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         LDLM_ERROR(lock, "%s AST failed (%d): evicting client %s@%s NID "LPX64
 313                    " (%s)", ast_type, rc, lock->l_export->exp_client_uuid.uuid,
 314                    conn->c_remote_uuid.uuid, conn->c_peer.peer_id.nid,
 315                    ptlrpc_peernid2str(&conn->c_peer, str));
 316         ptlrpc_fail_export(lock->l_export);
 317 }
 318
 319 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 320                                  struct ptlrpc_request *req, int rc,
 321                                  const char *ast_type)
 322 {
 323         struct ptlrpc_peer *peer = &req->rq_import->imp_connection->c_peer;
 324         char str[PTL_NALFMT_SIZE];
 325
 326         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 327                 LASSERT(lock->l_export);
 328                 if (lock->l_export->exp_libclient) {
 329                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 330                                    " timeout, just cancelling lock", ast_type,
 331                                    ptlrpc_peernid2str(peer, str));
 332                         ldlm_lock_cancel(lock);
 333                         rc = -ERESTART;
 334                 } else {
 335                         ldlm_del_waiting_lock(lock);
 336                         ldlm_failed_ast(lock, rc, ast_type);
 337                 }
 338         } else if (rc) {
 339                 if (rc == -EINVAL)
 340                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 341                                    " from %s AST - normal race",
 342                                    ptlrpc_peernid2str(peer, str),
 343                                    req->rq_repmsg->status, ast_type);
 344                 else
 345                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 346                                    "from %s AST", ptlrpc_peernid2str(peer, str),
 347                                    (req->rq_repmsg != NULL) ?
 348                                    req->rq_repmsg->status : 0, ast_type);
 349                 ldlm_lock_cancel(lock);
 350                 /* Server-side AST functions are called from ldlm_reprocess_all,
 351                  * which needs to be told to please restart its reprocessing. */
 352                 rc = -ERESTART;
 353         }
 354
 355         return rc;
 356 }
 357
 358 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 359                              struct ldlm_lock_desc *desc,
 360                              void *data, int flag)
 361 {
 362         struct ldlm_request *body;
 363         struct ptlrpc_request *req;
 364         int rc = 0, size = sizeof(*body);
 365         ENTRY;
 366
 367         if (flag == LDLM_CB_CANCELING) {
 368                 /* Don't need to do anything here. */
 369                 RETURN(0);
 370         }
 371
 372         LASSERT(lock);
 373
 374         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 375         if (lock->l_granted_mode != lock->l_req_mode) {
 376                 /* this blocking AST will be communicated as part of the
 377                  * completion AST instead */
 378                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 379                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");                 RETURN(0);
 380         }
 381
 382         if (lock->l_destroyed) {
 383                 /* What's the point? */
 384                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 385                 RETURN(0);
 386         }
 387
 388 #if 0
 389         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 390                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 391                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 392                 RETURN(-ETIMEDOUT);
 393         }
 394 #endif
 395
 396         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 397                               LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK,
 398                               1, &size, NULL);
 399         if (req == NULL) {
 400                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 401                 RETURN(-ENOMEM);
 402         }
 403
 404         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 405         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 406                sizeof(body->lock_handle1));
 407         memcpy(&body->lock_desc, desc, sizeof(*desc));
 408         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 409
 410         LDLM_DEBUG(lock, "server preparing blocking AST");
 411         req->rq_replen = lustre_msg_size(0, NULL);
 412
 413         if (lock->l_granted_mode == lock->l_req_mode)
 414                 ldlm_add_waiting_lock(lock);
 415         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 416
 417         req->rq_send_state = LUSTRE_IMP_FULL;
 418         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 419         rc = ptlrpc_queue_wait(req);
 420         if (rc != 0)
 421                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 422
 423         ptlrpc_req_finished(req);
 424
 425         RETURN(rc);
 426 }
 427
 428 /* XXX copied from ptlrpc/service.c */
 429 static long timeval_sub(struct timeval *large, struct timeval *small)
 430 {
 431         return (large->tv_sec - small->tv_sec) * 1000000 +
 432                 (large->tv_usec - small->tv_usec);
 433 }
 434
 435 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 436 {
 437         struct ldlm_request *body;
 438         struct ptlrpc_request *req;
 439         struct timeval granted_time;
 440         long total_enqueue_wait;
 441         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 442         ENTRY;
 443
 444         LASSERT(lock != NULL);
 445
 446         do_gettimeofday(&granted_time);
 447         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 448
 449         if (total_enqueue_wait / 1000000 > obd_timeout)
 450                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 451
 452         if (lock->l_resource->lr_lvb_len) {
 453                 buffers = 2;
 454                 size[1] = lock->l_resource->lr_lvb_len;
 455         }
 456
 457         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 458                               LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK,
 459                               buffers, size, NULL);
 460         if (req == NULL)
 461                 RETURN(-ENOMEM);
 462
 463         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 464         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 465                sizeof(body->lock_handle1));
 466         body->lock_flags = flags;
 467         ldlm_lock2desc(lock, &body->lock_desc);
 468
 469         if (buffers == 2) {
 470                 void *lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 471                                            lock->l_resource->lr_lvb_len);
 472                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 473                        lock->l_resource->lr_lvb_len);
 474         }
 475
 476         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 477                    total_enqueue_wait);
 478         req->rq_replen = lustre_msg_size(0, NULL);
 479
 480         req->rq_send_state = LUSTRE_IMP_FULL;
 481         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 482
 483         /* We only send real blocking ASTs after the lock is granted */
 484         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 485         if (lock->l_flags & LDLM_FL_AST_SENT) {
 486                 body->lock_flags |= LDLM_FL_AST_SENT;
 487                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 488         }
 489         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 490
 491         rc = ptlrpc_queue_wait(req);
 492         if (rc != 0)
 493                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 494
 495         ptlrpc_req_finished(req);
 496
 497         RETURN(rc);
 498 }
 499
 500 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 501 {
 502         struct ldlm_resource *res = lock->l_resource;
 503         struct ldlm_request *body;
 504         struct ptlrpc_request *req;
 505         int rc = 0, size = sizeof(*body);
 506         ENTRY;
 507
 508         LASSERT(lock != NULL);
 509
 510         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 511                               LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK,
 512                               1, &size, NULL);
 513         if (req == NULL)
 514                 RETURN(-ENOMEM);
 515
 516         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 517         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 518                sizeof(body->lock_handle1));
 519         ldlm_lock2desc(lock, &body->lock_desc);
 520
 521         size = lock->l_resource->lr_lvb_len;
 522         req->rq_replen = lustre_msg_size(1, &size);
 523
 524         req->rq_send_state = LUSTRE_IMP_FULL;
 525         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 526
 527         rc = ptlrpc_queue_wait(req);
 528         if (rc == -ELDLM_NO_LOCK_DATA)
 529                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 530         else if (rc != 0)
 531                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 532         else
 533                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 534                         (res, req->rq_repmsg, 0, 1);
 535         ptlrpc_req_finished(req);
 536         RETURN(rc);
 537 }
 538
 539 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 540                         ldlm_completion_callback completion_callback,
 541                         ldlm_blocking_callback blocking_callback,
 542                         ldlm_glimpse_callback glimpse_callback)
 543 {
 544         struct obd_device *obddev = req->rq_export->exp_obd;
 545         struct ldlm_reply *dlm_rep;
 546         struct ldlm_request *dlm_req;
 547         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 548         __u32 flags;
 549         ldlm_error_t err = ELDLM_OK;
 550         struct ldlm_lock *lock = NULL;
 551         void *cookie = NULL;
 552         ENTRY;
 553
 554         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 555
 556         dlm_req = lustre_swab_reqbuf (req, MDS_REQ_INTENT_LOCKREQ_OFF,
 557                                       sizeof (*dlm_req),
 558                                       lustre_swab_ldlm_request);
 559         if (dlm_req == NULL) {
 560                 CERROR ("Can't unpack dlm_req\n");
 561                 GOTO(out, rc = -EFAULT);
 562         }
 563
 564         flags = dlm_req->lock_flags;
 565
 566         /* The lock's callback data might be set in the policy function */
 567         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 568                                 dlm_req->lock_desc.l_resource.lr_name,
 569                                 dlm_req->lock_desc.l_resource.lr_type,
 570                                 dlm_req->lock_desc.l_req_mode,
 571                                 blocking_callback, completion_callback,
 572                                 glimpse_callback, NULL, 0);
 573         if (!lock)
 574                 GOTO(out, rc = -ENOMEM);
 575
 576         do_gettimeofday(&lock->l_enqueued_time);
 577         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 578                sizeof(lock->l_remote_handle));
 579         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 580
 581         LASSERT(req->rq_export);
 582         lock->l_export = class_export_get(req->rq_export);
 583         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 584         list_add(&lock->l_export_chain,
 585                  &lock->l_export->exp_ldlm_data.led_held_locks);
 586         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 587
 588         if (flags & LDLM_FL_HAS_INTENT) {
 589                 /* In this case, the reply buffer is allocated deep in
 590                  * local_lock_enqueue by the policy function. */
 591                 cookie = req;
 592         } else {
 593                 int buffers = 1;
 594                 if (lock->l_resource->lr_lvb_len) {
 595                         size[1] = lock->l_resource->lr_lvb_len;
 596                         buffers = 2;
 597                 }
 598
 599                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 600                         GOTO(out, rc = -ENOMEM);
 601
 602                 rc = lustre_pack_reply(req, buffers, size, NULL);
 603                 if (rc)
 604                         GOTO(out, rc);
 605         }
 606
 607         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 608                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 609                        sizeof(ldlm_policy_data_t));
 610         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 611                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 612                        sizeof(lock->l_req_extent));
 613
 614         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 615         if (err)
 616                 GOTO(out, err);
 617
 618         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 619         dlm_rep->lock_flags = flags;
 620
 621         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 622         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 623
 624         /* We never send a blocking AST until the lock is granted, but
 625          * we can tell it right now */
 626         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 627         if (lock->l_flags & LDLM_FL_AST_SENT) {
 628                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 629                 if (lock->l_granted_mode == lock->l_req_mode)
 630                         ldlm_add_waiting_lock(lock);
 631         }
 632         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 633
 634         EXIT;
 635  out:
 636         req->rq_status = err;
 637         if (req->rq_reply_state == NULL) {
 638                 err = lustre_pack_reply(req, 0, NULL, NULL);
 639                 if (rc == 0)
 640                         rc = err;
 641                 req->rq_status = rc;
 642         }
 643
 644         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 645          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 646         if (lock) {
 647                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 648                            "(err=%d, rc=%d)", err, rc);
 649
 650                 if (lock->l_resource->lr_lvb_len > 0 && rc == 0) {
 651                         void *lvb = lustre_msg_buf(req->rq_repmsg, 1,
 652                                                   lock->l_resource->lr_lvb_len);
 653                         LASSERT(lvb != NULL);
 654                         memcpy(lvb, lock->l_resource->lr_lvb_data,
 655                                lock->l_resource->lr_lvb_len);
 656                 }
 657
 658                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 659                         ldlm_reprocess_all(lock->l_resource);
 660                 LDLM_LOCK_PUT(lock);
 661         }
 662         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 663                           lock, rc);
 664
 665         return rc;
 666 }
 667
 668 int ldlm_handle_convert(struct ptlrpc_request *req)
 669 {
 670         struct ldlm_request *dlm_req;
 671         struct ldlm_reply *dlm_rep;
 672         struct ldlm_lock *lock;
 673         int rc, size = sizeof(*dlm_rep);
 674         ENTRY;
 675
 676         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 677                                       lustre_swab_ldlm_request);
 678         if (dlm_req == NULL) {
 679                 CERROR ("Can't unpack dlm_req\n");
 680                 RETURN (-EFAULT);
 681         }
 682
 683         rc = lustre_pack_reply(req, 1, &size, NULL);
 684         if (rc) {
 685                 CERROR("out of memory\n");
 686                 RETURN(-ENOMEM);
 687         }
 688         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 689         dlm_rep->lock_flags = dlm_req->lock_flags;
 690
 691         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 692         if (!lock) {
 693                 req->rq_status = EINVAL;
 694         } else {
 695                 LDLM_DEBUG(lock, "server-side convert handler START");
 696                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 697                                   &dlm_rep->lock_flags);
 698                 if (ldlm_del_waiting_lock(lock))
 699                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 700                 req->rq_status = 0;
 701         }
 702
 703         if (lock) {
 704                 ldlm_reprocess_all(lock->l_resource);
 705                 LDLM_DEBUG(lock, "server-side convert handler END");
 706                 LDLM_LOCK_PUT(lock);
 707         } else
 708                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 709
 710         RETURN(0);
 711 }
 712
 713 int ldlm_handle_cancel(struct ptlrpc_request *req)
 714 {
 715         struct ldlm_request *dlm_req;
 716         struct ldlm_lock *lock;
 717         struct ldlm_resource *res;
 718         char str[PTL_NALFMT_SIZE];
 719         int rc;
 720         ENTRY;
 721
 722         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 723                                       lustre_swab_ldlm_request);
 724         if (dlm_req == NULL) {
 725                 CERROR("bad request buffer for cancel\n");
 726                 RETURN(-EFAULT);
 727         }
 728
 729         rc = lustre_pack_reply(req, 0, NULL, NULL);
 730         if (rc) {
 731                 CERROR("out of memory\n");
 732                 RETURN(-ENOMEM);
 733         }
 734
 735         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 736         if (!lock) {
 737                 CERROR("received cancel for unknown lock cookie "LPX64
 738                        " from client %s nid %s\n",
 739                        dlm_req->lock_handle1.cookie,
 740                        req->rq_export->exp_client_uuid.uuid,
 741                        ptlrpc_peernid2str(&req->rq_peer, str));
 742                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 743                                   "(cookie "LPU64")",
 744                                   dlm_req->lock_handle1.cookie);
 745                 req->rq_status = ESTALE;
 746         } else {
 747                 LDLM_DEBUG(lock, "server-side cancel handler START");
 748                 res = lock->l_resource;
 749                 if (res && res->lr_namespace->ns_lvbo &&
 750                     res->lr_namespace->ns_lvbo->lvbo_update) {
 751                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 752                                 (res, NULL, 0, 0);
 753                                 //(res, req->rq_reqmsg, 1);
 754                 }
 755
 756                 ldlm_lock_cancel(lock);
 757                 if (ldlm_del_waiting_lock(lock))
 758                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 759                 req->rq_status = rc;
 760         }
 761
 762         if (ptlrpc_reply(req) != 0)
 763                 LBUG();
 764
 765         if (lock) {
 766                 ldlm_reprocess_all(lock->l_resource);
 767                 LDLM_DEBUG(lock, "server-side cancel handler END");
 768                 LDLM_LOCK_PUT(lock);
 769         }
 770
 771         RETURN(0);
 772 }
 773
 774 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 775                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 776 {
 777         int do_ast;
 778         ENTRY;
 779
 780         l_lock(&ns->ns_lock);
 781         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 782
 783         lock->l_flags |= LDLM_FL_CBPENDING;
 784         do_ast = (!lock->l_readers && !lock->l_writers);
 785
 786         if (do_ast) {
 787                 LDLM_DEBUG(lock, "already unused, calling "
 788                            "callback (%p)", lock->l_blocking_ast);
 789                 if (lock->l_blocking_ast != NULL) {
 790                         l_unlock(&ns->ns_lock);
 791                         l_check_no_ns_lock(ns);
 792                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 793                                              LDLM_CB_BLOCKING);
 794                         l_lock(&ns->ns_lock);
 795                 }
 796         } else {
 797                 LDLM_DEBUG(lock, "Lock still has references, will be"
 798                            " cancelled later");
 799         }
 800
 801         LDLM_DEBUG(lock, "client blocking callback handler END");
 802         l_unlock(&ns->ns_lock);
 803         LDLM_LOCK_PUT(lock);
 804         EXIT;
 805 }
 806
 807 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 808                                     struct ldlm_namespace *ns,
 809                                     struct ldlm_request *dlm_req,
 810                                     struct ldlm_lock *lock)
 811 {
 812         LIST_HEAD(ast_list);
 813         ENTRY;
 814
 815         l_lock(&ns->ns_lock);
 816         LDLM_DEBUG(lock, "client completion callback handler START");
 817
 818         /* If we receive the completion AST before the actual enqueue returned,
 819          * then we might need to switch lock modes, resources, or extents. */
 820         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 821                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 822                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 823         }
 824
 825         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 826                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 827                        sizeof(lock->l_policy_data));
 828                 LDLM_DEBUG(lock, "completion AST, new policy data");
 829         }
 830
 831         ldlm_resource_unlink_lock(lock);
 832         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 833                    &lock->l_resource->lr_name,
 834                    sizeof(lock->l_resource->lr_name)) != 0) {
 835                 ldlm_lock_change_resource(ns, lock,
 836                                          dlm_req->lock_desc.l_resource.lr_name);
 837                 LDLM_DEBUG(lock, "completion AST, new resource");
 838         }
 839
 840         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 841                 lock->l_flags |= LDLM_FL_CBPENDING;
 842                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 843         }
 844
 845         if (lock->l_lvb_len) {
 846                 void *lvb;
 847                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 848                                          lock->l_lvb_swabber);
 849                 if (lvb == NULL) {
 850                         LDLM_ERROR(lock, "completion AST did not contain "
 851                                    "expected LVB!");
 852                 } else {
 853                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 854                 }
 855         }
 856
 857         lock->l_resource->lr_tmp = &ast_list;
 858         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 859         lock->l_resource->lr_tmp = NULL;
 860         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 861         l_unlock(&ns->ns_lock);
 862         LDLM_LOCK_PUT(lock);
 863
 864         ldlm_run_ast_work(ns, &ast_list);
 865
 866         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 867                           lock);
 868         EXIT;
 869 }
 870
 871 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 872                                     struct ldlm_namespace *ns,
 873                                     struct ldlm_request *dlm_req,
 874                                     struct ldlm_lock *lock)
 875 {
 876         int rc = -ENOSYS;
 877         ENTRY;
 878
 879         l_lock(&ns->ns_lock);
 880         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 881
 882         if (lock->l_glimpse_ast != NULL) {
 883                 l_unlock(&ns->ns_lock);
 884                 l_check_no_ns_lock(ns);
 885                 rc = lock->l_glimpse_ast(lock, req);
 886                 l_lock(&ns->ns_lock);
 887         }
 888
 889         if (req->rq_repmsg != NULL) {
 890                 ptlrpc_reply(req);
 891         } else {
 892                 req->rq_status = rc;
 893                 ptlrpc_error(req);
 894         }
 895
 896         if (lock->l_granted_mode == LCK_PW &&
 897             !lock->l_readers && !lock->l_writers &&
 898             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 899 #ifdef __KERNEL__
 900                 ldlm_bl_to_thread(ns, NULL, lock);
 901                 l_unlock(&ns->ns_lock);
 902 #else
 903                 l_unlock(&ns->ns_lock);
 904                 ldlm_handle_bl_callback(ns, NULL, lock);
 905 #endif
 906                 EXIT;
 907                 return;
 908         }
 909
 910         l_unlock(&ns->ns_lock);
 911         LDLM_LOCK_PUT(lock);
 912         EXIT;
 913 }
 914
 915 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 916 {
 917         req->rq_status = rc;
 918         if (req->rq_reply_state == NULL) {
 919                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 920                 if (rc)
 921                         return rc;
 922         }
 923         return ptlrpc_reply(req);
 924 }
 925
 926 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
 927                       struct ldlm_lock *lock)
 928 {
 929 #ifdef __KERNEL__
 930         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
 931         struct ldlm_bl_work_item *blwi;
 932         ENTRY;
 933
 934         OBD_ALLOC(blwi, sizeof(*blwi));
 935         if (blwi == NULL)
 936                 RETURN(-ENOMEM);
 937
 938         blwi->blwi_ns = ns;
 939         if (ld != NULL)
 940                 blwi->blwi_ld = *ld;
 941         blwi->blwi_lock = lock;
 942
 943         spin_lock(&blp->blp_lock);
 944         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
 945         wake_up(&blp->blp_waitq);
 946         spin_unlock(&blp->blp_lock);
 947 #else
 948         LBUG();
 949 #endif
 950
 951         RETURN(0);
 952 }
 953
 954 static int ldlm_msg_check_version(struct lustre_msg *msg)
 955 {
 956         int rc;
 957
 958         switch (msg->opc) {
 959         case LDLM_ENQUEUE:
 960         case LDLM_CONVERT:
 961         case LDLM_CANCEL:
 962         case LDLM_BL_CALLBACK:
 963         case LDLM_CP_CALLBACK:
 964         case LDLM_GL_CALLBACK:
 965                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
 966                 if (rc)
 967                         CERROR("bad opc %u version %08x, expecting %08x\n",
 968                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
 969                 break;
 970         case OBD_LOG_CANCEL:
 971         case LLOG_ORIGIN_HANDLE_OPEN:
 972         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
 973         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
 974         case LLOG_ORIGIN_HANDLE_READ_HEADER:
 975         case LLOG_ORIGIN_HANDLE_CLOSE:
 976         case LLOG_CATINFO:
 977                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
 978                 if (rc)
 979                         CERROR("bad opc %u version %08x, expecting %08x\n",
 980                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
 981                 break;
 982         default:
 983                 CERROR("LDLM unknown opcode %d\n", msg->opc);
 984                 rc = -ENOTSUPP;
 985                 break;
 986         }
 987
 988         return rc;
 989 }
 990
 991 static int ldlm_callback_handler(struct ptlrpc_request *req)
 992 {
 993         struct ldlm_namespace *ns;
 994         struct ldlm_request *dlm_req;
 995         struct ldlm_lock *lock;
 996         char str[PTL_NALFMT_SIZE];
 997         int rc;
 998         ENTRY;
 999
1000         rc = ldlm_msg_check_version(req->rq_reqmsg);
1001         if (rc) {
1002                 CERROR("LDLM_CB drop mal-formed request\n");
1003                 RETURN(rc);
1004         }
1005
1006         /* Requests arrive in sender's byte order.  The ptlrpc service
1007          * handler has already checked and, if necessary, byte-swapped the
1008          * incoming request message body, but I am responsible for the
1009          * message buffers. */
1010
1011         if (req->rq_export == NULL) {
1012                 struct ldlm_request *dlm_req;
1013
1014                 CDEBUG(D_RPCTRACE, "operation %d from nid %s with bad "
1015                        "export cookie "LPX64" (ptl req %d/rep %d); this is "
1016                        "normal if this node rebooted with a lock held\n",
1017                        req->rq_reqmsg->opc,
1018                        ptlrpc_peernid2str(&req->rq_peer, str),
1019                        req->rq_reqmsg->handle.cookie,
1020                        req->rq_request_portal, req->rq_reply_portal);
1021
1022                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1023                                              lustre_swab_ldlm_request);
1024                 if (dlm_req != NULL)
1025                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1026                                dlm_req->lock_handle1.cookie);
1027
1028                 ldlm_callback_reply(req, -ENOTCONN);
1029                 RETURN(0);
1030         }
1031
1032         LASSERT(req->rq_export != NULL);
1033         LASSERT(req->rq_export->exp_obd != NULL);
1034
1035         switch(req->rq_reqmsg->opc) {
1036         case LDLM_BL_CALLBACK:
1037                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1038                 break;
1039         case LDLM_CP_CALLBACK:
1040                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1041                 break;
1042         case LDLM_GL_CALLBACK:
1043                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1044                 break;
1045         case OBD_LOG_CANCEL:
1046                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1047                 rc = llog_origin_handle_cancel(req);
1048                 ldlm_callback_reply(req, rc);
1049                 RETURN(0);
1050         case LLOG_ORIGIN_HANDLE_OPEN:
1051                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1052                 rc = llog_origin_handle_open(req);
1053                 ldlm_callback_reply(req, rc);
1054                 RETURN(0);
1055         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1056                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1057                 rc = llog_origin_handle_next_block(req);
1058                 ldlm_callback_reply(req, rc);
1059                 RETURN(0);
1060         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1061                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1062                 rc = llog_origin_handle_prev_block(req);
1063                 ldlm_callback_reply(req, rc);
1064                 RETURN(0);
1065         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1066                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1067                 rc = llog_origin_handle_read_header(req);
1068                 ldlm_callback_reply(req, rc);
1069                 RETURN(0);
1070         case LLOG_ORIGIN_HANDLE_CLOSE:
1071                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1072                 rc = llog_origin_handle_close(req);
1073                 ldlm_callback_reply(req, rc);
1074                 RETURN(0);
1075         default:
1076                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1077                 ldlm_callback_reply(req, -EPROTO);
1078                 RETURN(0);
1079         }
1080
1081         ns = req->rq_export->exp_obd->obd_namespace;
1082         LASSERT(ns != NULL);
1083
1084         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1085                                       lustre_swab_ldlm_request);
1086         if (dlm_req == NULL) {
1087                 CERROR ("can't unpack dlm_req\n");
1088                 ldlm_callback_reply (req, -EPROTO);
1089                 RETURN (0);
1090         }
1091
1092         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1093         if (!lock) {
1094                 LDLM_DEBUG_NOLOCK("callback on lock "LPX64" - lock "
1095                                   "disappeared\n",dlm_req->lock_handle1.cookie);
1096                 ldlm_callback_reply(req, -EINVAL);
1097                 RETURN(0);
1098         }
1099
1100         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1101         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1102
1103         /* We want the ost thread to get this reply so that it can respond
1104          * to ost requests (write cache writeback) that might be triggered
1105          * in the callback.
1106          *
1107          * But we'd also like to be able to indicate in the reply that we're
1108          * cancelling right now, because it's unused, or have an intent result
1109          * in the reply, so we might have to push the responsibility for sending
1110          * the reply down into the AST handlers, alas. */
1111
1112         switch (req->rq_reqmsg->opc) {
1113         case LDLM_BL_CALLBACK:
1114                 CDEBUG(D_INODE, "blocking ast\n");
1115 #ifdef __KERNEL__
1116                 rc = ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock);
1117                 ldlm_callback_reply(req, rc);
1118 #else
1119                 rc = 0;
1120                 ldlm_callback_reply(req, rc);
1121                 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1122 #endif
1123                 break;
1124         case LDLM_CP_CALLBACK:
1125                 CDEBUG(D_INODE, "completion ast\n");
1126                 ldlm_callback_reply(req, 0);
1127                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1128                 break;
1129         case LDLM_GL_CALLBACK:
1130                 CDEBUG(D_INODE, "glimpse ast\n");
1131                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1132                 break;
1133         default:
1134                 LBUG();                         /* checked above */
1135         }
1136
1137         RETURN(0);
1138 }
1139
1140 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1141 {
1142         int rc = 0;
1143         ENTRY;
1144
1145         rc = ldlm_msg_check_version(req->rq_reqmsg);
1146         if (rc) {
1147                 CERROR("LDLM_CL drop mal-formed request\n");
1148                 RETURN(rc);
1149         }
1150
1151         /* Requests arrive in sender's byte order.  The ptlrpc service
1152          * handler has already checked and, if necessary, byte-swapped the
1153          * incoming request message body, but I am responsible for the
1154          * message buffers. */
1155
1156         if (req->rq_export == NULL) {
1157                 struct ldlm_request *dlm_req;
1158                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
1159                        req->rq_reqmsg->opc, req->rq_request_portal,
1160                        req->rq_reply_portal);
1161                 CERROR("--> export cookie: "LPX64"\n",
1162                        req->rq_reqmsg->handle.cookie);
1163                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1164                                              lustre_swab_ldlm_request);
1165                 if (dlm_req != NULL)
1166                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1167                 ldlm_callback_reply(req, -ENOTCONN);
1168                 RETURN(0);
1169         }
1170
1171         switch (req->rq_reqmsg->opc) {
1172
1173         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1174         case LDLM_CANCEL:
1175                 CDEBUG(D_INODE, "cancel\n");
1176                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1177                 rc = ldlm_handle_cancel(req);
1178                 break;
1179         default:
1180                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1181                 ldlm_callback_reply(req, -EINVAL);
1182         }
1183
1184         RETURN(rc);
1185 }
1186
1187 #ifdef __KERNEL__
1188 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1189 {
1190         struct ldlm_bl_work_item *blwi = NULL;
1191
1192         spin_lock(&blp->blp_lock);
1193         if (!list_empty(&blp->blp_list)) {
1194                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1195                                   blwi_entry);
1196                 list_del(&blwi->blwi_entry);
1197         }
1198         spin_unlock(&blp->blp_lock);
1199
1200         return blwi;
1201 }
1202
1203 struct ldlm_bl_thread_data {
1204         int                     bltd_num;
1205         struct ldlm_bl_pool     *bltd_blp;
1206 };
1207
1208 static int ldlm_bl_thread_main(void *arg)
1209 {
1210         struct ldlm_bl_thread_data *bltd = arg;
1211         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1212         unsigned long flags;
1213         ENTRY;
1214
1215         /* XXX boiler-plate */
1216         {
1217                 char name[sizeof(current->comm)];
1218                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1219                          bltd->bltd_num);
1220                 kportal_daemonize(name);
1221         }
1222         SIGNAL_MASK_LOCK(current, flags);
1223         sigfillset(&current->blocked);
1224         RECALC_SIGPENDING;
1225         SIGNAL_MASK_UNLOCK(current, flags);
1226
1227         atomic_inc(&blp->blp_num_threads);
1228         complete(&blp->blp_comp);
1229
1230         while(1) {
1231                 struct l_wait_info lwi = { 0 };
1232                 struct ldlm_bl_work_item *blwi = NULL;
1233
1234                 l_wait_event_exclusive(blp->blp_waitq,
1235                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1236                                        &lwi);
1237
1238                 if (blwi->blwi_ns == NULL)
1239                         break;
1240
1241                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1242                                         blwi->blwi_lock);
1243                 OBD_FREE(blwi, sizeof(*blwi));
1244         }
1245
1246         atomic_dec(&blp->blp_num_threads);
1247         complete(&blp->blp_comp);
1248         RETURN(0);
1249 }
1250
1251 #endif
1252
1253 static int ldlm_setup(void);
1254 static int ldlm_cleanup(int force);
1255
1256 int ldlm_get_ref(void)
1257 {
1258         int rc = 0;
1259         down(&ldlm_ref_sem);
1260         if (++ldlm_refcount == 1) {
1261                 rc = ldlm_setup();
1262                 if (rc)
1263                         ldlm_refcount--;
1264         }
1265         up(&ldlm_ref_sem);
1266
1267         RETURN(rc);
1268 }
1269
1270 void ldlm_put_ref(int force)
1271 {
1272         down(&ldlm_ref_sem);
1273         if (ldlm_refcount == 1) {
1274                 int rc = ldlm_cleanup(force);
1275                 if (rc)
1276                         CERROR("ldlm_cleanup failed: %d\n", rc);
1277                 else
1278                         ldlm_refcount--;
1279         } else {
1280                 ldlm_refcount--;
1281         }
1282         up(&ldlm_ref_sem);
1283
1284         EXIT;
1285 }
1286
1287 static int ldlm_setup(void)
1288 {
1289         struct ldlm_bl_pool *blp;
1290         int rc = 0;
1291 #ifdef __KERNEL__
1292         int i;
1293 #endif
1294         ENTRY;
1295
1296         if (ldlm_state != NULL)
1297                 RETURN(-EALREADY);
1298
1299         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1300         if (ldlm_state == NULL)
1301                 RETURN(-ENOMEM);
1302
1303 #ifdef __KERNEL__
1304         rc = ldlm_proc_setup();
1305         if (rc != 0)
1306                 GOTO(out_free, rc);
1307 #endif
1308
1309         ldlm_state->ldlm_cb_service =
1310                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1311                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1312                                 ldlm_callback_handler, "ldlm_cbd",
1313                                 ldlm_svc_proc_dir);
1314
1315         if (!ldlm_state->ldlm_cb_service) {
1316                 CERROR("failed to start service\n");
1317                 GOTO(out_proc, rc = -ENOMEM);
1318         }
1319
1320         ldlm_state->ldlm_cancel_service =
1321                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1322                                 LDLM_CANCEL_REQUEST_PORTAL,
1323                                 LDLM_CANCEL_REPLY_PORTAL,
1324                                 ldlm_cancel_handler, "ldlm_canceld",
1325                                 ldlm_svc_proc_dir);
1326
1327         if (!ldlm_state->ldlm_cancel_service) {
1328                 CERROR("failed to start service\n");
1329                 GOTO(out_proc, rc = -ENOMEM);
1330         }
1331
1332         OBD_ALLOC(blp, sizeof(*blp));
1333         if (blp == NULL)
1334                 GOTO(out_proc, rc = -ENOMEM);
1335         ldlm_state->ldlm_bl_pool = blp;
1336
1337         atomic_set(&blp->blp_num_threads, 0);
1338         init_waitqueue_head(&blp->blp_waitq);
1339         spin_lock_init(&blp->blp_lock);
1340
1341         INIT_LIST_HEAD(&blp->blp_list);
1342
1343 #ifdef __KERNEL__
1344         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1345                 struct ldlm_bl_thread_data bltd = {
1346                         .bltd_num = i,
1347                         .bltd_blp = blp,
1348                 };
1349                 init_completion(&blp->blp_comp);
1350                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1351                 if (rc < 0) {
1352                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1353                         GOTO(out_thread, rc);
1354                 }
1355                 wait_for_completion(&blp->blp_comp);
1356         }
1357
1358         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1359                                     LDLM_NUM_THREADS, "ldlm_cn");
1360         if (rc)
1361                 GOTO(out_thread, rc);
1362
1363         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1364                                     LDLM_NUM_THREADS, "ldlm_cb");
1365         if (rc)
1366                 GOTO(out_thread, rc);
1367
1368         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1369         spin_lock_init(&expired_lock_thread.elt_lock);
1370         expired_lock_thread.elt_state = ELT_STOPPED;
1371         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1372
1373         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1374         if (rc < 0) {
1375                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1376                 GOTO(out_thread, rc);
1377         }
1378
1379         wait_event(expired_lock_thread.elt_waitq,
1380                    expired_lock_thread.elt_state == ELT_READY);
1381
1382         INIT_LIST_HEAD(&waiting_locks_list);
1383         spin_lock_init(&waiting_locks_spinlock);
1384         waiting_locks_timer.function = waiting_locks_callback;
1385         waiting_locks_timer.data = 0;
1386         init_timer(&waiting_locks_timer);
1387 #endif
1388
1389         RETURN(0);
1390
1391 #ifdef __KERNEL__
1392  out_thread:
1393         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1394         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1395 #endif
1396
1397  out_proc:
1398 #ifdef __KERNEL__
1399         ldlm_proc_cleanup();
1400  out_free:
1401 #endif
1402         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1403         ldlm_state = NULL;
1404         return rc;
1405 }
1406
1407 static int ldlm_cleanup(int force)
1408 {
1409 #ifdef __KERNEL__
1410         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1411 #endif
1412         ENTRY;
1413
1414         if (!list_empty(&ldlm_namespace_list)) {
1415                 CERROR("ldlm still has namespaces; clean these up first.\n");
1416                 ldlm_dump_all_namespaces();
1417                 RETURN(-EBUSY);
1418         }
1419
1420 #ifdef __KERNEL__
1421         while (atomic_read(&blp->blp_num_threads) > 0) {
1422                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1423
1424                 init_completion(&blp->blp_comp);
1425
1426                 spin_lock(&blp->blp_lock);
1427                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1428                 wake_up(&blp->blp_waitq);
1429                 spin_unlock(&blp->blp_lock);
1430
1431                 wait_for_completion(&blp->blp_comp);
1432         }
1433         OBD_FREE(blp, sizeof(*blp));
1434
1435         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1436         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1437         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1438         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1439         ldlm_proc_cleanup();
1440
1441         expired_lock_thread.elt_state = ELT_TERMINATE;
1442         wake_up(&expired_lock_thread.elt_waitq);
1443         wait_event(expired_lock_thread.elt_waitq,
1444                    expired_lock_thread.elt_state == ELT_STOPPED);
1445
1446 #endif
1447
1448         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1449         ldlm_state = NULL;
1450
1451         RETURN(0);
1452 }
1453
1454 int __init ldlm_init(void)
1455 {
1456         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1457                                                sizeof(struct ldlm_resource), 0,
1458                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1459         if (ldlm_resource_slab == NULL)
1460                 return -ENOMEM;
1461
1462         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1463                                            sizeof(struct ldlm_lock), 0,
1464                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1465         if (ldlm_lock_slab == NULL) {
1466                 kmem_cache_destroy(ldlm_resource_slab);
1467                 return -ENOMEM;
1468         }
1469
1470         l_lock_init(&ldlm_handle_lock);
1471
1472         return 0;
1473 }
1474
1475 void __exit ldlm_exit(void)
1476 {
1477         if ( ldlm_refcount )
1478                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1479         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
1480                 CERROR("couldn't free ldlm resource slab\n");
1481         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
1482                 CERROR("couldn't free ldlm lock slab\n");
1483 }
1484
1485 /* ldlm_flock.c */
1486 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1487
1488 /* ldlm_extent.c */
1489 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1490
1491 /* ldlm_lock.c */
1492 EXPORT_SYMBOL(ldlm_get_processing_policy);
1493 EXPORT_SYMBOL(ldlm_lock2desc);
1494 EXPORT_SYMBOL(ldlm_register_intent);
1495 EXPORT_SYMBOL(ldlm_lockname);
1496 EXPORT_SYMBOL(ldlm_typename);
1497 EXPORT_SYMBOL(ldlm_lock2handle);
1498 EXPORT_SYMBOL(__ldlm_handle2lock);
1499 EXPORT_SYMBOL(ldlm_lock_get);
1500 EXPORT_SYMBOL(ldlm_lock_put);
1501 EXPORT_SYMBOL(ldlm_lock_match);
1502 EXPORT_SYMBOL(ldlm_lock_cancel);
1503 EXPORT_SYMBOL(ldlm_lock_addref);
1504 EXPORT_SYMBOL(ldlm_lock_decref);
1505 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1506 EXPORT_SYMBOL(ldlm_lock_change_resource);
1507 EXPORT_SYMBOL(ldlm_lock_set_data);
1508 EXPORT_SYMBOL(ldlm_it2str);
1509 EXPORT_SYMBOL(ldlm_lock_dump);
1510 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1511 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1512 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1513 EXPORT_SYMBOL(ldlm_lock_allow_match);
1514
1515 /* ldlm_request.c */
1516 EXPORT_SYMBOL(ldlm_completion_ast);
1517 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1518 EXPORT_SYMBOL(ldlm_cli_convert);
1519 EXPORT_SYMBOL(ldlm_cli_enqueue);
1520 EXPORT_SYMBOL(ldlm_cli_cancel);
1521 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1522 EXPORT_SYMBOL(ldlm_replay_locks);
1523 EXPORT_SYMBOL(ldlm_resource_foreach);
1524 EXPORT_SYMBOL(ldlm_namespace_foreach);
1525 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1526 EXPORT_SYMBOL(ldlm_change_cbdata);
1527
1528 /* ldlm_lockd.c */
1529 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1530 EXPORT_SYMBOL(ldlm_server_completion_ast);
1531 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1532 EXPORT_SYMBOL(ldlm_handle_enqueue);
1533 EXPORT_SYMBOL(ldlm_handle_cancel);
1534 EXPORT_SYMBOL(ldlm_handle_convert);
1535 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1536 EXPORT_SYMBOL(ldlm_get_ref);
1537 EXPORT_SYMBOL(ldlm_put_ref);
1538
1539 #if 0
1540 /* ldlm_test.c */
1541 EXPORT_SYMBOL(ldlm_test);
1542 EXPORT_SYMBOL(ldlm_regression_start);
1543 EXPORT_SYMBOL(ldlm_regression_stop);
1544 #endif
1545
1546 /* ldlm_resource.c */
1547 EXPORT_SYMBOL(ldlm_namespace_new);
1548 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1549 EXPORT_SYMBOL(ldlm_namespace_free);
1550 EXPORT_SYMBOL(ldlm_namespace_dump);
1551 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1552 EXPORT_SYMBOL(ldlm_resource_get);
1553 EXPORT_SYMBOL(ldlm_resource_putref);
1554
1555 /* l_lock.c */
1556 EXPORT_SYMBOL(l_lock);
1557 EXPORT_SYMBOL(l_unlock);
1558
1559 /* ldlm_lib.c */
1560 EXPORT_SYMBOL(client_obd_setup);
1561 EXPORT_SYMBOL(client_obd_cleanup);
1562 EXPORT_SYMBOL(client_connect_import);
1563 EXPORT_SYMBOL(client_disconnect_export);
1564 EXPORT_SYMBOL(target_start_recovery_thread);
1565 EXPORT_SYMBOL(target_stop_recovery_thread);
1566 EXPORT_SYMBOL(target_handle_connect);
1567 EXPORT_SYMBOL(target_destroy_export);
1568 EXPORT_SYMBOL(target_cancel_recovery_timer);
1569 EXPORT_SYMBOL(target_send_reply);
1570 EXPORT_SYMBOL(target_queue_recovery_request);
1571 EXPORT_SYMBOL(target_handle_ping);
1572 EXPORT_SYMBOL(target_handle_disconnect);
1573 EXPORT_SYMBOL(target_queue_final_reply);