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 <libcfs/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 list_head ldlm_namespace_list;
  46
  47 static DECLARE_MUTEX(ldlm_ref_sem);
  48 static int ldlm_refcount;
  49 /* LDLM state */
  50
  51 static struct ldlm_state *ldlm_state;
  52
  53 static inline unsigned long
  54 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 int ldlm_add_waiting_lock(struct ldlm_lock *lock);
  99
 100 static inline int have_expired_locks(void)
 101 {
 102         int need_to_run;
 103
 104         spin_lock_bh(&expired_lock_thread.elt_lock);
 105         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 106         spin_unlock_bh(&expired_lock_thread.elt_lock);
 107
 108         RETURN(need_to_run);
 109 }
 110
 111 static int expired_lock_main(void *arg)
 112 {
 113         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 114         struct l_wait_info lwi = { 0 };
 115         unsigned long flags;
 116
 117         ENTRY;
 118         lock_kernel();
 119         kportal_daemonize("ldlm_elt");
 120
 121         SIGNAL_MASK_LOCK(current, flags);
 122         sigfillset(&current->blocked);
 123         RECALC_SIGPENDING;
 124         SIGNAL_MASK_UNLOCK(current, flags);
 125
 126         unlock_kernel();
 127
 128         expired_lock_thread.elt_state = ELT_READY;
 129         wake_up(&expired_lock_thread.elt_waitq);
 130
 131         while (1) {
 132                 l_wait_event(expired_lock_thread.elt_waitq,
 133                              have_expired_locks() ||
 134                              expired_lock_thread.elt_state == ELT_TERMINATE,
 135                              &lwi);
 136
 137                 spin_lock_bh(&expired_lock_thread.elt_lock);
 138                 while (!list_empty(expired)) {
 139                         struct obd_export *export;
 140                         struct ldlm_lock *lock;
 141
 142                         lock = list_entry(expired->next, struct ldlm_lock,
 143                                           l_pending_chain);
 144                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
 145                             (void *)lock >= LP_POISON) {
 146                                 CERROR("free lock on elt list %p\n", lock);
 147                                 LBUG();
 148                         }
 149                         list_del_init(&lock->l_pending_chain);
 150                         if ((void *)lock->l_export < LP_POISON + PAGE_SIZE &&
 151                             (void *)lock->l_export >= LP_POISON + PAGE_SIZE) {
 152                                 CERROR("lock with free export on elt list %p\n",
 153                                        export);
 154                                 lock->l_export = NULL;
 155                                 LDLM_ERROR(lock, "free export\n");
 156                                 continue;
 157                         }
 158                         export = class_export_get(lock->l_export);
 159                         spin_unlock_bh(&expired_lock_thread.elt_lock);
 160
 161                         ptlrpc_fail_export(export);
 162                         class_export_put(export);
 163                         spin_lock_bh(&expired_lock_thread.elt_lock);
 164                 }
 165                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 166
 167                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 168                         break;
 169         }
 170
 171         expired_lock_thread.elt_state = ELT_STOPPED;
 172         wake_up(&expired_lock_thread.elt_waitq);
 173         RETURN(0);
 174 }
 175
 176 static void waiting_locks_callback(unsigned long unused)
 177 {
 178         struct ldlm_lock *lock;
 179         char str[PTL_NALFMT_SIZE];
 180
 181         if (obd_dump_on_timeout)
 182                 portals_debug_dumplog();
 183
 184 repeat:
 185         spin_lock_bh(&waiting_locks_spinlock);
 186         while (!list_empty(&waiting_locks_list)) {
 187                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 188                                   l_pending_chain);
 189
 190                 if ((lock->l_callback_timeout > jiffies) ||
 191                     (lock->l_req_mode == LCK_GROUP))
 192                         break;
 193
 194                 if (ptlrpc_check_suspend()) {
 195                         /* there is a case when we talk to one mds, holding
 196                          * lock from another mds. this way we easily can get
 197                          * here, if second mds is being recovered. so, we
 198                          * suspend timeouts. bug 6019 */
 199
 200                         LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
 201                                    lock->l_export->exp_client_uuid.uuid,
 202                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
 203                                    ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str));
 204
 205                         list_del_init(&lock->l_pending_chain);
 206                         spin_unlock_bh(&waiting_locks_spinlock);
 207                         ldlm_add_waiting_lock(lock);
 208
 209                         goto repeat;
 210                 }
 211
 212                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 213                            "%s@%s nid %s ",
 214                            lock->l_export->exp_client_uuid.uuid,
 215                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 216                            ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str));
 217
 218                 spin_lock_bh(&expired_lock_thread.elt_lock);
 219                 list_del(&lock->l_pending_chain);
 220                 list_add(&lock->l_pending_chain,
 221                          &expired_lock_thread.elt_expired_locks);
 222                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 223                 wake_up(&expired_lock_thread.elt_waitq);
 224         }
 225
 226         /*
 227          * Make sure the timer will fire again if we have any locks
 228          * left.
 229          */
 230         if (!list_empty(&waiting_locks_list)) {
 231                 unsigned long timeout_rounded;
 232                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 233                                   l_pending_chain);
 234                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 235                 mod_timer(&waiting_locks_timer, timeout_rounded);
 236         }
 237         spin_unlock_bh(&waiting_locks_spinlock);
 238 }
 239
 240 /*
 241  * Indicate that we're waiting for a client to call us back cancelling a given
 242  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 243  * timer to fire appropriately.  (We round up to the next second, to avoid
 244  * floods of timer firings during periods of high lock contention and traffic).
 245  */
 246 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 247 {
 248         unsigned long timeout_rounded;
 249
 250         spin_lock_bh(&waiting_locks_spinlock);
 251         if (!list_empty(&lock->l_pending_chain)) {
 252                 LDLM_DEBUG(lock, "not re-adding to wait list");
 253                 spin_unlock_bh(&waiting_locks_spinlock);
 254                 return 0;
 255         }
 256         LDLM_DEBUG(lock, "adding to wait list");
 257
 258         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 259
 260         timeout_rounded = round_timeout(lock->l_callback_timeout);
 261
 262         if (timeout_rounded < waiting_locks_timer.expires ||
 263             !timer_pending(&waiting_locks_timer)) {
 264                 mod_timer(&waiting_locks_timer, timeout_rounded);
 265         }
 266         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 267         spin_unlock_bh(&waiting_locks_spinlock);
 268         return 1;
 269 }
 270
 271 /*
 272  * Remove a lock from the pending list, likely because it had its cancellation
 273  * callback arrive without incident.  This adjusts the lock-timeout timer if
 274  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 275  */
 276 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 277 {
 278         struct list_head *list_next;
 279
 280         if (lock->l_export == NULL) {
 281                 /* We don't have a "waiting locks list" on clients. */
 282                 LDLM_DEBUG(lock, "client lock: no-op");
 283                 return 0;
 284         }
 285
 286         spin_lock_bh(&waiting_locks_spinlock);
 287
 288         if (list_empty(&lock->l_pending_chain)) {
 289                 spin_unlock_bh(&waiting_locks_spinlock);
 290                 LDLM_DEBUG(lock, "wasn't waiting");
 291                 return 0;
 292         }
 293
 294         list_next = lock->l_pending_chain.next;
 295         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 296                 /* Removing the head of the list, adjust timer. */
 297                 if (list_next == &waiting_locks_list) {
 298                         /* No more, just cancel. */
 299                         del_timer(&waiting_locks_timer);
 300                 } else {
 301                         struct ldlm_lock *next;
 302                         next = list_entry(list_next, struct ldlm_lock,
 303                                           l_pending_chain);
 304                         mod_timer(&waiting_locks_timer,
 305                                   round_timeout(next->l_callback_timeout));
 306                 }
 307         }
 308
 309         spin_lock_bh(&expired_lock_thread.elt_lock);
 310         list_del_init(&lock->l_pending_chain);
 311         spin_unlock_bh(&expired_lock_thread.elt_lock);
 312
 313         spin_unlock_bh(&waiting_locks_spinlock);
 314         LDLM_DEBUG(lock, "removed");
 315         return 1;
 316 }
 317
 318 #else /* !__KERNEL__ */
 319
 320 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 321 {
 322         RETURN(1);
 323 }
 324
 325 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 326 {
 327         RETURN(0);
 328 }
 329
 330 #endif /* __KERNEL__ */
 331
 332 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,const char *ast_type)
 333 {
 334         struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 335         char str[PTL_NALFMT_SIZE];
 336
 337         LDLM_ERROR(lock, "%s AST failed (%d): evicting client %s@%s NID "LPX64
 338                    " (%s)", ast_type, rc, lock->l_export->exp_client_uuid.uuid,
 339                    conn->c_remote_uuid.uuid, conn->c_peer.peer_id.nid,
 340                    ptlrpc_peernid2str(&conn->c_peer, str));
 341
 342         if (obd_dump_on_timeout)
 343                 portals_debug_dumplog();
 344         ptlrpc_fail_export(lock->l_export);
 345 }
 346
 347 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 348                                  struct ptlrpc_request *req, int rc,
 349                                  const char *ast_type)
 350 {
 351         struct ptlrpc_peer *peer = &req->rq_import->imp_connection->c_peer;
 352         char str[PTL_NALFMT_SIZE];
 353
 354         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 355                 LASSERT(lock->l_export);
 356                 if (lock->l_export->exp_libclient) {
 357                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 358                                    " timeout, just cancelling lock", ast_type,
 359                                    ptlrpc_peernid2str(peer, str));
 360                         ldlm_lock_cancel(lock);
 361                         rc = -ERESTART;
 362                 } else {
 363                         ldlm_del_waiting_lock(lock);
 364                         ldlm_failed_ast(lock, rc, ast_type);
 365                 }
 366         } else if (rc) {
 367                 if (rc == -EINVAL)
 368                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 369                                    " from %s AST - normal race",
 370                                    ptlrpc_peernid2str(peer, str),
 371                                    req->rq_repmsg->status, ast_type);
 372                 else
 373                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 374                                    "from %s AST", ptlrpc_peernid2str(peer, str),
 375                                    (req->rq_repmsg != NULL) ?
 376                                    req->rq_repmsg->status : 0, ast_type);
 377                 ldlm_lock_cancel(lock);
 378                 /* Server-side AST functions are called from ldlm_reprocess_all,
 379                  * which needs to be told to please restart its reprocessing. */
 380                 rc = -ERESTART;
 381         }
 382
 383         return rc;
 384 }
 385
 386 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 387                              struct ldlm_lock_desc *desc,
 388                              void *data, int flag)
 389 {
 390         struct ldlm_request *body;
 391         struct ptlrpc_request *req;
 392         int rc = 0, size = sizeof(*body);
 393         ENTRY;
 394
 395         if (flag == LDLM_CB_CANCELING) {
 396                 /* Don't need to do anything here. */
 397                 RETURN(0);
 398         }
 399
 400         LASSERTF(lock->l_export->exp_obd->obd_recovering == 0,
 401                  "BUG 6063: lock collide during recovery");
 402
 403         LASSERT(lock);
 404
 405 #if 0
 406         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 407                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 408                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 409                 RETURN(-ETIMEDOUT);
 410         }
 411 #endif
 412
 413         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 414                               LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK,
 415                               1, &size, NULL);
 416         if (req == NULL)
 417                 RETURN(-ENOMEM);
 418
 419         lock_res_and_lock(lock);
 420         if (lock->l_granted_mode != lock->l_req_mode) {
 421                 /* this blocking AST will be communicated as part of the
 422                  * completion AST instead */
 423                 unlock_res_and_lock(lock);
 424                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 425                 ptlrpc_req_finished(req);
 426                 RETURN(0);
 427         }
 428
 429         if (lock->l_destroyed) {
 430                 /* What's the point? */
 431                 unlock_res_and_lock(lock);
 432                 ptlrpc_req_finished(req);
 433                 RETURN(0);
 434         }
 435
 436         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 437         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 438                sizeof(body->lock_handle1));
 439         memcpy(&body->lock_desc, desc, sizeof(*desc));
 440         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 441
 442         LDLM_DEBUG(lock, "server preparing blocking AST");
 443         req->rq_replen = lustre_msg_size(0, NULL);
 444
 445         if (lock->l_granted_mode == lock->l_req_mode)
 446                 ldlm_add_waiting_lock(lock);
 447         unlock_res_and_lock(lock);
 448
 449         req->rq_send_state = LUSTRE_IMP_FULL;
 450         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 451         rc = ptlrpc_queue_wait(req);
 452         if (rc != 0)
 453                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 454
 455         ptlrpc_req_finished(req);
 456
 457         RETURN(rc);
 458 }
 459
 460 /* XXX copied from ptlrpc/service.c */
 461 static long timeval_sub(struct timeval *large, struct timeval *small)
 462 {
 463         return (large->tv_sec - small->tv_sec) * 1000000 +
 464                 (large->tv_usec - small->tv_usec);
 465 }
 466
 467 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 468 {
 469         struct ldlm_request *body;
 470         struct ptlrpc_request *req;
 471         struct timeval granted_time;
 472         long total_enqueue_wait;
 473         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 474         ENTRY;
 475
 476         LASSERT(lock != NULL);
 477
 478         do_gettimeofday(&granted_time);
 479         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 480
 481         if (total_enqueue_wait / 1000000 > obd_timeout)
 482                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 483
 484         lock_res_and_lock(lock);
 485         if (lock->l_resource->lr_lvb_len) {
 486                 buffers = 2;
 487                 size[1] = lock->l_resource->lr_lvb_len;
 488         }
 489         unlock_res_and_lock(lock);
 490
 491         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 492                               LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK,
 493                               buffers, size, NULL);
 494         if (req == NULL)
 495                 RETURN(-ENOMEM);
 496
 497         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 498         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 499                sizeof(body->lock_handle1));
 500         body->lock_flags = flags;
 501         ldlm_lock2desc(lock, &body->lock_desc);
 502
 503         if (buffers == 2) {
 504                 void *lvb;
 505
 506                 lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 507                                      lock->l_resource->lr_lvb_len);
 508                 lock_res_and_lock(lock);
 509                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 510                        lock->l_resource->lr_lvb_len);
 511                 unlock_res_and_lock(lock);
 512         }
 513
 514         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 515                    total_enqueue_wait);
 516         req->rq_replen = lustre_msg_size(0, NULL);
 517
 518         req->rq_send_state = LUSTRE_IMP_FULL;
 519         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 520
 521         /* We only send real blocking ASTs after the lock is granted */
 522         lock_res_and_lock(lock);
 523         if (lock->l_flags & LDLM_FL_AST_SENT) {
 524                 body->lock_flags |= LDLM_FL_AST_SENT;
 525                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 526         }
 527         unlock_res_and_lock(lock);
 528
 529         rc = ptlrpc_queue_wait(req);
 530         if (rc != 0)
 531                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 532
 533         ptlrpc_req_finished(req);
 534
 535         RETURN(rc);
 536 }
 537
 538 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 539 {
 540         struct ldlm_resource *res;
 541         struct ldlm_request *body;
 542         struct ptlrpc_request *req;
 543         int rc = 0, size = sizeof(*body);
 544         ENTRY;
 545
 546         LASSERT(lock != NULL);
 547
 548         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 549                               LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK,
 550                               1, &size, NULL);
 551         if (req == NULL)
 552                 RETURN(-ENOMEM);
 553
 554         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 555         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 556                sizeof(body->lock_handle1));
 557         ldlm_lock2desc(lock, &body->lock_desc);
 558
 559         lock_res_and_lock(lock);
 560         size = lock->l_resource->lr_lvb_len;
 561         unlock_res_and_lock(lock);
 562         res = lock->l_resource;
 563         req->rq_replen = lustre_msg_size(1, &size);
 564
 565         req->rq_send_state = LUSTRE_IMP_FULL;
 566         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 567
 568         rc = ptlrpc_queue_wait(req);
 569         if (rc == -ELDLM_NO_LOCK_DATA)
 570                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 571         else if (rc != 0)
 572                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 573         else
 574                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 575                         (res, req->rq_repmsg, 0, 1);
 576         ptlrpc_req_finished(req);
 577         RETURN(rc);
 578 }
 579
 580 static struct ldlm_lock *
 581 find_existing_lock(struct obd_export *exp, struct lustre_handle *remote_hdl)
 582 {
 583         struct list_head *iter;
 584
 585         spin_lock(&exp->exp_ldlm_data.led_lock);
 586         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
 587                 struct ldlm_lock *lock;
 588                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 589                 if (lock->l_remote_handle.cookie == remote_hdl->cookie) {
 590                         LDLM_LOCK_GET(lock);
 591                         spin_unlock(&exp->exp_ldlm_data.led_lock);
 592                         return lock;
 593                 }
 594         }
 595         spin_unlock(&exp->exp_ldlm_data.led_lock);
 596         return NULL;
 597 }
 598
 599 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 600                         ldlm_completion_callback completion_callback,
 601                         ldlm_blocking_callback blocking_callback,
 602                         ldlm_glimpse_callback glimpse_callback)
 603 {
 604         struct obd_device *obddev = req->rq_export->exp_obd;
 605         struct ldlm_reply *dlm_rep;
 606         struct ldlm_request *dlm_req;
 607         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 608         __u32 flags;
 609         ldlm_error_t err = ELDLM_OK;
 610         struct ldlm_lock *lock = NULL;
 611         void *cookie = NULL;
 612         ENTRY;
 613
 614         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 615
 616         dlm_req = lustre_swab_reqbuf(req, MDS_REQ_INTENT_LOCKREQ_OFF,
 617                                      sizeof (*dlm_req),
 618                                      lustre_swab_ldlm_request);
 619         if (dlm_req == NULL) {
 620                 CERROR ("Can't unpack dlm_req\n");
 621                 GOTO(out, rc = -EFAULT);
 622         }
 623
 624         flags = dlm_req->lock_flags;
 625
 626         LASSERT(req->rq_export);
 627
 628         if (flags & LDLM_FL_REPLAY) {
 629                 lock = find_existing_lock(req->rq_export,
 630                                           &dlm_req->lock_handle1);
 631                 if (lock != NULL) {
 632                         DEBUG_REQ(D_HA, req, "found existing lock cookie "LPX64,
 633                                   lock->l_handle.h_cookie);
 634                         GOTO(existing_lock, rc = 0);
 635                 }
 636         }
 637
 638         /* The lock's callback data might be set in the policy function */
 639         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 640                                 dlm_req->lock_desc.l_resource.lr_name,
 641                                 dlm_req->lock_desc.l_resource.lr_type,
 642                                 dlm_req->lock_desc.l_req_mode,
 643                                 blocking_callback, completion_callback,
 644                                 glimpse_callback, NULL, 0);
 645         if (!lock)
 646                 GOTO(out, rc = -ENOMEM);
 647
 648         do_gettimeofday(&lock->l_enqueued_time);
 649         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 650                sizeof(lock->l_remote_handle));
 651         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 652
 653         LASSERT(req->rq_export);
 654         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
 655         if (req->rq_export->exp_failed) {
 656                 LDLM_ERROR(lock,"lock on destroyed export %p\n",req->rq_export);
 657                 GOTO(out, err = -ENOTCONN);
 658         }
 659
 660         lock->l_export = class_export_get(req->rq_export);
 661         spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
 662         list_add(&lock->l_export_chain,
 663                  &lock->l_export->exp_ldlm_data.led_held_locks);
 664         spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
 665
 666 existing_lock:
 667
 668         if (flags & LDLM_FL_HAS_INTENT) {
 669                 /* In this case, the reply buffer is allocated deep in
 670                  * local_lock_enqueue by the policy function. */
 671                 cookie = req;
 672         } else {
 673                 int buffers = 1;
 674                 lock_res_and_lock(lock);
 675                 if (lock->l_resource->lr_lvb_len) {
 676                         size[1] = lock->l_resource->lr_lvb_len;
 677                         buffers = 2;
 678                 }
 679                 unlock_res_and_lock(lock);
 680                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 681                         GOTO(out, rc = -ENOMEM);
 682
 683                 rc = lustre_pack_reply(req, buffers, size, NULL);
 684                 if (rc)
 685                         GOTO(out, rc);
 686         }
 687
 688         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 689                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 690                        sizeof(ldlm_policy_data_t));
 691         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 692                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 693                        sizeof(lock->l_req_extent));
 694
 695         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock,
 696                                 cookie, (int *)&flags);
 697         if (err)
 698                 GOTO(out, err);
 699
 700         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*dlm_rep));
 701         dlm_rep->lock_flags = flags;
 702
 703         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 704         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 705
 706         /* We never send a blocking AST until the lock is granted, but
 707          * we can tell it right now */
 708         lock_res_and_lock(lock);
 709         if (lock->l_flags & LDLM_FL_AST_SENT) {
 710                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 711                 if (lock->l_granted_mode == lock->l_req_mode)
 712                         ldlm_add_waiting_lock(lock);
 713         }
 714         unlock_res_and_lock(lock);
 715
 716         EXIT;
 717  out:
 718         req->rq_status = err;
 719         if (req->rq_reply_state == NULL) {
 720                 err = lustre_pack_reply(req, 0, NULL, NULL);
 721                 if (rc == 0)
 722                         rc = err;
 723                 req->rq_status = rc;
 724         }
 725
 726         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 727          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 728         if (lock) {
 729                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 730                            "(err=%d, rc=%d)", err, rc);
 731
 732                 if (rc == 0) {
 733                         lock_res_and_lock(lock);
 734                         size[1] = lock->l_resource->lr_lvb_len;
 735                         if (size[1] > 0) {
 736                                 void *lvb = lustre_msg_buf(req->rq_repmsg,
 737                                                            1, size[1]);
 738                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
 739                                          req, lock);
 740
 741                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 742                                        size[1]);
 743                         }
 744                         unlock_res_and_lock(lock);
 745                 } else {
 746                         ldlm_lock_destroy(lock);
 747                 }
 748
 749                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 750                         ldlm_reprocess_all(lock->l_resource);
 751                 LDLM_LOCK_PUT(lock);
 752         }
 753         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 754                           lock, rc);
 755
 756         return rc;
 757 }
 758
 759 int ldlm_handle_convert(struct ptlrpc_request *req)
 760 {
 761         struct ldlm_request *dlm_req;
 762         struct ldlm_reply *dlm_rep;
 763         struct ldlm_lock *lock;
 764         int rc, size = sizeof(*dlm_rep);
 765         ENTRY;
 766
 767         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 768                                      lustre_swab_ldlm_request);
 769         if (dlm_req == NULL) {
 770                 CERROR ("Can't unpack dlm_req\n");
 771                 RETURN (-EFAULT);
 772         }
 773
 774         rc = lustre_pack_reply(req, 1, &size, NULL);
 775         if (rc) {
 776                 CERROR("out of memory\n");
 777                 RETURN(-ENOMEM);
 778         }
 779         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 780         dlm_rep->lock_flags = dlm_req->lock_flags;
 781
 782         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 783         if (!lock) {
 784                 req->rq_status = EINVAL;
 785         } else {
 786                 void *res = NULL;
 787
 788                 LDLM_DEBUG(lock, "server-side convert handler START");
 789
 790                 do_gettimeofday(&lock->l_enqueued_time);
 791                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 792                                         (int *)&dlm_rep->lock_flags);
 793                 if (res) {
 794                         if (ldlm_del_waiting_lock(lock))
 795                                 LDLM_DEBUG(lock, "converted waiting lock");
 796                         req->rq_status = 0;
 797                 } else {
 798                         req->rq_status = EDEADLOCK;
 799                 }
 800         }
 801
 802         if (lock) {
 803                 if (!req->rq_status)
 804                         ldlm_reprocess_all(lock->l_resource);
 805                 LDLM_DEBUG(lock, "server-side convert handler END");
 806                 LDLM_LOCK_PUT(lock);
 807         } else
 808                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 809
 810         RETURN(0);
 811 }
 812
 813 int ldlm_handle_cancel(struct ptlrpc_request *req)
 814 {
 815         struct ldlm_request *dlm_req;
 816         struct ldlm_lock *lock;
 817         struct ldlm_resource *res;
 818         int rc;
 819         ENTRY;
 820
 821         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 822                                       lustre_swab_ldlm_request);
 823         if (dlm_req == NULL) {
 824                 CERROR("bad request buffer for cancel\n");
 825                 RETURN(-EFAULT);
 826         }
 827
 828         rc = lustre_pack_reply(req, 0, NULL, NULL);
 829         if (rc) {
 830                 CERROR("out of memory\n");
 831                 RETURN(-ENOMEM);
 832         }
 833
 834         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 835         if (!lock) {
 836                 CERROR("received cancel for unknown lock cookie "LPX64
 837                        " from client %s id %s\n",
 838                        dlm_req->lock_handle1.cookie,
 839                        req->rq_export->exp_client_uuid.uuid,
 840                        req->rq_peerstr);
 841                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 842                                   "(cookie "LPU64")",
 843                                   dlm_req->lock_handle1.cookie);
 844                 req->rq_status = ESTALE;
 845         } else {
 846                 LDLM_DEBUG(lock, "server-side cancel handler START");
 847                 res = lock->l_resource;
 848                 if (res && res->lr_namespace->ns_lvbo &&
 849                     res->lr_namespace->ns_lvbo->lvbo_update) {
 850                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 851                                 (res, NULL, 0, 0);
 852                                 //(res, req->rq_reqmsg, 1);
 853                 }
 854
 855                 ldlm_lock_cancel(lock);
 856                 if (ldlm_del_waiting_lock(lock))
 857                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 858                 req->rq_status = rc;
 859         }
 860
 861         if (ptlrpc_reply(req) != 0)
 862                 LBUG();
 863
 864         if (lock) {
 865                 ldlm_reprocess_all(lock->l_resource);
 866                 LDLM_DEBUG(lock, "server-side cancel handler END");
 867                 LDLM_LOCK_PUT(lock);
 868         }
 869
 870         RETURN(0);
 871 }
 872
 873 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 874                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 875 {
 876         int do_ast;
 877         ENTRY;
 878
 879         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 880
 881         lock_res_and_lock(lock);
 882         lock->l_flags |= LDLM_FL_CBPENDING;
 883         do_ast = (!lock->l_readers && !lock->l_writers);
 884         unlock_res_and_lock(lock);
 885
 886         if (do_ast) {
 887                 LDLM_DEBUG(lock, "already unused, calling "
 888                            "callback (%p)", lock->l_blocking_ast);
 889                 if (lock->l_blocking_ast != NULL)
 890                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 891                                              LDLM_CB_BLOCKING);
 892         } else {
 893                 LDLM_DEBUG(lock, "Lock still has references, will be"
 894                            " cancelled later");
 895         }
 896
 897         LDLM_DEBUG(lock, "client blocking callback handler END");
 898         LDLM_LOCK_PUT(lock);
 899         EXIT;
 900 }
 901
 902 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 903                                     struct ldlm_namespace *ns,
 904                                     struct ldlm_request *dlm_req,
 905                                     struct ldlm_lock *lock)
 906 {
 907         LIST_HEAD(ast_list);
 908         ENTRY;
 909
 910         LDLM_DEBUG(lock, "client completion callback handler START");
 911
 912         lock_res_and_lock(lock);
 913
 914         /* If we receive the completion AST before the actual enqueue returned,
 915          * then we might need to switch lock modes, resources, or extents. */
 916         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 917                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 918                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 919         }
 920
 921         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 922                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 923                        sizeof(lock->l_policy_data));
 924                 LDLM_DEBUG(lock, "completion AST, new policy data");
 925         }
 926
 927         ldlm_resource_unlink_lock(lock);
 928         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 929                    &lock->l_resource->lr_name,
 930                    sizeof(lock->l_resource->lr_name)) != 0) {
 931                 unlock_res_and_lock(lock);
 932                 ldlm_lock_change_resource(ns, lock,
 933                                          dlm_req->lock_desc.l_resource.lr_name);
 934                 LDLM_DEBUG(lock, "completion AST, new resource");
 935                 CERROR("change resource!\n");
 936                 lock_res_and_lock(lock);
 937         }
 938
 939         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 940                 lock->l_flags |= LDLM_FL_CBPENDING;
 941                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 942         }
 943
 944         if (lock->l_lvb_len) {
 945                 void *lvb;
 946                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 947                                          lock->l_lvb_swabber);
 948                 if (lvb == NULL) {
 949                         LDLM_ERROR(lock, "completion AST did not contain "
 950                                    "expected LVB!");
 951                 } else {
 952                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 953                 }
 954         }
 955
 956         ldlm_grant_lock(lock, &ast_list);
 957         unlock_res_and_lock(lock);
 958
 959         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 960         LDLM_LOCK_PUT(lock);
 961
 962         ldlm_run_cp_ast_work(&ast_list);
 963
 964         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 965                           lock);
 966         EXIT;
 967 }
 968
 969 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 970                                     struct ldlm_namespace *ns,
 971                                     struct ldlm_request *dlm_req,
 972                                     struct ldlm_lock *lock)
 973 {
 974         int rc = -ENOSYS;
 975         ENTRY;
 976
 977         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 978
 979         if (lock->l_glimpse_ast != NULL)
 980                 rc = lock->l_glimpse_ast(lock, req);
 981
 982         if (req->rq_repmsg != NULL) {
 983                 ptlrpc_reply(req);
 984         } else {
 985                 req->rq_status = rc;
 986                 ptlrpc_error(req);
 987         }
 988
 989         lock_res_and_lock(lock);
 990         if (lock->l_granted_mode == LCK_PW &&
 991             !lock->l_readers && !lock->l_writers &&
 992             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 993                 unlock_res_and_lock(lock);
 994                 if (ldlm_bl_to_thread(ns, NULL, lock))
 995                         ldlm_handle_bl_callback(ns, NULL, lock);
 996
 997                 EXIT;
 998                 return;
 999         }
1000         unlock_res_and_lock(lock);
1001         LDLM_LOCK_PUT(lock);
1002         EXIT;
1003 }
1004
1005 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1006 {
1007         req->rq_status = rc;
1008         if (req->rq_reply_state == NULL) {
1009                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1010                 if (rc)
1011                         return rc;
1012         }
1013         return ptlrpc_reply(req);
1014 }
1015
1016 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1017                       struct ldlm_lock *lock)
1018 {
1019 #ifdef __KERNEL__
1020         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1021         struct ldlm_bl_work_item *blwi;
1022         ENTRY;
1023
1024         OBD_ALLOC(blwi, sizeof(*blwi));
1025         if (blwi == NULL)
1026                 RETURN(-ENOMEM);
1027
1028         blwi->blwi_ns = ns;
1029         if (ld != NULL)
1030                 blwi->blwi_ld = *ld;
1031         blwi->blwi_lock = lock;
1032
1033         spin_lock(&blp->blp_lock);
1034         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1035         wake_up(&blp->blp_waitq);
1036         spin_unlock(&blp->blp_lock);
1037
1038         RETURN(0);
1039 #else
1040         RETURN(-ENOSYS);
1041 #endif
1042
1043 }
1044
1045 static int ldlm_msg_check_version(struct lustre_msg *msg)
1046 {
1047         int rc;
1048
1049         switch (msg->opc) {
1050         case LDLM_ENQUEUE:
1051         case LDLM_CONVERT:
1052         case LDLM_CANCEL:
1053         case LDLM_BL_CALLBACK:
1054         case LDLM_CP_CALLBACK:
1055         case LDLM_GL_CALLBACK:
1056                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1057                 if (rc)
1058                         CERROR("bad opc %u version %08x, expecting %08x\n",
1059                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
1060                 break;
1061         case OBD_LOG_CANCEL:
1062         case LLOG_ORIGIN_HANDLE_OPEN:
1063         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1064         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1065         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1066         case LLOG_ORIGIN_HANDLE_CLOSE:
1067         case LLOG_CATINFO:
1068                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1069                 if (rc)
1070                         CERROR("bad opc %u version %08x, expecting %08x\n",
1071                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
1072                 break;
1073         default:
1074                 CERROR("LDLM unknown opcode %d\n", msg->opc);
1075                 rc = -ENOTSUPP;
1076                 break;
1077         }
1078
1079         return rc;
1080 }
1081
1082 static int ldlm_callback_handler(struct ptlrpc_request *req)
1083 {
1084         struct ldlm_namespace *ns;
1085         struct ldlm_request *dlm_req;
1086         struct ldlm_lock *lock;
1087         int rc;
1088         ENTRY;
1089
1090         rc = ldlm_msg_check_version(req->rq_reqmsg);
1091         if (rc) {
1092                 CERROR("LDLM_CB drop mal-formed request\n");
1093                 RETURN(rc);
1094         }
1095
1096         /* Requests arrive in sender's byte order.  The ptlrpc service
1097          * handler has already checked and, if necessary, byte-swapped the
1098          * incoming request message body, but I am responsible for the
1099          * message buffers. */
1100
1101         if (req->rq_export == NULL) {
1102                 struct ldlm_request *dlm_req;
1103
1104                 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1105                        "export cookie "LPX64"; this is "
1106                        "normal if this node rebooted with a lock held\n",
1107                        req->rq_reqmsg->opc,
1108                        req->rq_peerstr,
1109                        req->rq_reqmsg->handle.cookie);
1110                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1111                                              lustre_swab_ldlm_request);
1112                 if (dlm_req != NULL)
1113                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1114                                dlm_req->lock_handle1.cookie);
1115
1116                 ldlm_callback_reply(req, -ENOTCONN);
1117                 RETURN(0);
1118         }
1119
1120         LASSERT(req->rq_export != NULL);
1121         LASSERT(req->rq_export->exp_obd != NULL);
1122
1123         switch(req->rq_reqmsg->opc) {
1124         case LDLM_BL_CALLBACK:
1125                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1126                 break;
1127         case LDLM_CP_CALLBACK:
1128                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1129                 break;
1130         case LDLM_GL_CALLBACK:
1131                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1132                 break;
1133         case OBD_LOG_CANCEL:
1134                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1135                 rc = llog_origin_handle_cancel(req);
1136                 ldlm_callback_reply(req, rc);
1137                 RETURN(0);
1138         case LLOG_ORIGIN_HANDLE_OPEN:
1139                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1140                 rc = llog_origin_handle_open(req);
1141                 ldlm_callback_reply(req, rc);
1142                 RETURN(0);
1143         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1144                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1145                 rc = llog_origin_handle_next_block(req);
1146                 ldlm_callback_reply(req, rc);
1147                 RETURN(0);
1148         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1149                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1150                 rc = llog_origin_handle_prev_block(req);
1151                 ldlm_callback_reply(req, rc);
1152                 RETURN(0);
1153         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1154                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1155                 rc = llog_origin_handle_read_header(req);
1156                 ldlm_callback_reply(req, rc);
1157                 RETURN(0);
1158         case LLOG_ORIGIN_HANDLE_CLOSE:
1159                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1160                 rc = llog_origin_handle_close(req);
1161                 ldlm_callback_reply(req, rc);
1162                 RETURN(0);
1163         default:
1164                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1165                 ldlm_callback_reply(req, -EPROTO);
1166                 RETURN(0);
1167         }
1168
1169         ns = req->rq_export->exp_obd->obd_namespace;
1170         LASSERT(ns != NULL);
1171
1172         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1173                                       lustre_swab_ldlm_request);
1174         if (dlm_req == NULL) {
1175                 CERROR ("can't unpack dlm_req\n");
1176                 ldlm_callback_reply (req, -EPROTO);
1177                 RETURN (0);
1178         }
1179
1180         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1181         if (!lock) {
1182                 LDLM_DEBUG_NOLOCK("callback on lock "LPX64" - lock "
1183                                   "disappeared\n",dlm_req->lock_handle1.cookie);
1184                 ldlm_callback_reply(req, -EINVAL);
1185                 RETURN(0);
1186         }
1187
1188         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1189         lock_res_and_lock(lock);
1190         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1191         unlock_res_and_lock(lock);
1192
1193         /* We want the ost thread to get this reply so that it can respond
1194          * to ost requests (write cache writeback) that might be triggered
1195          * in the callback.
1196          *
1197          * But we'd also like to be able to indicate in the reply that we're
1198          * cancelling right now, because it's unused, or have an intent result
1199          * in the reply, so we might have to push the responsibility for sending
1200          * the reply down into the AST handlers, alas. */
1201
1202         switch (req->rq_reqmsg->opc) {
1203         case LDLM_BL_CALLBACK:
1204                 CDEBUG(D_INODE, "blocking ast\n");
1205                 ldlm_callback_reply(req, 0);
1206                 if (ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock))
1207                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1208
1209                 break;
1210         case LDLM_CP_CALLBACK:
1211                 CDEBUG(D_INODE, "completion ast\n");
1212                 ldlm_callback_reply(req, 0);
1213                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1214                 break;
1215         case LDLM_GL_CALLBACK:
1216                 CDEBUG(D_INODE, "glimpse ast\n");
1217                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1218                 break;
1219         default:
1220                 LBUG();                         /* checked above */
1221         }
1222
1223         RETURN(0);
1224 }
1225
1226 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1227 {
1228         int rc = 0;
1229         ENTRY;
1230
1231         rc = ldlm_msg_check_version(req->rq_reqmsg);
1232         if (rc) {
1233                 CERROR("LDLM_CL drop mal-formed request\n");
1234                 RETURN(rc);
1235         }
1236
1237         /* Requests arrive in sender's byte order.  The ptlrpc service
1238          * handler has already checked and, if necessary, byte-swapped the
1239          * incoming request message body, but I am responsible for the
1240          * message buffers. */
1241
1242         if (req->rq_export == NULL) {
1243                 struct ldlm_request *dlm_req;
1244                 CERROR("operation %d with bad export from %s\n",
1245                        req->rq_reqmsg->opc,
1246                        req->rq_peerstr);
1247                 CERROR("--> export cookie: "LPX64"\n",
1248                        req->rq_reqmsg->handle.cookie);
1249                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1250                                              lustre_swab_ldlm_request);
1251                 if (dlm_req != NULL)
1252                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1253                 ldlm_callback_reply(req, -ENOTCONN);
1254                 RETURN(0);
1255         }
1256
1257         switch (req->rq_reqmsg->opc) {
1258
1259         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1260         case LDLM_CANCEL:
1261                 CDEBUG(D_INODE, "cancel\n");
1262                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1263                 rc = ldlm_handle_cancel(req);
1264                 break;
1265         default:
1266                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1267                 ldlm_callback_reply(req, -EINVAL);
1268         }
1269
1270         RETURN(rc);
1271 }
1272
1273 #ifdef __KERNEL__
1274 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1275 {
1276         struct ldlm_bl_work_item *blwi = NULL;
1277
1278         spin_lock(&blp->blp_lock);
1279         if (!list_empty(&blp->blp_list)) {
1280                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1281                                   blwi_entry);
1282                 list_del(&blwi->blwi_entry);
1283         }
1284         spin_unlock(&blp->blp_lock);
1285
1286         return blwi;
1287 }
1288
1289 struct ldlm_bl_thread_data {
1290         int                     bltd_num;
1291         struct ldlm_bl_pool     *bltd_blp;
1292 };
1293
1294 static int ldlm_bl_thread_main(void *arg)
1295 {
1296         struct ldlm_bl_thread_data *bltd = arg;
1297         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1298         unsigned long flags;
1299         ENTRY;
1300
1301         /* XXX boiler-plate */
1302         {
1303                 char name[sizeof(current->comm)];
1304                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1305                          bltd->bltd_num);
1306                 kportal_daemonize(name);
1307         }
1308         SIGNAL_MASK_LOCK(current, flags);
1309         sigfillset(&current->blocked);
1310         RECALC_SIGPENDING;
1311         SIGNAL_MASK_UNLOCK(current, flags);
1312
1313         atomic_inc(&blp->blp_num_threads);
1314         complete(&blp->blp_comp);
1315
1316         while(1) {
1317                 struct l_wait_info lwi = { 0 };
1318                 struct ldlm_bl_work_item *blwi = NULL;
1319
1320                 l_wait_event_exclusive(blp->blp_waitq,
1321                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1322                                        &lwi);
1323
1324                 if (blwi->blwi_ns == NULL)
1325                         break;
1326
1327                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1328                                         blwi->blwi_lock);
1329                 OBD_FREE(blwi, sizeof(*blwi));
1330         }
1331
1332         atomic_dec(&blp->blp_num_threads);
1333         complete(&blp->blp_comp);
1334         RETURN(0);
1335 }
1336
1337 #endif
1338
1339 static int ldlm_setup(void);
1340 static int ldlm_cleanup(int force);
1341
1342 int ldlm_get_ref(void)
1343 {
1344         int rc = 0;
1345         down(&ldlm_ref_sem);
1346         if (++ldlm_refcount == 1) {
1347                 rc = ldlm_setup();
1348                 if (rc)
1349                         ldlm_refcount--;
1350         }
1351         up(&ldlm_ref_sem);
1352
1353         RETURN(rc);
1354 }
1355
1356 void ldlm_put_ref(int force)
1357 {
1358         down(&ldlm_ref_sem);
1359         if (ldlm_refcount == 1) {
1360                 int rc = ldlm_cleanup(force);
1361                 if (rc)
1362                         CERROR("ldlm_cleanup failed: %d\n", rc);
1363                 else
1364                         ldlm_refcount--;
1365         } else {
1366                 ldlm_refcount--;
1367         }
1368         up(&ldlm_ref_sem);
1369
1370         EXIT;
1371 }
1372
1373 static int ldlm_setup(void)
1374 {
1375         struct ldlm_bl_pool *blp;
1376         int rc = 0;
1377 #ifdef __KERNEL__
1378         int i;
1379 #endif
1380         ENTRY;
1381
1382         if (ldlm_state != NULL)
1383                 RETURN(-EALREADY);
1384
1385         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1386         if (ldlm_state == NULL)
1387                 RETURN(-ENOMEM);
1388
1389 #ifdef __KERNEL__
1390         rc = ldlm_proc_setup();
1391         if (rc != 0)
1392                 GOTO(out_free, rc);
1393 #endif
1394
1395         ldlm_state->ldlm_cb_service =
1396                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1397                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1398                                 1500, ldlm_callback_handler, "ldlm_cbd",
1399                                 ldlm_svc_proc_dir);
1400
1401         if (!ldlm_state->ldlm_cb_service) {
1402                 CERROR("failed to start service\n");
1403                 GOTO(out_proc, rc = -ENOMEM);
1404         }
1405
1406         ldlm_state->ldlm_cancel_service =
1407                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1408                                 LDLM_CANCEL_REQUEST_PORTAL,
1409                                 LDLM_CANCEL_REPLY_PORTAL, 30000,
1410                                 ldlm_cancel_handler, "ldlm_canceld",
1411                                 ldlm_svc_proc_dir);
1412
1413         if (!ldlm_state->ldlm_cancel_service) {
1414                 CERROR("failed to start service\n");
1415                 GOTO(out_proc, rc = -ENOMEM);
1416         }
1417
1418         OBD_ALLOC(blp, sizeof(*blp));
1419         if (blp == NULL)
1420                 GOTO(out_proc, rc = -ENOMEM);
1421         ldlm_state->ldlm_bl_pool = blp;
1422
1423         atomic_set(&blp->blp_num_threads, 0);
1424         init_waitqueue_head(&blp->blp_waitq);
1425         spin_lock_init(&blp->blp_lock);
1426
1427         INIT_LIST_HEAD(&blp->blp_list);
1428
1429 #ifdef __KERNEL__
1430         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1431                 struct ldlm_bl_thread_data bltd = {
1432                         .bltd_num = i,
1433                         .bltd_blp = blp,
1434                 };
1435                 init_completion(&blp->blp_comp);
1436                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1437                 if (rc < 0) {
1438                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1439                         GOTO(out_thread, rc);
1440                 }
1441                 wait_for_completion(&blp->blp_comp);
1442         }
1443
1444         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1445                                     LDLM_NUM_THREADS, "ldlm_cn");
1446         if (rc)
1447                 GOTO(out_thread, rc);
1448
1449         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1450                                     LDLM_NUM_THREADS, "ldlm_cb");
1451         if (rc)
1452                 GOTO(out_thread, rc);
1453
1454         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1455         spin_lock_init(&expired_lock_thread.elt_lock);
1456         expired_lock_thread.elt_state = ELT_STOPPED;
1457         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1458
1459         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1460         if (rc < 0) {
1461                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1462                 GOTO(out_thread, rc);
1463         }
1464
1465         wait_event(expired_lock_thread.elt_waitq,
1466                    expired_lock_thread.elt_state == ELT_READY);
1467
1468         INIT_LIST_HEAD(&waiting_locks_list);
1469         spin_lock_init(&waiting_locks_spinlock);
1470         waiting_locks_timer.function = waiting_locks_callback;
1471         waiting_locks_timer.data = 0;
1472         init_timer(&waiting_locks_timer);
1473 #endif
1474
1475         RETURN(0);
1476
1477 #ifdef __KERNEL__
1478  out_thread:
1479         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1480         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1481 #endif
1482
1483  out_proc:
1484 #ifdef __KERNEL__
1485         ldlm_proc_cleanup();
1486  out_free:
1487 #endif
1488         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1489         ldlm_state = NULL;
1490         return rc;
1491 }
1492
1493 static int ldlm_cleanup(int force)
1494 {
1495 #ifdef __KERNEL__
1496         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1497 #endif
1498         ENTRY;
1499
1500         if (!list_empty(&ldlm_namespace_list)) {
1501                 CERROR("ldlm still has namespaces; clean these up first.\n");
1502                 ldlm_dump_all_namespaces(D_DLMTRACE);
1503                 RETURN(-EBUSY);
1504         }
1505
1506 #ifdef __KERNEL__
1507         while (atomic_read(&blp->blp_num_threads) > 0) {
1508                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1509
1510                 init_completion(&blp->blp_comp);
1511
1512                 spin_lock(&blp->blp_lock);
1513                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1514                 wake_up(&blp->blp_waitq);
1515                 spin_unlock(&blp->blp_lock);
1516
1517                 wait_for_completion(&blp->blp_comp);
1518         }
1519         OBD_FREE(blp, sizeof(*blp));
1520
1521         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1522         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1523         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1524         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1525         ldlm_proc_cleanup();
1526
1527         expired_lock_thread.elt_state = ELT_TERMINATE;
1528         wake_up(&expired_lock_thread.elt_waitq);
1529         wait_event(expired_lock_thread.elt_waitq,
1530                    expired_lock_thread.elt_state == ELT_STOPPED);
1531 #else
1532         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1533         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1534 #endif
1535
1536         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1537         ldlm_state = NULL;
1538
1539         RETURN(0);
1540 }
1541
1542 int __init ldlm_init(void)
1543 {
1544         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1545                                                sizeof(struct ldlm_resource), 0,
1546                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1547         if (ldlm_resource_slab == NULL)
1548                 return -ENOMEM;
1549
1550         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1551                                            sizeof(struct ldlm_lock), 0,
1552                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1553         if (ldlm_lock_slab == NULL) {
1554                 kmem_cache_destroy(ldlm_resource_slab);
1555                 return -ENOMEM;
1556         }
1557
1558         return 0;
1559 }
1560
1561 void __exit ldlm_exit(void)
1562 {
1563         if ( ldlm_refcount )
1564                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1565         LASSERTF(kmem_cache_destroy(ldlm_resource_slab) == 0,
1566                  "couldn't free ldlm resource slab\n");
1567         LASSERTF(kmem_cache_destroy(ldlm_lock_slab) == 0,
1568                  "couldn't free ldlm lock slab\n");
1569
1570 }
1571
1572 /* ldlm_flock.c */
1573 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1574
1575 /* ldlm_extent.c */
1576 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1577
1578 /* ldlm_lock.c */
1579 EXPORT_SYMBOL(ldlm_get_processing_policy);
1580 EXPORT_SYMBOL(ldlm_lock2desc);
1581 EXPORT_SYMBOL(ldlm_register_intent);
1582 EXPORT_SYMBOL(ldlm_lockname);
1583 EXPORT_SYMBOL(ldlm_typename);
1584 EXPORT_SYMBOL(ldlm_lock2handle);
1585 EXPORT_SYMBOL(__ldlm_handle2lock);
1586 EXPORT_SYMBOL(ldlm_lock_get);
1587 EXPORT_SYMBOL(ldlm_lock_put);
1588 EXPORT_SYMBOL(ldlm_lock_match);
1589 EXPORT_SYMBOL(ldlm_lock_cancel);
1590 EXPORT_SYMBOL(ldlm_lock_addref);
1591 EXPORT_SYMBOL(ldlm_lock_decref);
1592 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1593 EXPORT_SYMBOL(ldlm_lock_change_resource);
1594 EXPORT_SYMBOL(ldlm_lock_set_data);
1595 EXPORT_SYMBOL(ldlm_it2str);
1596 EXPORT_SYMBOL(ldlm_lock_dump);
1597 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1598 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1599 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1600 EXPORT_SYMBOL(ldlm_lock_allow_match);
1601
1602 /* ldlm_request.c */
1603 EXPORT_SYMBOL(ldlm_completion_ast);
1604 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1605 EXPORT_SYMBOL(ldlm_cli_convert);
1606 EXPORT_SYMBOL(ldlm_cli_enqueue);
1607 EXPORT_SYMBOL(ldlm_cli_cancel);
1608 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1609 EXPORT_SYMBOL(ldlm_replay_locks);
1610 EXPORT_SYMBOL(ldlm_resource_foreach);
1611 EXPORT_SYMBOL(ldlm_namespace_foreach);
1612 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1613 EXPORT_SYMBOL(ldlm_change_cbdata);
1614
1615 /* ldlm_lockd.c */
1616 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1617 EXPORT_SYMBOL(ldlm_server_completion_ast);
1618 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1619 EXPORT_SYMBOL(ldlm_handle_enqueue);
1620 EXPORT_SYMBOL(ldlm_handle_cancel);
1621 EXPORT_SYMBOL(ldlm_handle_convert);
1622 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1623 EXPORT_SYMBOL(ldlm_get_ref);
1624 EXPORT_SYMBOL(ldlm_put_ref);
1625
1626 #if 0
1627 /* ldlm_test.c */
1628 EXPORT_SYMBOL(ldlm_test);
1629 EXPORT_SYMBOL(ldlm_regression_start);
1630 EXPORT_SYMBOL(ldlm_regression_stop);
1631 #endif
1632
1633 /* ldlm_resource.c */
1634 EXPORT_SYMBOL(ldlm_namespace_new);
1635 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1636 EXPORT_SYMBOL(ldlm_namespace_free);
1637 EXPORT_SYMBOL(ldlm_namespace_dump);
1638 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1639 EXPORT_SYMBOL(ldlm_resource_get);
1640 EXPORT_SYMBOL(ldlm_resource_putref);
1641
1642 /* ldlm_lib.c */
1643 EXPORT_SYMBOL(client_import_add_conn);
1644 EXPORT_SYMBOL(client_import_del_conn);
1645 EXPORT_SYMBOL(client_obd_setup);
1646 EXPORT_SYMBOL(client_obd_cleanup);
1647 EXPORT_SYMBOL(client_connect_import);
1648 EXPORT_SYMBOL(client_disconnect_export);
1649 EXPORT_SYMBOL(target_start_recovery_thread);
1650 EXPORT_SYMBOL(target_stop_recovery_thread);
1651 EXPORT_SYMBOL(target_handle_connect);
1652 EXPORT_SYMBOL(target_cleanup_recovery);
1653 EXPORT_SYMBOL(target_destroy_export);
1654 EXPORT_SYMBOL(target_cancel_recovery_timer);
1655 EXPORT_SYMBOL(target_send_reply);
1656 EXPORT_SYMBOL(target_queue_recovery_request);
1657 EXPORT_SYMBOL(target_handle_ping);
1658 EXPORT_SYMBOL(target_handle_disconnect);
1659
1660 /* l_lock.c */
1661 EXPORT_SYMBOL(lock_res_and_lock);
1662 EXPORT_SYMBOL(unlock_res_and_lock);
1663
1664