lustre/ldlm/ldlm_lockd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #ifndef EXPORT_SYMTAB
  25 # define EXPORT_SYMTAB
  26 #endif
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28
  29 #ifdef __KERNEL__
  30 # include <linux/module.h>
  31 # include <linux/slab.h>
  32 # include <linux/init.h>
  33 # include <linux/wait.h>
  34 #else
  35 # include <liblustre.h>
  36 #endif
  37
  38 #include <linux/lustre_dlm.h>
  39 #include <linux/obd_class.h>
  40 #include <portals/list.h>
  41 #include "ldlm_internal.h"
  42
  43 extern kmem_cache_t *ldlm_resource_slab;
  44 extern kmem_cache_t *ldlm_lock_slab;
  45 extern struct lustre_lock ldlm_handle_lock;
  46 extern struct list_head ldlm_namespace_list;
  47
  48 static DECLARE_MUTEX(ldlm_ref_sem);
  49 static int ldlm_refcount;
  50
  51 /* LDLM state */
  52
  53 static struct ldlm_state *ldlm_state;
  54
  55 inline unsigned long round_timeout(unsigned long timeout)
  56 {
  57         return ((timeout / HZ) + 1) * HZ;
  58 }
  59
  60 #ifdef __KERNEL__
  61 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
  62 static spinlock_t waiting_locks_spinlock;
  63 static struct list_head waiting_locks_list;
  64 static struct timer_list waiting_locks_timer;
  65
  66 static struct expired_lock_thread {
  67         wait_queue_head_t         elt_waitq;
  68         int                       elt_state;
  69         struct list_head          elt_expired_locks;
  70 } expired_lock_thread;
  71 #endif
  72
  73 #define ELT_STOPPED   0
  74 #define ELT_READY     1
  75 #define ELT_TERMINATE 2
  76
  77 struct ldlm_bl_pool {
  78         spinlock_t              blp_lock;
  79         struct list_head        blp_list;
  80         wait_queue_head_t       blp_waitq;
  81         atomic_t                blp_num_threads;
  82         struct completion       blp_comp;
  83 };
  84
  85 struct ldlm_bl_work_item {
  86         struct list_head        blwi_entry;
  87         struct ldlm_namespace   *blwi_ns;
  88         struct ldlm_lock_desc   blwi_ld;
  89         struct ldlm_lock        *blwi_lock;
  90 };
  91
  92 #ifdef __KERNEL__
  93
  94 static inline int have_expired_locks(void)
  95 {
  96         int need_to_run;
  97
  98         spin_lock_bh(&waiting_locks_spinlock);
  99         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 100         spin_unlock_bh(&waiting_locks_spinlock);
 101
 102         RETURN(need_to_run);
 103 }
 104
 105 static int expired_lock_main(void *arg)
 106 {
 107         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 108         struct l_wait_info lwi = { 0 };
 109         unsigned long flags;
 110
 111         ENTRY;
 112         lock_kernel();
 113         kportal_daemonize("ldlm_elt");
 114
 115         SIGNAL_MASK_LOCK(current, flags);
 116         sigfillset(&current->blocked);
 117         RECALC_SIGPENDING;
 118         SIGNAL_MASK_UNLOCK(current, flags);
 119
 120         unlock_kernel();
 121
 122         expired_lock_thread.elt_state = ELT_READY;
 123         wake_up(&expired_lock_thread.elt_waitq);
 124
 125         while (1) {
 126                 l_wait_event(expired_lock_thread.elt_waitq,
 127                              have_expired_locks() ||
 128                              expired_lock_thread.elt_state == ELT_TERMINATE,
 129                              &lwi);
 130
 131                 spin_lock_bh(&waiting_locks_spinlock);
 132                 while (!list_empty(expired)) {
 133                         struct obd_export *export;
 134                         struct ldlm_lock *lock;
 135
 136                         lock = list_entry(expired->next, struct ldlm_lock,
 137                                           l_pending_chain);
 138                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
 139                             (void *)lock >= LP_POISON) {
 140                                 CERROR("free lock on elt list %p\n", lock);
 141                                 LBUG();
 142                         }
 143                         list_del_init(&lock->l_pending_chain);
 144                         if ((void *)lock->l_export < LP_POISON + PAGE_SIZE &&
 145                             (void *)lock->l_export >= LP_POISON + PAGE_SIZE) {
 146                                 CERROR("lock with free export on elt list %p\n",
 147                                        export);
 148                                 lock->l_export = NULL;
 149                                 LDLM_ERROR(lock, "free export\n");
 150                                 continue;
 151                         }
 152                         export = class_export_get(lock->l_export);
 153                         spin_unlock_bh(&waiting_locks_spinlock);
 154
 155                         ptlrpc_fail_export(export);
 156                         class_export_put(export);
 157                         spin_lock_bh(&waiting_locks_spinlock);
 158                 }
 159                 spin_unlock_bh(&waiting_locks_spinlock);
 160
 161                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 162                         break;
 163         }
 164
 165         expired_lock_thread.elt_state = ELT_STOPPED;
 166         wake_up(&expired_lock_thread.elt_waitq);
 167         RETURN(0);
 168 }
 169
 170 static void waiting_locks_callback(unsigned long unused)
 171 {
 172         struct ldlm_lock *lock, *last = NULL;
 173         char str[PTL_NALFMT_SIZE];
 174
 175         spin_lock_bh(&waiting_locks_spinlock);
 176         while (!list_empty(&waiting_locks_list)) {
 177
 178                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 179                                   l_pending_chain);
 180
 181                 if (lock->l_callback_timeout > jiffies)
 182                         break;
 183
 184                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 185                            "%s@%s nid %s ",lock->l_export->exp_client_uuid.uuid,
 186                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 187                            ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer,str));
 188
 189                 if (lock == last) {
 190                         LDLM_ERROR(lock, "waiting on lock multiple times");
 191                         CERROR("wll %p .prev %p, l_pending.next %p .prev %p\n",
 192                                waiting_locks_list.next, waiting_locks_list.prev,
 193                                lock->l_pending_chain.next,
 194                                lock->l_pending_chain.prev);
 195                         spin_unlock_bh(&waiting_locks_spinlock);
 196                         LBUG();
 197                 }
 198                 last = lock;
 199
 200                 list_del(&lock->l_pending_chain);
 201                 list_add(&lock->l_pending_chain,
 202                          &expired_lock_thread.elt_expired_locks);
 203
 204                 wake_up(&expired_lock_thread.elt_waitq);
 205         }
 206
 207         /*
 208          * Make sure the timer will fire again if we have any locks
 209          * left.
 210          */
 211         if (!list_empty(&waiting_locks_list)) {
 212                 unsigned long timeout_rounded;
 213                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 214                                   l_pending_chain);
 215                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 216                 mod_timer(&waiting_locks_timer, timeout_rounded);
 217         }
 218         spin_unlock_bh(&waiting_locks_spinlock);
 219 }
 220
 221 /*
 222  * Indicate that we're waiting for a client to call us back cancelling a given
 223  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 224  * timer to fire appropriately.  (We round up to the next second, to avoid
 225  * floods of timer firings during periods of high lock contention and traffic).
 226  */
 227 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 228 {
 229         unsigned long timeout_rounded;
 230
 231         spin_lock_bh(&waiting_locks_spinlock);
 232         if (!list_empty(&lock->l_pending_chain)) {
 233                 spin_unlock_bh(&waiting_locks_spinlock);
 234                 LDLM_DEBUG(lock, "not re-adding to wait list");
 235                 return 0;
 236         }
 237
 238         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 239
 240         timeout_rounded = round_timeout(lock->l_callback_timeout);
 241
 242         if (timeout_rounded < waiting_locks_timer.expires ||
 243             !timer_pending(&waiting_locks_timer)) {
 244                 mod_timer(&waiting_locks_timer, timeout_rounded);
 245         }
 246         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 247         spin_unlock_bh(&waiting_locks_spinlock);
 248         LDLM_DEBUG(lock, "adding to wait list");
 249         return 1;
 250 }
 251
 252 /*
 253  * Remove a lock from the pending list, likely because it had its cancellation
 254  * callback arrive without incident.  This adjusts the lock-timeout timer if
 255  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 256  */
 257 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 258 {
 259         struct list_head *list_next;
 260
 261         if (lock->l_export == NULL) {
 262                 /* We don't have a "waiting locks list" on clients. */
 263                 LDLM_DEBUG(lock, "client lock: no-op");
 264                 return 0;
 265         }
 266
 267         spin_lock_bh(&waiting_locks_spinlock);
 268
 269         if (list_empty(&lock->l_pending_chain)) {
 270                 spin_unlock_bh(&waiting_locks_spinlock);
 271                 LDLM_DEBUG(lock, "wasn't waiting");
 272                 return 0;
 273         }
 274
 275         list_next = lock->l_pending_chain.next;
 276         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 277                 /* Removing the head of the list, adjust timer. */
 278                 if (list_next == &waiting_locks_list) {
 279                         /* No more, just cancel. */
 280                         del_timer(&waiting_locks_timer);
 281                 } else {
 282                         struct ldlm_lock *next;
 283                         next = list_entry(list_next, struct ldlm_lock,
 284                                           l_pending_chain);
 285                         mod_timer(&waiting_locks_timer,
 286                                   round_timeout(next->l_callback_timeout));
 287                 }
 288         }
 289         list_del_init(&lock->l_pending_chain);
 290
 291         spin_unlock_bh(&waiting_locks_spinlock);
 292         LDLM_DEBUG(lock, "removed");
 293         return 1;
 294 }
 295
 296 #else /* !__KERNEL__ */
 297
 298 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 299 {
 300         RETURN(1);
 301 }
 302
 303 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 304 {
 305         RETURN(0);
 306 }
 307
 308 #endif /* __KERNEL__ */
 309
 310 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,const char *ast_type)
 311 {
 312         struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 313         char str[PTL_NALFMT_SIZE];
 314
 315         LDLM_ERROR(lock, "%s AST failed (%d): evicting client %s@%s NID "LPX64
 316                    " (%s)", ast_type, rc, lock->l_export->exp_client_uuid.uuid,
 317                    conn->c_remote_uuid.uuid, conn->c_peer.peer_id.nid,
 318                    ptlrpc_peernid2str(&conn->c_peer, str));
 319
 320         ptlrpc_fail_export(lock->l_export);
 321 }
 322
 323 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 324                                  struct ptlrpc_request *req, int rc,
 325                                  const char *ast_type)
 326 {
 327         struct ptlrpc_peer *peer = &req->rq_import->imp_connection->c_peer;
 328         char str[PTL_NALFMT_SIZE];
 329
 330         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 331                 LASSERT(lock->l_export);
 332                 if (lock->l_export->exp_libclient) {
 333                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 334                                    " timeout, just cancelling lock", ast_type,
 335                                    ptlrpc_peernid2str(peer, str));
 336                         ldlm_lock_cancel(lock);
 337                         rc = -ERESTART;
 338                 } else {
 339                         ldlm_del_waiting_lock(lock);
 340                         ldlm_failed_ast(lock, rc, ast_type);
 341                 }
 342         } else if (rc) {
 343                 if (rc == -EINVAL)
 344                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 345                                    " from %s AST - normal race",
 346                                    ptlrpc_peernid2str(peer, str),
 347                                    req->rq_repmsg->status, ast_type);
 348                 else
 349                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 350                                    "from %s AST", ptlrpc_peernid2str(peer, str),
 351                                    (req->rq_repmsg != NULL) ?
 352                                    req->rq_repmsg->status : 0, ast_type);
 353                 ldlm_lock_cancel(lock);
 354                 /* Server-side AST functions are called from ldlm_reprocess_all,
 355                  * which needs to be told to please restart its reprocessing. */
 356                 rc = -ERESTART;
 357         }
 358
 359         return rc;
 360 }
 361
 362 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 363                              struct ldlm_lock_desc *desc,
 364                              void *data, int flag)
 365 {
 366         struct ldlm_request *body;
 367         struct ptlrpc_request *req;
 368         int rc = 0, size = sizeof(*body);
 369         ENTRY;
 370
 371         if (flag == LDLM_CB_CANCELING) {
 372                 /* Don't need to do anything here. */
 373                 RETURN(0);
 374         }
 375
 376         LASSERT(lock);
 377
 378         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 379         if (lock->l_granted_mode != lock->l_req_mode) {
 380                 /* this blocking AST will be communicated as part of the
 381                  * completion AST instead */
 382                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 383                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 384                 RETURN(0);
 385         }
 386
 387         if (lock->l_destroyed) {
 388                 /* What's the point? */
 389                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 390                 RETURN(0);
 391         }
 392
 393 #if 0
 394         if (CURRENT_SECONDS - lock->l_export->exp_last_request_time > 30){
 395                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 396                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 397                 RETURN(-ETIMEDOUT);
 398         }
 399 #endif
 400
 401         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 402                               LDLM_BL_CALLBACK, 1, &size, NULL);
 403         if (req == NULL) {
 404                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 405                 RETURN(-ENOMEM);
 406         }
 407
 408         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 409         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 410                sizeof(body->lock_handle1));
 411         memcpy(&body->lock_desc, desc, sizeof(*desc));
 412         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 413
 414         LDLM_DEBUG(lock, "server preparing blocking AST");
 415         req->rq_replen = lustre_msg_size(0, NULL);
 416
 417         if (lock->l_granted_mode == lock->l_req_mode)
 418                 ldlm_add_waiting_lock(lock);
 419         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 420
 421         req->rq_send_state = LUSTRE_IMP_FULL;
 422         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 423         rc = ptlrpc_queue_wait(req);
 424         if (rc != 0)
 425                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 426
 427         ptlrpc_req_finished(req);
 428
 429         RETURN(rc);
 430 }
 431
 432 /* XXX copied from ptlrpc/service.c */
 433 static long timeval_sub(struct timeval *large, struct timeval *small)
 434 {
 435         return (large->tv_sec - small->tv_sec) * 1000000 +
 436                 (large->tv_usec - small->tv_usec);
 437 }
 438
 439 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 440 {
 441         struct ldlm_request *body;
 442         struct ptlrpc_request *req;
 443         struct timeval granted_time;
 444         long total_enqueue_wait;
 445         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 446         ENTRY;
 447
 448         LASSERT(lock != NULL);
 449
 450         do_gettimeofday(&granted_time);
 451         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 452
 453         if (total_enqueue_wait / 1000000 > obd_timeout)
 454                 LDLM_ERROR(lock, "enqueue wait took %luus", total_enqueue_wait);
 455
 456         down(&lock->l_resource->lr_lvb_sem);
 457         if (lock->l_resource->lr_lvb_len) {
 458                 buffers = 2;
 459                 size[1] = lock->l_resource->lr_lvb_len;
 460         }
 461         up(&lock->l_resource->lr_lvb_sem);
 462
 463         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 464                               LDLM_CP_CALLBACK, buffers, size, NULL);
 465         if (req == NULL)
 466                 RETURN(-ENOMEM);
 467
 468         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 469         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 470                sizeof(body->lock_handle1));
 471         body->lock_flags = flags;
 472         ldlm_lock2desc(lock, &body->lock_desc);
 473
 474         if (buffers == 2) {
 475                 void *lvb;
 476
 477                 down(&lock->l_resource->lr_lvb_sem);
 478                 lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 479                                      lock->l_resource->lr_lvb_len);
 480                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 481                        lock->l_resource->lr_lvb_len);
 482                 up(&lock->l_resource->lr_lvb_sem);
 483         }
 484
 485         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 486                    total_enqueue_wait);
 487         req->rq_replen = lustre_msg_size(0, NULL);
 488
 489         req->rq_send_state = LUSTRE_IMP_FULL;
 490         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 491
 492         /* We only send real blocking ASTs after the lock is granted */
 493         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 494         if (lock->l_flags & LDLM_FL_AST_SENT) {
 495                 body->lock_flags |= LDLM_FL_AST_SENT;
 496                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 497         }
 498         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 499
 500         rc = ptlrpc_queue_wait(req);
 501         if (rc != 0)
 502                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 503
 504         ptlrpc_req_finished(req);
 505
 506         RETURN(rc);
 507 }
 508
 509 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 510 {
 511         struct ldlm_resource *res = lock->l_resource;
 512         struct ldlm_request *body;
 513         struct ptlrpc_request *req;
 514         int rc = 0, size = sizeof(*body);
 515         ENTRY;
 516
 517         LASSERT(lock != NULL);
 518
 519         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 520                               LDLM_GL_CALLBACK, 1, &size, NULL);
 521         if (req == NULL)
 522                 RETURN(-ENOMEM);
 523
 524         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 525         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 526                sizeof(body->lock_handle1));
 527         ldlm_lock2desc(lock, &body->lock_desc);
 528
 529         down(&lock->l_resource->lr_lvb_sem);
 530         size = lock->l_resource->lr_lvb_len;
 531         up(&lock->l_resource->lr_lvb_sem);
 532         req->rq_replen = lustre_msg_size(1, &size);
 533
 534         req->rq_send_state = LUSTRE_IMP_FULL;
 535         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 536
 537         rc = ptlrpc_queue_wait(req);
 538         if (rc == -ELDLM_NO_LOCK_DATA)
 539                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 540         else if (rc != 0)
 541                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 542         else
 543                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 544                         (res, req->rq_repmsg, 0, 1);
 545         ptlrpc_req_finished(req);
 546         RETURN(rc);
 547 }
 548
 549 static struct ldlm_lock *
 550 find_existing_lock(struct obd_export *exp, struct lustre_handle *remote_hdl)
 551 {
 552         struct obd_device *obd = exp->exp_obd;
 553         struct list_head *iter;
 554
 555         l_lock(&obd->obd_namespace->ns_lock);
 556         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
 557                 struct ldlm_lock *lock;
 558                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 559                 if (lock->l_remote_handle.cookie == remote_hdl->cookie) {
 560                         LDLM_LOCK_GET(lock);
 561                         l_unlock(&obd->obd_namespace->ns_lock);
 562                         return lock;
 563                 }
 564         }
 565         l_unlock(&obd->obd_namespace->ns_lock);
 566         return NULL;
 567 }
 568
 569
 570 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 571                         ldlm_completion_callback completion_callback,
 572                         ldlm_blocking_callback blocking_callback,
 573                         ldlm_glimpse_callback glimpse_callback)
 574 {
 575         struct obd_device *obddev = req->rq_export->exp_obd;
 576         struct ldlm_reply *dlm_rep;
 577         struct ldlm_request *dlm_req;
 578         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 579         __u32 flags;
 580         ldlm_error_t err = ELDLM_OK;
 581         struct ldlm_lock *lock = NULL;
 582         void *cookie = NULL;
 583         ENTRY;
 584
 585         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 586
 587         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 588                                       lustre_swab_ldlm_request);
 589         if (dlm_req == NULL) {
 590                 CERROR ("Can't unpack dlm_req\n");
 591                 GOTO(out, rc = -EFAULT);
 592         }
 593
 594         flags = dlm_req->lock_flags;
 595
 596         LASSERT(req->rq_export);
 597
 598         if (flags & LDLM_FL_REPLAY) {
 599                 lock = find_existing_lock(req->rq_export,
 600                                           &dlm_req->lock_handle1);
 601                 if (lock != NULL) {
 602                         DEBUG_REQ(D_HA, req, "found existing lock cookie "LPX64,
 603                                   lock->l_handle.h_cookie);
 604                         GOTO(existing_lock, rc = 0);
 605                 }
 606         }
 607
 608         /* The lock's callback data might be set in the policy function */
 609         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 610                                 dlm_req->lock_desc.l_resource.lr_name,
 611                                 dlm_req->lock_desc.l_resource.lr_type,
 612                                 dlm_req->lock_desc.l_req_mode,
 613                                 blocking_callback, completion_callback,
 614                                 glimpse_callback, NULL, 0);
 615         if (!lock)
 616                 GOTO(out, rc = -ENOMEM);
 617
 618         do_gettimeofday(&lock->l_enqueued_time);
 619         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 620                sizeof(lock->l_remote_handle));
 621         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 622
 623         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
 624         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 625         if (req->rq_export->exp_failed) {
 626                 LDLM_ERROR(lock,"lock on destroyed export %p\n",req->rq_export);
 627                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 628                 GOTO(out, rc = -ENOTCONN);
 629         }
 630         lock->l_export = class_export_get(req->rq_export);
 631         list_add(&lock->l_export_chain,
 632                  &lock->l_export->exp_ldlm_data.led_held_locks);
 633         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 634
 635 existing_lock:
 636
 637         if (flags & LDLM_FL_HAS_INTENT) {
 638                 /* In this case, the reply buffer is allocated deep in
 639                  * local_lock_enqueue by the policy function. */
 640                 cookie = req;
 641         } else {
 642                 int buffers = 1;
 643
 644                 down(&lock->l_resource->lr_lvb_sem);
 645                 if (lock->l_resource->lr_lvb_len) {
 646                         size[1] = lock->l_resource->lr_lvb_len;
 647                         buffers = 2;
 648                 }
 649                 up(&lock->l_resource->lr_lvb_sem);
 650
 651                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 652                         GOTO(out, rc = -ENOMEM);
 653
 654                 rc = lustre_pack_reply(req, buffers, size, NULL);
 655                 if (rc)
 656                         GOTO(out, rc);
 657         }
 658
 659         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 660                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 661                        sizeof(ldlm_policy_data_t));
 662         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 663                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 664                        sizeof(lock->l_req_extent));
 665
 666         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 667         if (err)
 668                 GOTO(out, err);
 669
 670         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 671         dlm_rep->lock_flags = flags;
 672
 673         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 674         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 675
 676         /* We never send a blocking AST until the lock is granted, but
 677          * we can tell it right now */
 678         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 679         if (lock->l_flags & LDLM_FL_AST_SENT) {
 680                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 681                 if (lock->l_granted_mode == lock->l_req_mode)
 682                         ldlm_add_waiting_lock(lock);
 683         }
 684         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 685
 686         EXIT;
 687  out:
 688         req->rq_status = err;
 689         if (req->rq_reply_state == NULL) {
 690                 err = lustre_pack_reply(req, 0, NULL, NULL);
 691                 if (rc == 0)
 692                         rc = err;
 693                 req->rq_status = rc;
 694         }
 695
 696         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 697          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 698         if (lock) {
 699                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 700                            "(err=%d, rc=%d)", err, rc);
 701
 702                 if (rc == 0) {
 703                         down(&lock->l_resource->lr_lvb_sem);
 704                         size[1] = lock->l_resource->lr_lvb_len;
 705                         if (size[1] > 0) {
 706                                 void *lvb = lustre_msg_buf(req->rq_repmsg,
 707                                                            1, size[1]);
 708                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
 709                                          req, lock);
 710
 711                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 712                                        size[1]);
 713                         }
 714                         up(&lock->l_resource->lr_lvb_sem);
 715                 } else {
 716                         ldlm_lock_destroy(lock);
 717                 }
 718
 719                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 720                         ldlm_reprocess_all(lock->l_resource);
 721                 LDLM_LOCK_PUT(lock);
 722         }
 723         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 724                           lock, rc);
 725
 726         return rc;
 727 }
 728
 729 int ldlm_handle_convert(struct ptlrpc_request *req)
 730 {
 731         struct ldlm_request *dlm_req;
 732         struct ldlm_reply *dlm_rep;
 733         struct ldlm_lock *lock;
 734         int rc, size = sizeof(*dlm_rep);
 735         ENTRY;
 736
 737         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 738                                       lustre_swab_ldlm_request);
 739         if (dlm_req == NULL) {
 740                 CERROR ("Can't unpack dlm_req\n");
 741                 RETURN (-EFAULT);
 742         }
 743
 744         rc = lustre_pack_reply(req, 1, &size, NULL);
 745         if (rc) {
 746                 CERROR("out of memory\n");
 747                 RETURN(-ENOMEM);
 748         }
 749         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 750         dlm_rep->lock_flags = dlm_req->lock_flags;
 751
 752         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 753         if (!lock) {
 754                 req->rq_status = EINVAL;
 755         } else {
 756                 LDLM_DEBUG(lock, "server-side convert handler START");
 757                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 758                                   &dlm_rep->lock_flags);
 759                 if (ldlm_del_waiting_lock(lock))
 760                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 761                 req->rq_status = 0;
 762         }
 763
 764         if (lock) {
 765                 ldlm_reprocess_all(lock->l_resource);
 766                 LDLM_DEBUG(lock, "server-side convert handler END");
 767                 LDLM_LOCK_PUT(lock);
 768         } else
 769                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 770
 771         RETURN(0);
 772 }
 773
 774 int ldlm_handle_cancel(struct ptlrpc_request *req)
 775 {
 776         struct ldlm_request *dlm_req;
 777         struct ldlm_lock *lock;
 778         struct ldlm_resource *res;
 779         int rc;
 780         ENTRY;
 781
 782         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 783                                       lustre_swab_ldlm_request);
 784         if (dlm_req == NULL) {
 785                 CERROR("bad request buffer for cancel\n");
 786                 RETURN(-EFAULT);
 787         }
 788
 789         rc = lustre_pack_reply(req, 0, NULL, NULL);
 790         if (rc) {
 791                 CERROR("out of memory\n");
 792                 RETURN(-ENOMEM);
 793         }
 794
 795         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 796         if (!lock) {
 797                 CERROR("received cancel for unknown lock cookie "LPX64
 798                        " from client %s id %s\n",
 799                        dlm_req->lock_handle1.cookie,
 800                        req->rq_export->exp_client_uuid.uuid,
 801                        req->rq_peerstr);
 802                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 803                                   "(cookie "LPU64")",
 804                                   dlm_req->lock_handle1.cookie);
 805                 req->rq_status = ESTALE;
 806         } else {
 807                 LDLM_DEBUG(lock, "server-side cancel handler START");
 808                 res = lock->l_resource;
 809                 if (res && res->lr_namespace->ns_lvbo &&
 810                     res->lr_namespace->ns_lvbo->lvbo_update) {
 811                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 812                                 (res, NULL, 0, 0);
 813                                 //(res, req->rq_reqmsg, 1);
 814                 }
 815
 816                 ldlm_lock_cancel(lock);
 817                 if (ldlm_del_waiting_lock(lock))
 818                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 819                 req->rq_status = rc;
 820         }
 821
 822         if (ptlrpc_reply(req) != 0)
 823                 LBUG();
 824
 825         if (lock) {
 826                 ldlm_reprocess_all(lock->l_resource);
 827                 LDLM_DEBUG(lock, "server-side cancel handler END");
 828                 LDLM_LOCK_PUT(lock);
 829         }
 830
 831         RETURN(0);
 832 }
 833
 834 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 835                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 836 {
 837         int do_ast;
 838         ENTRY;
 839
 840         l_lock(&ns->ns_lock);
 841         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 842
 843         lock->l_flags |= LDLM_FL_CBPENDING;
 844         do_ast = (!lock->l_readers && !lock->l_writers);
 845
 846         if (do_ast) {
 847                 LDLM_DEBUG(lock, "already unused, calling "
 848                            "callback (%p)", lock->l_blocking_ast);
 849                 if (lock->l_blocking_ast != NULL) {
 850                         l_unlock(&ns->ns_lock);
 851                         l_check_no_ns_lock(ns);
 852                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 853                                              LDLM_CB_BLOCKING);
 854                         l_lock(&ns->ns_lock);
 855                 }
 856         } else {
 857                 LDLM_DEBUG(lock, "Lock still has references, will be"
 858                            " cancelled later");
 859         }
 860
 861         LDLM_DEBUG(lock, "client blocking callback handler END");
 862         l_unlock(&ns->ns_lock);
 863         LDLM_LOCK_PUT(lock);
 864         EXIT;
 865 }
 866
 867 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 868                                     struct ldlm_namespace *ns,
 869                                     struct ldlm_request *dlm_req,
 870                                     struct ldlm_lock *lock)
 871 {
 872         LIST_HEAD(ast_list);
 873         ENTRY;
 874
 875         l_lock(&ns->ns_lock);
 876         LDLM_DEBUG(lock, "client completion callback handler START");
 877
 878         /* If we receive the completion AST before the actual enqueue returned,
 879          * then we might need to switch lock modes, resources, or extents. */
 880         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 881                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 882                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 883         }
 884
 885         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 886                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 887                        sizeof(lock->l_policy_data));
 888                 LDLM_DEBUG(lock, "completion AST, new policy data");
 889         }
 890
 891         ldlm_resource_unlink_lock(lock);
 892         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 893                    &lock->l_resource->lr_name,
 894                    sizeof(lock->l_resource->lr_name)) != 0) {
 895                 ldlm_lock_change_resource(ns, lock,
 896                                          dlm_req->lock_desc.l_resource.lr_name);
 897                 LDLM_DEBUG(lock, "completion AST, new resource");
 898         }
 899
 900         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 901                 lock->l_flags |= LDLM_FL_CBPENDING;
 902                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 903         }
 904
 905         if (lock->l_lvb_len) {
 906                 void *lvb;
 907                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 908                                          lock->l_lvb_swabber);
 909                 if (lvb == NULL) {
 910                         LDLM_ERROR(lock, "completion AST did not contain "
 911                                    "expected LVB!");
 912                 } else {
 913                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 914                 }
 915         }
 916
 917         lock->l_resource->lr_tmp = &ast_list;
 918         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 919         lock->l_resource->lr_tmp = NULL;
 920         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 921         l_unlock(&ns->ns_lock);
 922         LDLM_LOCK_PUT(lock);
 923
 924         ldlm_run_ast_work(ns, &ast_list);
 925
 926         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 927                           lock);
 928         EXIT;
 929 }
 930
 931 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 932                                     struct ldlm_namespace *ns,
 933                                     struct ldlm_request *dlm_req,
 934                                     struct ldlm_lock *lock)
 935 {
 936         int rc = -ENOSYS;
 937         ENTRY;
 938
 939         l_lock(&ns->ns_lock);
 940         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 941
 942         if (lock->l_glimpse_ast != NULL) {
 943                 l_unlock(&ns->ns_lock);
 944                 l_check_no_ns_lock(ns);
 945                 rc = lock->l_glimpse_ast(lock, req);
 946                 l_lock(&ns->ns_lock);
 947         }
 948
 949         if (req->rq_repmsg != NULL) {
 950                 ptlrpc_reply(req);
 951         } else {
 952                 req->rq_status = rc;
 953                 ptlrpc_error(req);
 954         }
 955
 956         l_unlock(&ns->ns_lock);
 957         if (lock->l_granted_mode == LCK_PW &&
 958             !lock->l_readers && !lock->l_writers &&
 959             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 960                 if (ldlm_bl_to_thread(ns, NULL, lock))
 961                         ldlm_handle_bl_callback(ns, NULL, lock);
 962                 EXIT;
 963                 return;
 964         }
 965
 966         LDLM_LOCK_PUT(lock);
 967         EXIT;
 968 }
 969
 970 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 971 {
 972         req->rq_status = rc;
 973         if (req->rq_reply_state == NULL) {
 974                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 975                 if (rc)
 976                         return rc;
 977         }
 978         return ptlrpc_reply(req);
 979 }
 980
 981 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
 982                       struct ldlm_lock *lock)
 983 {
 984 #ifdef __KERNEL__
 985         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
 986         struct ldlm_bl_work_item *blwi;
 987         ENTRY;
 988
 989         OBD_ALLOC(blwi, sizeof(*blwi));
 990         if (blwi == NULL)
 991                 RETURN(-ENOMEM);
 992
 993         blwi->blwi_ns = ns;
 994         if (ld != NULL)
 995                 blwi->blwi_ld = *ld;
 996         blwi->blwi_lock = lock;
 997
 998         spin_lock(&blp->blp_lock);
 999         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1000         wake_up(&blp->blp_waitq);
1001         spin_unlock(&blp->blp_lock);
1002
1003         RETURN(0);
1004 #else
1005         RETURN(-ENOSYS);
1006 #endif
1007 }
1008
1009 static int ldlm_callback_handler(struct ptlrpc_request *req)
1010 {
1011         struct ldlm_namespace *ns;
1012         struct ldlm_request *dlm_req;
1013         struct ldlm_lock *lock;
1014         int rc;
1015         ENTRY;
1016
1017         /* Requests arrive in sender's byte order.  The ptlrpc service
1018          * handler has already checked and, if necessary, byte-swapped the
1019          * incoming request message body, but I am responsible for the
1020          * message buffers. */
1021
1022         if (req->rq_export == NULL) {
1023                 struct ldlm_request *dlm_req;
1024
1025                 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1026                        "export cookie "LPX64"; this is "
1027                        "normal if this node rebooted with a lock held\n",
1028                        req->rq_reqmsg->opc,
1029                        req->rq_peerstr,
1030                        req->rq_reqmsg->handle.cookie);
1031
1032                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1033                                              lustre_swab_ldlm_request);
1034                 if (dlm_req != NULL)
1035                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1036                                dlm_req->lock_handle1.cookie);
1037
1038                 ldlm_callback_reply(req, -ENOTCONN);
1039                 RETURN(0);
1040         }
1041
1042         LASSERT(req->rq_export != NULL);
1043         LASSERT(req->rq_export->exp_obd != NULL);
1044
1045         switch(req->rq_reqmsg->opc) {
1046         case LDLM_BL_CALLBACK:
1047                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1048                 break;
1049         case LDLM_CP_CALLBACK:
1050                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1051                 break;
1052         case LDLM_GL_CALLBACK:
1053                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1054                 break;
1055         case OBD_LOG_CANCEL:
1056                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1057                 rc = llog_origin_handle_cancel(req);
1058                 ldlm_callback_reply(req, rc);
1059                 RETURN(0);
1060         case LLOG_ORIGIN_HANDLE_CREATE:
1061                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1062                 rc = llog_origin_handle_create(req);
1063                 ldlm_callback_reply(req, rc);
1064                 RETURN(0);
1065         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1066                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1067                 rc = llog_origin_handle_next_block(req);
1068                 ldlm_callback_reply(req, rc);
1069                 RETURN(0);
1070         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1071                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1072                 rc = llog_origin_handle_read_header(req);
1073                 ldlm_callback_reply(req, rc);
1074                 RETURN(0);
1075         case LLOG_ORIGIN_HANDLE_CLOSE:
1076                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1077                 rc = llog_origin_handle_close(req);
1078                 ldlm_callback_reply(req, rc);
1079                 RETURN(0);
1080         default:
1081                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1082                 ldlm_callback_reply(req, -EPROTO);
1083                 RETURN(0);
1084         }
1085
1086         ns = req->rq_export->exp_obd->obd_namespace;
1087         LASSERT(ns != NULL);
1088
1089         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1090                                       lustre_swab_ldlm_request);
1091         if (dlm_req == NULL) {
1092                 CERROR ("can't unpack dlm_req\n");
1093                 ldlm_callback_reply (req, -EPROTO);
1094                 RETURN (0);
1095         }
1096
1097         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1098         if (!lock) {
1099                 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
1100                        dlm_req->lock_handle1.cookie);
1101                 ldlm_callback_reply(req, -EINVAL);
1102                 RETURN(0);
1103         }
1104
1105         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1106         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1107
1108         /* We want the ost thread to get this reply so that it can respond
1109          * to ost requests (write cache writeback) that might be triggered
1110          * in the callback.
1111          *
1112          * But we'd also like to be able to indicate in the reply that we're
1113          * cancelling right now, because it's unused, or have an intent result
1114          * in the reply, so we might have to push the responsibility for sending
1115          * the reply down into the AST handlers, alas. */
1116
1117         switch (req->rq_reqmsg->opc) {
1118         case LDLM_BL_CALLBACK:
1119                 CDEBUG(D_INODE, "blocking ast\n");
1120                 ldlm_callback_reply(req, 0);
1121                 if (ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock))
1122                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
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;
1143         ENTRY;
1144
1145         /* Requests arrive in sender's byte order.  The ptlrpc service
1146          * handler has already checked and, if necessary, byte-swapped the
1147          * incoming request message body, but I am responsible for the
1148          * message buffers. */
1149
1150         if (req->rq_export == NULL) {
1151                 struct ldlm_request *dlm_req;
1152                 CERROR("operation %d with bad export from %s\n",
1153                        req->rq_reqmsg->opc,
1154                        req->rq_peerstr);
1155                 CERROR("--> export cookie: "LPX64"\n",
1156                        req->rq_reqmsg->handle.cookie);
1157                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1158                                              lustre_swab_ldlm_request);
1159                 if (dlm_req != NULL)
1160                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1161                 ldlm_callback_reply(req, -ENOTCONN);
1162                 RETURN(0);
1163         }
1164
1165         switch (req->rq_reqmsg->opc) {
1166
1167         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1168         case LDLM_CANCEL:
1169                 CDEBUG(D_INODE, "cancel\n");
1170                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1171                 rc = ldlm_handle_cancel(req);
1172                 if (rc)
1173                         break;
1174                 RETURN(0);
1175
1176         default:
1177                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1178                 ldlm_callback_reply(req, -EINVAL);
1179         }
1180
1181         RETURN(0);
1182 }
1183
1184 #ifdef __KERNEL__
1185 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1186 {
1187         struct ldlm_bl_work_item *blwi = NULL;
1188
1189         spin_lock(&blp->blp_lock);
1190         if (!list_empty(&blp->blp_list)) {
1191                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1192                                   blwi_entry);
1193                 list_del(&blwi->blwi_entry);
1194         }
1195         spin_unlock(&blp->blp_lock);
1196
1197         return blwi;
1198 }
1199
1200 struct ldlm_bl_thread_data {
1201         int                     bltd_num;
1202         struct ldlm_bl_pool     *bltd_blp;
1203 };
1204
1205 static int ldlm_bl_thread_main(void *arg)
1206 {
1207         struct ldlm_bl_thread_data *bltd = arg;
1208         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1209         unsigned long flags;
1210         ENTRY;
1211
1212         /* XXX boiler-plate */
1213         {
1214                 char name[sizeof(current->comm)];
1215                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1216                          bltd->bltd_num);
1217                 kportal_daemonize(name);
1218         }
1219         SIGNAL_MASK_LOCK(current, flags);
1220         sigfillset(&current->blocked);
1221         RECALC_SIGPENDING;
1222         SIGNAL_MASK_UNLOCK(current, flags);
1223
1224         atomic_inc(&blp->blp_num_threads);
1225         complete(&blp->blp_comp);
1226
1227         while(1) {
1228                 struct l_wait_info lwi = { 0 };
1229                 struct ldlm_bl_work_item *blwi = NULL;
1230
1231                 l_wait_event_exclusive(blp->blp_waitq,
1232                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1233                                        &lwi);
1234
1235                 if (blwi->blwi_ns == NULL)
1236                         break;
1237
1238                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1239                                         blwi->blwi_lock);
1240                 OBD_FREE(blwi, sizeof(*blwi));
1241         }
1242
1243         atomic_dec(&blp->blp_num_threads);
1244         complete(&blp->blp_comp);
1245         RETURN(0);
1246 }
1247
1248 #endif
1249
1250 static int ldlm_setup(void);
1251 static int ldlm_cleanup(int force);
1252
1253 int ldlm_get_ref(void)
1254 {
1255         int rc = 0;
1256         down(&ldlm_ref_sem);
1257         if (++ldlm_refcount == 1) {
1258                 rc = ldlm_setup();
1259                 if (rc)
1260                         ldlm_refcount--;
1261         }
1262         up(&ldlm_ref_sem);
1263
1264         RETURN(rc);
1265 }
1266
1267 void ldlm_put_ref(int force)
1268 {
1269         down(&ldlm_ref_sem);
1270         if (ldlm_refcount == 1) {
1271                 int rc = ldlm_cleanup(force);
1272                 if (rc)
1273                         CERROR("ldlm_cleanup failed: %d\n", rc);
1274                 else
1275                         ldlm_refcount--;
1276         } else {
1277                 ldlm_refcount--;
1278         }
1279         up(&ldlm_ref_sem);
1280
1281         EXIT;
1282 }
1283
1284 static int ldlm_setup(void)
1285 {
1286         struct ldlm_bl_pool *blp;
1287         int rc = 0;
1288 #ifdef __KERNEL__
1289         int i;
1290 #endif
1291         ENTRY;
1292
1293         if (ldlm_state != NULL)
1294                 RETURN(-EALREADY);
1295
1296         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1297         if (ldlm_state == NULL)
1298                 RETURN(-ENOMEM);
1299
1300 #ifdef __KERNEL__
1301         rc = ldlm_proc_setup();
1302         if (rc != 0)
1303                 GOTO(out_free, rc);
1304 #endif
1305
1306         ldlm_state->ldlm_cb_service =
1307                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1308                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1309                                 1500, ldlm_callback_handler, "ldlm_cbd",
1310                                 ldlm_svc_proc_dir, NULL);
1311
1312         if (!ldlm_state->ldlm_cb_service) {
1313                 CERROR("failed to start service\n");
1314                 GOTO(out_proc, rc = -ENOMEM);
1315         }
1316
1317         ldlm_state->ldlm_cancel_service =
1318                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1319                                 LDLM_CANCEL_REQUEST_PORTAL,
1320                                 LDLM_CANCEL_REPLY_PORTAL, 30000,
1321                                 ldlm_cancel_handler, "ldlm_canceld",
1322                                 ldlm_svc_proc_dir, NULL);
1323
1324         if (!ldlm_state->ldlm_cancel_service) {
1325                 CERROR("failed to start service\n");
1326                 GOTO(out_proc, rc = -ENOMEM);
1327         }
1328
1329         OBD_ALLOC(blp, sizeof(*blp));
1330         if (blp == NULL)
1331                 GOTO(out_proc, rc = -ENOMEM);
1332         ldlm_state->ldlm_bl_pool = blp;
1333
1334         atomic_set(&blp->blp_num_threads, 0);
1335         init_waitqueue_head(&blp->blp_waitq);
1336         spin_lock_init(&blp->blp_lock);
1337
1338         INIT_LIST_HEAD(&blp->blp_list);
1339
1340 #ifdef __KERNEL__
1341         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1342                 struct ldlm_bl_thread_data bltd = {
1343                         .bltd_num = i,
1344                         .bltd_blp = blp,
1345                 };
1346                 init_completion(&blp->blp_comp);
1347                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1348                 if (rc < 0) {
1349                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1350                         GOTO(out_thread, rc);
1351                 }
1352                 wait_for_completion(&blp->blp_comp);
1353         }
1354
1355         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1356                                     LDLM_NUM_THREADS, "ldlm_cn");
1357         if (rc)
1358                 GOTO(out_thread, rc);
1359
1360         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1361                                     LDLM_NUM_THREADS, "ldlm_cb");
1362         if (rc)
1363                 GOTO(out_thread, rc);
1364
1365         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1366         expired_lock_thread.elt_state = ELT_STOPPED;
1367         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1368
1369         INIT_LIST_HEAD(&waiting_locks_list);
1370         spin_lock_init(&waiting_locks_spinlock);
1371         waiting_locks_timer.function = waiting_locks_callback;
1372         waiting_locks_timer.data = 0;
1373         init_timer(&waiting_locks_timer);
1374
1375         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1376         if (rc < 0) {
1377                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1378                 GOTO(out_thread, rc);
1379         }
1380
1381         wait_event(expired_lock_thread.elt_waitq,
1382                    expired_lock_thread.elt_state == ELT_READY);
1383 #endif
1384
1385         RETURN(0);
1386
1387 #ifdef __KERNEL__
1388  out_thread:
1389         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1390         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1391 #endif
1392
1393  out_proc:
1394 #ifdef __KERNEL__
1395         ldlm_proc_cleanup();
1396  out_free:
1397 #endif
1398         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1399         ldlm_state = NULL;
1400         return rc;
1401 }
1402
1403 static int ldlm_cleanup(int force)
1404 {
1405 #ifdef __KERNEL__
1406         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1407 #endif
1408         ENTRY;
1409
1410         if (!list_empty(&ldlm_namespace_list)) {
1411                 CERROR("ldlm still has namespaces; clean these up first.\n");
1412                 ldlm_dump_all_namespaces(D_DLMTRACE);
1413                 RETURN(-EBUSY);
1414         }
1415
1416 #ifdef __KERNEL__
1417         while (atomic_read(&blp->blp_num_threads) > 0) {
1418                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1419
1420                 init_completion(&blp->blp_comp);
1421
1422                 spin_lock(&blp->blp_lock);
1423                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1424                 wake_up(&blp->blp_waitq);
1425                 spin_unlock(&blp->blp_lock);
1426
1427                 wait_for_completion(&blp->blp_comp);
1428         }
1429         OBD_FREE(blp, sizeof(*blp));
1430
1431         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1432         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1433         ldlm_proc_cleanup();
1434
1435         expired_lock_thread.elt_state = ELT_TERMINATE;
1436         wake_up(&expired_lock_thread.elt_waitq);
1437         wait_event(expired_lock_thread.elt_waitq,
1438                    expired_lock_thread.elt_state == ELT_STOPPED);
1439
1440 #endif
1441
1442         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1443         ldlm_state = NULL;
1444
1445         RETURN(0);
1446 }
1447
1448 int __init ldlm_init(void)
1449 {
1450         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1451                                                sizeof(struct ldlm_resource), 0,
1452                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1453         if (ldlm_resource_slab == NULL)
1454                 return -ENOMEM;
1455
1456         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1457                                            sizeof(struct ldlm_lock), 0,
1458                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1459         if (ldlm_lock_slab == NULL) {
1460                 kmem_cache_destroy(ldlm_resource_slab);
1461                 return -ENOMEM;
1462         }
1463
1464         l_lock_init(&ldlm_handle_lock);
1465
1466         return 0;
1467 }
1468
1469 void __exit ldlm_exit(void)
1470 {
1471         if ( ldlm_refcount )
1472                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1473         LASSERTF(kmem_cache_destroy(ldlm_resource_slab) == 0,
1474                  "couldn't free ldlm resource slab\n");
1475         LASSERTF(kmem_cache_destroy(ldlm_lock_slab) == 0,
1476                  "couldn't free ldlm lock slab\n");
1477 }
1478
1479 /* ldlm_flock.c */
1480 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1481
1482 /* ldlm_extent.c */
1483 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1484
1485 /* ldlm_lock.c */
1486 EXPORT_SYMBOL(ldlm_get_processing_policy);
1487 EXPORT_SYMBOL(ldlm_lock2desc);
1488 EXPORT_SYMBOL(ldlm_register_intent);
1489 EXPORT_SYMBOL(ldlm_lockname);
1490 EXPORT_SYMBOL(ldlm_typename);
1491 EXPORT_SYMBOL(ldlm_lock2handle);
1492 EXPORT_SYMBOL(__ldlm_handle2lock);
1493 EXPORT_SYMBOL(ldlm_lock_get);
1494 EXPORT_SYMBOL(ldlm_lock_put);
1495 EXPORT_SYMBOL(ldlm_lock_match);
1496 EXPORT_SYMBOL(ldlm_lock_cancel);
1497 EXPORT_SYMBOL(ldlm_lock_addref);
1498 EXPORT_SYMBOL(ldlm_lock_decref);
1499 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1500 EXPORT_SYMBOL(ldlm_lock_change_resource);
1501 EXPORT_SYMBOL(ldlm_lock_set_data);
1502 EXPORT_SYMBOL(ldlm_it2str);
1503 EXPORT_SYMBOL(ldlm_lock_dump);
1504 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1505 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1506 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1507 EXPORT_SYMBOL(ldlm_lock_allow_match);
1508
1509 /* ldlm_request.c */
1510 EXPORT_SYMBOL(ldlm_completion_ast);
1511 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1512 EXPORT_SYMBOL(ldlm_cli_convert);
1513 EXPORT_SYMBOL(ldlm_cli_enqueue);
1514 EXPORT_SYMBOL(ldlm_cli_cancel);
1515 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1516 EXPORT_SYMBOL(ldlm_replay_locks);
1517 EXPORT_SYMBOL(ldlm_resource_foreach);
1518 EXPORT_SYMBOL(ldlm_namespace_foreach);
1519 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1520 EXPORT_SYMBOL(ldlm_change_cbdata);
1521
1522 /* ldlm_lockd.c */
1523 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1524 EXPORT_SYMBOL(ldlm_server_completion_ast);
1525 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1526 EXPORT_SYMBOL(ldlm_handle_enqueue);
1527 EXPORT_SYMBOL(ldlm_handle_cancel);
1528 EXPORT_SYMBOL(ldlm_handle_convert);
1529 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1530 EXPORT_SYMBOL(ldlm_get_ref);
1531 EXPORT_SYMBOL(ldlm_put_ref);
1532
1533 #if 0
1534 /* ldlm_test.c */
1535 EXPORT_SYMBOL(ldlm_test);
1536 EXPORT_SYMBOL(ldlm_regression_start);
1537 EXPORT_SYMBOL(ldlm_regression_stop);
1538 #endif
1539
1540 /* ldlm_resource.c */
1541 EXPORT_SYMBOL(ldlm_namespace_new);
1542 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1543 EXPORT_SYMBOL(ldlm_namespace_free);
1544 EXPORT_SYMBOL(ldlm_namespace_dump);
1545 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1546 EXPORT_SYMBOL(ldlm_resource_get);
1547 EXPORT_SYMBOL(ldlm_resource_putref);
1548
1549 /* l_lock.c */
1550 EXPORT_SYMBOL(l_lock);
1551 EXPORT_SYMBOL(l_unlock);
1552
1553 /* ldlm_lib.c */
1554 EXPORT_SYMBOL(client_obd_setup);
1555 EXPORT_SYMBOL(client_obd_cleanup);
1556 EXPORT_SYMBOL(client_connect_import);
1557 EXPORT_SYMBOL(client_disconnect_export);
1558 EXPORT_SYMBOL(target_abort_recovery);
1559 EXPORT_SYMBOL(target_cleanup_recovery);
1560 EXPORT_SYMBOL(target_handle_connect);
1561 EXPORT_SYMBOL(target_destroy_export);
1562 EXPORT_SYMBOL(target_cancel_recovery_timer);
1563 EXPORT_SYMBOL(target_send_reply);
1564 EXPORT_SYMBOL(target_queue_recovery_request);
1565 EXPORT_SYMBOL(target_handle_ping);
1566 EXPORT_SYMBOL(target_handle_disconnect);
1567 EXPORT_SYMBOL(target_queue_final_reply);