(ldlm_change_cbdata);
1549
1550 /* ldlm_lockd.c */
1551 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1552 EXPORT_SYMBOL(ldlm_server_completion_ast);
1553 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1554 EXPORT_SYMBOL(ldlm_handle_enqueue);
1555 EXPORT_SYMBOL(ldlm_handle_cancel);
1556 EXPORT_SYMBOL(ldlm_handle_convert);
1557 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1558 EXPORT_SYMBOL(ldlm_get_ref);
1559 EXPORT_SYMBOL(ldlm_put_ref);
1560
1561 #if 0
1562 /* ldlm_test.c */
1563 EXPORT_SYMBOL(ldlm_test);
1564 EXPORT_SYMBOL(ldlm_regression_start);
1565 EXPORT_SYMBOL(ldlm_regression_stop);
1566 #endif
1567
1568 /* ldlm_resource.c */
1569 EXPORT_SYMBOL(ldlm_namespace_new);
1570 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1571 EXPORT_SYMBOL(ldlm_namespace_free);
1572 EXPORT_SYMBOL(ldlm_namespace_dump);
1573 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1574 EXPORT_SYMBOL(ldlm_resource_get);
1575 EXPORT_SYMBOL(ldlm_resource_putref);
1576
1577 /* l_lock.c */
1578 EXPORT_SYMBOL(l_lock);
1579 EXPORT_SYMBOL(l_unlock);
1580
1581 /* ldlm_lib.c */
1582 EXPORT_SYMBOL(client_import_add_conn);
1583 EXPORT_SYMBOL(client_import_del_conn);
1584 EXPORT_SYMBOL(client_obd_setup);
1585 EXPORT_SYMBOL(client_obd_cleanup);
1586 EXPORT_SYMBOL(client_connect_import);
1587 EXPORT_SYMBOL(client_disconnect_export);
1588 EXPORT_SYMBOL(target_start_recovery_thread);
1589 EXPORT_SYMBOL(target_stop_recovery_thread);
1590 EXPORT_SYMBOL(target_handle_connect);
1591 EXPORT_SYMBOL(target_cleanup_recovery);
1592 EXPORT_SYMBOL(target_destroy_export);
1593 EXPORT_SYMBOL(target_cancel_recovery_timer);
1594 EXPORT_SYMBOL(target_send_reply);
1595 EXPORT_SYMBOL(target_queue_recovery_request);
1596 EXPORT_SYMBOL(target_handle_ping);
1597 EXPORT_SYMBOL(target_handle_disconnect);
1598 EXPORT_SYMBOL(target_queue_final_reply);