lustre/ldlm/ldlm_lockd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #ifndef EXPORT_SYMTAB
  25 # define EXPORT_SYMTAB
  26 #endif
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28
  29 #ifdef __KERNEL__
  30 # include <linux/module.h>
  31 # include <linux/slab.h>
  32 # include <linux/init.h>
  33 # include <linux/wait.h>
  34 #else
  35 # include <liblustre.h>
  36 #endif
  37
  38 #include <linux/lustre_dlm.h>
  39 #include <linux/obd_class.h>
  40 #include <portals/list.h>
  41 #include "ldlm_internal.h"
  42
  43 extern kmem_cache_t *ldlm_resource_slab;
  44 extern kmem_cache_t *ldlm_lock_slab;
  45 extern struct lustre_lock ldlm_handle_lock;
  46 extern struct list_head ldlm_namespace_list;
  47 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
  48 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
  49
  50 static DECLARE_MUTEX(ldlm_ref_sem);
  51 static int ldlm_refcount = 0;
  52
  53 /* LDLM state */
  54
  55 static struct ldlm_state *ldlm_state;
  56
  57 inline unsigned long round_timeout(unsigned long timeout)
  58 {
  59         return ((timeout / HZ) + 1) * HZ;
  60 }
  61
  62 #ifdef __KERNEL__
  63 /* XXX should this be per-ldlm? */
  64 static struct list_head waiting_locks_list;
  65 static spinlock_t waiting_locks_spinlock;
  66 static struct timer_list waiting_locks_timer;
  67
  68 static struct expired_lock_thread {
  69         wait_queue_head_t         elt_waitq;
  70         int                       elt_state;
  71         struct list_head          elt_expired_locks;
  72         spinlock_t                elt_lock;
  73 } expired_lock_thread;
  74 #endif
  75
  76 #define ELT_STOPPED   0
  77 #define ELT_READY     1
  78 #define ELT_TERMINATE 2
  79
  80 struct ldlm_bl_pool {
  81         spinlock_t              blp_lock;
  82         struct list_head        blp_list;
  83         wait_queue_head_t       blp_waitq;
  84         atomic_t                blp_num_threads;
  85         struct completion       blp_comp;
  86 };
  87
  88 struct ldlm_bl_work_item {
  89         struct list_head        blwi_entry;
  90         struct ldlm_namespace   *blwi_ns;
  91         struct ldlm_lock_desc   blwi_ld;
  92         struct ldlm_lock        *blwi_lock;
  93 };
  94
  95 #ifdef __KERNEL__
  96
  97 static inline int have_expired_locks(void)
  98 {
  99         int need_to_run;
 100
 101         spin_lock_bh(&expired_lock_thread.elt_lock);
 102         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 103         spin_unlock_bh(&expired_lock_thread.elt_lock);
 104
 105         RETURN(need_to_run);
 106 }
 107
 108 static int expired_lock_main(void *arg)
 109 {
 110         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 111         struct l_wait_info lwi = { 0 };
 112         unsigned long flags;
 113
 114         ENTRY;
 115         lock_kernel();
 116         kportal_daemonize("ldlm_elt");
 117
 118         SIGNAL_MASK_LOCK(current, flags);
 119         sigfillset(&current->blocked);
 120         RECALC_SIGPENDING;
 121         SIGNAL_MASK_UNLOCK(current, flags);
 122
 123         unlock_kernel();
 124
 125         expired_lock_thread.elt_state = ELT_READY;
 126         wake_up(&expired_lock_thread.elt_waitq);
 127
 128         while (1) {
 129                 struct list_head *tmp, *n, work_list;
 130                 l_wait_event(expired_lock_thread.elt_waitq,
 131                              have_expired_locks() ||
 132                              expired_lock_thread.elt_state == ELT_TERMINATE,
 133                              &lwi);
 134
 135                 spin_lock_bh(&expired_lock_thread.elt_lock);
 136                 while (!list_empty(expired)) {
 137                         struct ldlm_lock *lock;
 138
 139                         list_add(&work_list, expired);
 140                         list_del_init(expired);
 141
 142                         list_for_each_entry(lock, &work_list, l_pending_chain) {
 143                                 LDLM_DEBUG(lock, "moving to work list");
 144                         }
 145
 146                         spin_unlock_bh(&expired_lock_thread.elt_lock);
 147
 148
 149                         list_for_each_safe(tmp, n, &work_list) {
 150                                  lock = list_entry(tmp, struct ldlm_lock,
 151                                                    l_pending_chain);
 152                                  ptlrpc_fail_export(lock->l_export);
 153                         }
 154
 155
 156                         if (!list_empty(&work_list)) {
 157                                 list_for_each_entry(lock, &work_list, l_pending_chain) {
 158                                         LDLM_ERROR(lock, "still on work list!");
 159                                 }
 160                         }
 161                         LASSERTF (list_empty(&work_list),
 162                                   "some exports not failed properly\n");
 163
 164                         spin_lock_bh(&expired_lock_thread.elt_lock);
 165                 }
 166                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 167
 168                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 169                         break;
 170         }
 171
 172         expired_lock_thread.elt_state = ELT_STOPPED;
 173         wake_up(&expired_lock_thread.elt_waitq);
 174         RETURN(0);
 175 }
 176
 177 static void waiting_locks_callback(unsigned long unused)
 178 {
 179         struct ldlm_lock *lock;
 180         char str[PTL_NALFMT_SIZE];
 181
 182         spin_lock_bh(&waiting_locks_spinlock);
 183         while (!list_empty(&waiting_locks_list)) {
 184                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 185                                   l_pending_chain);
 186
 187                 if ((lock->l_callback_timeout > jiffies) ||
 188                     (lock->l_req_mode == LCK_CW))
 189                         break;
 190
 191                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 192                            "%s@%s nid "LPX64" (%s) ",
 193                            lock->l_export->exp_client_uuid.uuid,
 194                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 195                            lock->l_export->exp_connection->c_peer.peer_nid,
 196                            portals_nid2str(lock->l_export->exp_connection->c_peer.peer_ni->pni_number,
 197                                            lock->l_export->exp_connection->c_peer.peer_nid,
 198                                            str));
 199
 200                 spin_lock_bh(&expired_lock_thread.elt_lock);
 201                 list_del(&lock->l_pending_chain);
 202                 list_add(&lock->l_pending_chain,
 203                          &expired_lock_thread.elt_expired_locks);
 204                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 205                 wake_up(&expired_lock_thread.elt_waitq);
 206         }
 207
 208         /*
 209          * Make sure the timer will fire again if we have any locks
 210          * left.
 211          */
 212         if (!list_empty(&waiting_locks_list)) {
 213                 unsigned long timeout_rounded;
 214                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 215                                   l_pending_chain);
 216                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 217                 mod_timer(&waiting_locks_timer, timeout_rounded);
 218         }
 219         spin_unlock_bh(&waiting_locks_spinlock);
 220 }
 221
 222 /*
 223  * Indicate that we're waiting for a client to call us back cancelling a given
 224  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 225  * timer to fire appropriately.  (We round up to the next second, to avoid
 226  * floods of timer firings during periods of high lock contention and traffic).
 227  */
 228 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 229 {
 230         unsigned long timeout_rounded;
 231
 232         spin_lock_bh(&waiting_locks_spinlock);
 233         if (!list_empty(&lock->l_pending_chain)) {
 234                 LDLM_DEBUG(lock, "not re-adding to wait list");
 235                 spin_unlock_bh(&waiting_locks_spinlock);
 236                 return 0;
 237         }
 238         LDLM_DEBUG(lock, "adding to wait list");
 239
 240         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 241
 242         timeout_rounded = round_timeout(lock->l_callback_timeout);
 243
 244         if (timeout_rounded < waiting_locks_timer.expires ||
 245             !timer_pending(&waiting_locks_timer)) {
 246                 mod_timer(&waiting_locks_timer, timeout_rounded);
 247         }
 248         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 249         spin_unlock_bh(&waiting_locks_spinlock);
 250         return 1;
 251 }
 252
 253 /*
 254  * Remove a lock from the pending list, likely because it had its cancellation
 255  * callback arrive without incident.  This adjusts the lock-timeout timer if
 256  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 257  */
 258 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 259 {
 260         struct list_head *list_next;
 261
 262         if (lock->l_export == NULL) {
 263                 /* We don't have a "waiting locks list" on clients. */
 264                 LDLM_DEBUG(lock, "client lock: no-op");
 265                 return 0;
 266         }
 267
 268         spin_lock_bh(&waiting_locks_spinlock);
 269
 270         if (list_empty(&lock->l_pending_chain)) {
 271                 spin_unlock_bh(&waiting_locks_spinlock);
 272                 LDLM_DEBUG(lock, "wasn't waiting");
 273                 return 0;
 274         }
 275
 276         list_next = lock->l_pending_chain.next;
 277         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 278                 /* Removing the head of the list, adjust timer. */
 279                 if (list_next == &waiting_locks_list) {
 280                         /* No more, just cancel. */
 281                         del_timer(&waiting_locks_timer);
 282                 } else {
 283                         struct ldlm_lock *next;
 284                         next = list_entry(list_next, struct ldlm_lock,
 285                                           l_pending_chain);
 286                         mod_timer(&waiting_locks_timer,
 287                                   round_timeout(next->l_callback_timeout));
 288                 }
 289         }
 290
 291         spin_lock_bh(&expired_lock_thread.elt_lock);
 292         list_del_init(&lock->l_pending_chain);
 293         spin_unlock_bh(&expired_lock_thread.elt_lock);
 294
 295         spin_unlock_bh(&waiting_locks_spinlock);
 296         LDLM_DEBUG(lock, "removed");
 297         return 1;
 298 }
 299
 300 #else /* !__KERNEL__ */
 301
 302 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 303 {
 304         RETURN(1);
 305 }
 306
 307 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 308 {
 309         RETURN(0);
 310 }
 311
 312 #endif /* __KERNEL__ */
 313
 314 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, char *ast_type)
 315 {
 316         const struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 317         char str[PTL_NALFMT_SIZE];
 318
 319         CERROR("%s AST failed (%d) for res "LPU64"/"LPU64
 320                ", mode %s: evicting client %s@%s NID "LPX64" (%s)\n",
 321                ast_type, rc,
 322                lock->l_resource->lr_name.name[0],
 323                lock->l_resource->lr_name.name[1],
 324                ldlm_lockname[lock->l_granted_mode],
 325                lock->l_export->exp_client_uuid.uuid,
 326                conn->c_remote_uuid.uuid, conn->c_peer.peer_nid,
 327                portals_nid2str(conn->c_peer.peer_ni->pni_number,
 328                                conn->c_peer.peer_nid, str));
 329         ptlrpc_fail_export(lock->l_export);
 330 }
 331
 332 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 333                              struct ldlm_lock_desc *desc,
 334                              void *data, int flag)
 335 {
 336         struct ldlm_request *body;
 337         struct ptlrpc_request *req;
 338         int rc = 0, size = sizeof(*body);
 339         ENTRY;
 340
 341         if (flag == LDLM_CB_CANCELING) {
 342                 /* Don't need to do anything here. */
 343                 RETURN(0);
 344         }
 345
 346         LASSERT(lock);
 347
 348         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 349         if (lock->l_granted_mode != lock->l_req_mode) {
 350                 /* this blocking AST will be communicated as part of the
 351                  * completion AST instead */
 352                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 353                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");                 RETURN(0);
 354         }
 355
 356         if (lock->l_destroyed) {
 357                 /* What's the point? */
 358                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 359                 RETURN(0);
 360         }
 361
 362 #if 0
 363         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 364                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 365                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 366                 RETURN(-ETIMEDOUT);
 367         }
 368 #endif
 369
 370         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 371                               LDLM_BL_CALLBACK, 1, &size, NULL);
 372         if (req == NULL) {
 373                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 374                 RETURN(-ENOMEM);
 375         }
 376
 377         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 378         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 379                sizeof(body->lock_handle1));
 380         memcpy(&body->lock_desc, desc, sizeof(*desc));
 381         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 382
 383         LDLM_DEBUG(lock, "server preparing blocking AST");
 384         req->rq_replen = lustre_msg_size(0, NULL);
 385
 386         if (lock->l_granted_mode == lock->l_req_mode)
 387                 ldlm_add_waiting_lock(lock);
 388         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 389
 390         req->rq_send_state = LUSTRE_IMP_FULL;
 391         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 392         rc = ptlrpc_queue_wait(req);
 393         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 394                 LASSERT(lock->l_export);
 395                 if (lock->l_export->exp_libclient) {
 396                         CDEBUG(D_HA, "BLOCKING AST to liblustre client (nid "
 397                                LPU64") timeout, simply cancel lock 0x%p\n",
 398                                req->rq_peer.peer_nid, lock);
 399                         ldlm_lock_cancel(lock);
 400                         rc = -ERESTART;
 401                 } else {
 402                         ldlm_del_waiting_lock(lock);
 403                         ldlm_failed_ast(lock, rc, "blocking");
 404                 }
 405         } else if (rc) {
 406                 if (rc == -EINVAL)
 407                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 408                                "from blocking AST for lock %p--normal race\n",
 409                                req->rq_peer.peer_nid,
 410                                req->rq_repmsg->status, lock);
 411                 else if (rc == -ENOTCONN)
 412                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 413                                "from blocking AST for lock %p--this client was "
 414                                "probably rebooted while it held a lock, nothing"
 415                                " serious\n",req->rq_peer.peer_nid,
 416                                req->rq_repmsg->status, lock);
 417                 else
 418                         CDEBUG(D_ERROR, "client (nid "LPU64") returned %d "
 419                                "from blocking AST for lock %p\n",
 420                                req->rq_peer.peer_nid,
 421                                (req->rq_repmsg != NULL)?
 422                                req->rq_repmsg->status : 0,
 423                                lock);
 424                 LDLM_DEBUG(lock, "client sent rc %d rq_status %d from blocking "
 425                            "AST", rc, req->rq_status);
 426                 ldlm_lock_cancel(lock);
 427                 /* Server-side AST functions are called from ldlm_reprocess_all,
 428                  * which needs to be told to please restart its reprocessing. */
 429                 rc = -ERESTART;
 430         }
 431
 432         ptlrpc_req_finished(req);
 433
 434         RETURN(rc);
 435 }
 436
 437 /* XXX copied from ptlrpc/service.c */
 438 static long timeval_sub(struct timeval *large, struct timeval *small)
 439 {
 440         return (large->tv_sec - small->tv_sec) * 1000000 +
 441                 (large->tv_usec - small->tv_usec);
 442 }
 443
 444 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 445 {
 446         struct ldlm_request *body;
 447         struct ptlrpc_request *req;
 448         struct timeval granted_time;
 449         long total_enqueue_wait;
 450         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 451         ENTRY;
 452
 453         LASSERT(lock != NULL);
 454
 455         do_gettimeofday(&granted_time);
 456         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 457
 458         if (total_enqueue_wait / 1000000 > obd_timeout)
 459                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 460
 461         if (lock->l_resource->lr_lvb_len) {
 462                 buffers = 2;
 463                 size[1] = lock->l_resource->lr_lvb_len;
 464         }
 465
 466         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 467                               LDLM_CP_CALLBACK, buffers, size, NULL);
 468         if (req == NULL)
 469                 RETURN(-ENOMEM);
 470
 471         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 472         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 473                sizeof(body->lock_handle1));
 474         body->lock_flags = flags;
 475         ldlm_lock2desc(lock, &body->lock_desc);
 476
 477         if (buffers == 2) {
 478                 void *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         }
 483
 484         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 485                    total_enqueue_wait);
 486         req->rq_replen = lustre_msg_size(0, NULL);
 487
 488         req->rq_send_state = LUSTRE_IMP_FULL;
 489         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 490
 491         /* We only send real blocking ASTs after the lock is granted */
 492         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 493         if (lock->l_flags & LDLM_FL_AST_SENT) {
 494                 body->lock_flags |= LDLM_FL_AST_SENT;
 495                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 496         }
 497         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 498
 499         rc = ptlrpc_queue_wait(req);
 500         if ((rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) &&
 501              !lock->l_export->exp_libclient) {
 502                 ldlm_del_waiting_lock(lock);
 503                 ldlm_failed_ast(lock, rc, "completion");
 504         } else if (rc == -EINVAL) {
 505                 LDLM_DEBUG(lock, "lost the race -- client no longer has this "
 506                            "lock");
 507         } else if (rc) {
 508                 LDLM_ERROR(lock, "client sent rc %d rq_status %d from "
 509                            "completion AST", rc, req->rq_status);
 510                 ldlm_lock_cancel(lock);
 511                 /* Server-side AST functions are called from ldlm_reprocess_all,
 512                  * which needs to be told to please restart its reprocessing. */
 513                 rc = -ERESTART;
 514         }
 515         ptlrpc_req_finished(req);
 516
 517         RETURN(rc);
 518 }
 519
 520 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 521 {
 522         struct ldlm_resource *res = lock->l_resource;
 523         struct ldlm_request *body;
 524         struct ptlrpc_request *req;
 525         int rc = 0, size = sizeof(*body);
 526         ENTRY;
 527
 528         LASSERT(lock != NULL);
 529
 530         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 531                               LDLM_GL_CALLBACK, 1, &size, NULL);
 532         if (req == NULL)
 533                 RETURN(-ENOMEM);
 534
 535         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 536         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 537                sizeof(body->lock_handle1));
 538         ldlm_lock2desc(lock, &body->lock_desc);
 539
 540         size = lock->l_resource->lr_lvb_len;
 541         req->rq_replen = lustre_msg_size(1, &size);
 542
 543         req->rq_send_state = LUSTRE_IMP_FULL;
 544         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 545
 546         rc = ptlrpc_queue_wait(req);
 547         if ((rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) &&
 548             !lock->l_export->exp_libclient) {
 549                 ldlm_del_waiting_lock(lock);
 550                 ldlm_failed_ast(lock, rc, "glimpse");
 551         } else if (rc == -EINVAL) {
 552                 LDLM_DEBUG(lock, "lost the race -- client no longer has this "
 553                            "lock");
 554         } else if (rc == -ELDLM_NO_LOCK_DATA) {
 555                 LDLM_DEBUG(lock, "lost a race -- client has a lock, but no "
 556                            "inode");
 557         } else if (rc) {
 558                 LDLM_ERROR(lock, "client sent rc %d rq_status %d from "
 559                            "glimpse AST", rc, req->rq_status);
 560         } else {
 561                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 562                         (res, req->rq_repmsg, 0, 1);
 563         }
 564         ptlrpc_req_finished(req);
 565         RETURN(rc);
 566 }
 567
 568 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 569                         ldlm_completion_callback completion_callback,
 570                         ldlm_blocking_callback blocking_callback,
 571                         ldlm_glimpse_callback glimpse_callback)
 572 {
 573         struct obd_device *obddev = req->rq_export->exp_obd;
 574         struct ldlm_reply *dlm_rep;
 575         struct ldlm_request *dlm_req;
 576         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 577         __u32 flags;
 578         ldlm_error_t err = ELDLM_OK;
 579         struct ldlm_lock *lock = NULL;
 580         void *cookie = NULL;
 581         ENTRY;
 582
 583         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 584
 585         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 586                                       lustre_swab_ldlm_request);
 587         if (dlm_req == NULL) {
 588                 CERROR ("Can't unpack dlm_req\n");
 589                 GOTO(out, rc = -EFAULT);
 590         }
 591
 592         flags = dlm_req->lock_flags;
 593
 594         /* The lock's callback data might be set in the policy function */
 595         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 596                                 dlm_req->lock_desc.l_resource.lr_name,
 597                                 dlm_req->lock_desc.l_resource.lr_type,
 598                                 dlm_req->lock_desc.l_req_mode,
 599                                 blocking_callback, completion_callback,
 600                                 glimpse_callback, NULL, 0);
 601         if (!lock)
 602                 GOTO(out, rc = -ENOMEM);
 603
 604         do_gettimeofday(&lock->l_enqueued_time);
 605         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 606                sizeof(lock->l_remote_handle));
 607         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 608
 609         LASSERT(req->rq_export);
 610         lock->l_export = class_export_get(req->rq_export);
 611         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 612         list_add(&lock->l_export_chain,
 613                  &lock->l_export->exp_ldlm_data.led_held_locks);
 614         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 615
 616         if (flags & LDLM_FL_HAS_INTENT) {
 617                 /* In this case, the reply buffer is allocated deep in
 618                  * local_lock_enqueue by the policy function. */
 619                 cookie = req;
 620         } else {
 621                 int buffers = 1;
 622                 if (lock->l_resource->lr_lvb_len) {
 623                         size[1] = lock->l_resource->lr_lvb_len;
 624                         buffers = 2;
 625                 }
 626
 627                 rc = lustre_pack_reply(req, buffers, size, NULL);
 628                 if (rc)
 629                         GOTO(out, rc);
 630         }
 631
 632         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 633                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 634                        sizeof(ldlm_policy_data_t));
 635         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 636                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 637                        sizeof(lock->l_req_extent));
 638
 639         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 640         if (err)
 641                 GOTO(out, err);
 642
 643         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 644         dlm_rep->lock_flags = flags;
 645
 646         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 647         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 648
 649         /* We never send a blocking AST until the lock is granted, but
 650          * we can tell it right now */
 651         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 652         if (lock->l_flags & LDLM_FL_AST_SENT) {
 653                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 654                 if (lock->l_granted_mode == lock->l_req_mode)
 655                         ldlm_add_waiting_lock(lock);
 656         }
 657         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 658
 659         EXIT;
 660  out:
 661         req->rq_status = err;
 662         if (req->rq_reply_state == NULL) {
 663                 err = lustre_pack_reply(req, 0, NULL, NULL);
 664                 if (rc == 0)
 665                         rc = err;
 666         }
 667
 668         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 669          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 670         if (lock) {
 671                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 672                            "(err=%d, rc=%d)", err, rc);
 673
 674                 if (lock->l_resource->lr_lvb_len > 0) {
 675                         void *lvb = lustre_msg_buf(req->rq_repmsg, 1,
 676                                                   lock->l_resource->lr_lvb_len);
 677                         memcpy(lvb, lock->l_resource->lr_lvb_data,
 678                                lock->l_resource->lr_lvb_len);
 679                 }
 680
 681                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 682                         ldlm_reprocess_all(lock->l_resource);
 683                 LDLM_LOCK_PUT(lock);
 684         }
 685         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 686                           lock, rc);
 687
 688         return rc;
 689 }
 690
 691 int ldlm_handle_convert(struct ptlrpc_request *req)
 692 {
 693         struct ldlm_request *dlm_req;
 694         struct ldlm_reply *dlm_rep;
 695         struct ldlm_lock *lock;
 696         int rc, size = sizeof(*dlm_rep);
 697         ENTRY;
 698
 699         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 700                                       lustre_swab_ldlm_request);
 701         if (dlm_req == NULL) {
 702                 CERROR ("Can't unpack dlm_req\n");
 703                 RETURN (-EFAULT);
 704         }
 705
 706         rc = lustre_pack_reply(req, 1, &size, NULL);
 707         if (rc) {
 708                 CERROR("out of memory\n");
 709                 RETURN(-ENOMEM);
 710         }
 711         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 712         dlm_rep->lock_flags = dlm_req->lock_flags;
 713
 714         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 715         if (!lock) {
 716                 req->rq_status = EINVAL;
 717         } else {
 718                 LDLM_DEBUG(lock, "server-side convert handler START");
 719                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 720                                   &dlm_rep->lock_flags);
 721                 if (ldlm_del_waiting_lock(lock))
 722                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 723                 req->rq_status = 0;
 724         }
 725
 726         if (lock) {
 727                 ldlm_reprocess_all(lock->l_resource);
 728                 LDLM_DEBUG(lock, "server-side convert handler END");
 729                 LDLM_LOCK_PUT(lock);
 730         } else
 731                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 732
 733         RETURN(0);
 734 }
 735
 736 int ldlm_handle_cancel(struct ptlrpc_request *req)
 737 {
 738         struct ldlm_request *dlm_req;
 739         struct ldlm_lock *lock;
 740         struct ldlm_resource *res;
 741         char str[PTL_NALFMT_SIZE];
 742         int rc;
 743         ENTRY;
 744
 745         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 746                                       lustre_swab_ldlm_request);
 747         if (dlm_req == NULL) {
 748                 CERROR("bad request buffer for cancel\n");
 749                 RETURN(-EFAULT);
 750         }
 751
 752         rc = lustre_pack_reply(req, 0, NULL, NULL);
 753         if (rc) {
 754                 CERROR("out of memory\n");
 755                 RETURN(-ENOMEM);
 756         }
 757
 758         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 759         if (!lock) {
 760                 CERROR("received cancel for unknown lock cookie "LPX64
 761                        " from client %s nid "LPX64" (%s)\n",
 762                        dlm_req->lock_handle1.cookie,
 763                        req->rq_export->exp_client_uuid.uuid,
 764                        req->rq_peer.peer_nid,
 765                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 766                                        req->rq_peer.peer_nid, str));
 767                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 768                                   "(cookie "LPU64")",
 769                                   dlm_req->lock_handle1.cookie);
 770                 req->rq_status = ESTALE;
 771         } else {
 772                 LDLM_DEBUG(lock, "server-side cancel handler START");
 773                 res = lock->l_resource;
 774                 if (res && res->lr_namespace->ns_lvbo &&
 775                     res->lr_namespace->ns_lvbo->lvbo_update) {
 776                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 777                                 (res, NULL, 0, 0);
 778                                 //(res, req->rq_reqmsg, 1);
 779                 }
 780
 781                 ldlm_lock_cancel(lock);
 782                 if (ldlm_del_waiting_lock(lock))
 783                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 784                 req->rq_status = rc;
 785         }
 786
 787         if (ptlrpc_reply(req) != 0)
 788                 LBUG();
 789
 790         if (lock) {
 791                 ldlm_reprocess_all(lock->l_resource);
 792                 LDLM_DEBUG(lock, "server-side cancel handler END");
 793                 LDLM_LOCK_PUT(lock);
 794         }
 795
 796         RETURN(0);
 797 }
 798
 799 static void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 800                                     struct ldlm_lock_desc *ld,
 801                                     struct ldlm_lock *lock)
 802 {
 803         int do_ast;
 804         ENTRY;
 805
 806         l_lock(&ns->ns_lock);
 807         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 808
 809         lock->l_flags |= LDLM_FL_CBPENDING;
 810         do_ast = (!lock->l_readers && !lock->l_writers);
 811
 812         if (do_ast) {
 813                 LDLM_DEBUG(lock, "already unused, calling "
 814                            "callback (%p)", lock->l_blocking_ast);
 815                 if (lock->l_blocking_ast != NULL) {
 816                         l_unlock(&ns->ns_lock);
 817                         l_check_no_ns_lock(ns);
 818                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 819                                              LDLM_CB_BLOCKING);
 820                         l_lock(&ns->ns_lock);
 821                 }
 822         } else {
 823                 LDLM_DEBUG(lock, "Lock still has references, will be"
 824                            " cancelled later");
 825         }
 826
 827         LDLM_DEBUG(lock, "client blocking callback handler END");
 828         l_unlock(&ns->ns_lock);
 829         LDLM_LOCK_PUT(lock);
 830         EXIT;
 831 }
 832
 833 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 834                                     struct ldlm_namespace *ns,
 835                                     struct ldlm_request *dlm_req,
 836                                     struct ldlm_lock *lock)
 837 {
 838         LIST_HEAD(ast_list);
 839         ENTRY;
 840
 841         l_lock(&ns->ns_lock);
 842         LDLM_DEBUG(lock, "client completion callback handler START");
 843
 844         /* If we receive the completion AST before the actual enqueue returned,
 845          * then we might need to switch lock modes, resources, or extents. */
 846         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 847                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 848                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 849         }
 850
 851         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 852                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 853                        sizeof(lock->l_policy_data));
 854                 LDLM_DEBUG(lock, "completion AST, new policy data");
 855         }
 856
 857         ldlm_resource_unlink_lock(lock);
 858         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 859                    &lock->l_resource->lr_name,
 860                    sizeof(lock->l_resource->lr_name)) != 0) {
 861                 ldlm_lock_change_resource(ns, lock,
 862                                          dlm_req->lock_desc.l_resource.lr_name);
 863                 LDLM_DEBUG(lock, "completion AST, new resource");
 864         }
 865
 866         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 867                 lock->l_flags |= LDLM_FL_CBPENDING;
 868                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 869         }
 870
 871         if (lock->l_lvb_len) {
 872                 void *lvb;
 873                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 874                                          lock->l_lvb_swabber);
 875                 if (lvb == NULL) {
 876                         LDLM_ERROR(lock, "completion AST did not contain "
 877                                    "expected LVB!");
 878                 } else {
 879                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 880                 }
 881         }
 882
 883         lock->l_resource->lr_tmp = &ast_list;
 884         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 885         lock->l_resource->lr_tmp = NULL;
 886         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 887         l_unlock(&ns->ns_lock);
 888         LDLM_LOCK_PUT(lock);
 889
 890         ldlm_run_ast_work(ns, &ast_list);
 891
 892         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 893                           lock);
 894         EXIT;
 895 }
 896
 897 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 898                                     struct ldlm_namespace *ns,
 899                                     struct ldlm_request *dlm_req,
 900                                     struct ldlm_lock *lock)
 901 {
 902         int rc = -ENOSYS;
 903         ENTRY;
 904
 905         l_lock(&ns->ns_lock);
 906         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 907
 908         if (lock->l_glimpse_ast != NULL) {
 909                 l_unlock(&ns->ns_lock);
 910                 l_check_no_ns_lock(ns);
 911                 rc = lock->l_glimpse_ast(lock, req);
 912                 l_lock(&ns->ns_lock);
 913         }
 914
 915         if (req->rq_repmsg != NULL) {
 916                 ptlrpc_reply(req);
 917         } else {
 918                 req->rq_status = rc;
 919                 ptlrpc_error(req);
 920         }
 921
 922         if (lock->l_granted_mode == LCK_PW &&
 923             !lock->l_readers && !lock->l_writers &&
 924             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 925                 l_unlock(&ns->ns_lock);
 926                 ldlm_handle_bl_callback(ns, NULL, lock);
 927                 EXIT;
 928                 return;
 929         }
 930
 931         l_unlock(&ns->ns_lock);
 932         LDLM_LOCK_PUT(lock);
 933         EXIT;
 934 }
 935
 936 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 937 {
 938         req->rq_status = rc;
 939         if (req->rq_reply_state == NULL) {
 940                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 941                 if (rc)
 942                         return rc;
 943         }
 944         return ptlrpc_reply(req);
 945 }
 946
 947 #ifdef __KERNEL__
 948 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
 949                       struct ldlm_lock *lock)
 950 {
 951         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
 952         struct ldlm_bl_work_item *blwi;
 953         ENTRY;
 954
 955         OBD_ALLOC(blwi, sizeof(*blwi));
 956         if (blwi == NULL)
 957                 RETURN(-ENOMEM);
 958
 959         blwi->blwi_ns = ns;
 960         if (ld != NULL)
 961                 blwi->blwi_ld = *ld;
 962         blwi->blwi_lock = lock;
 963
 964         spin_lock(&blp->blp_lock);
 965         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
 966         wake_up(&blp->blp_waitq);
 967         spin_unlock(&blp->blp_lock);
 968
 969         RETURN(0);
 970 }
 971 #else
 972 /* XXX */
 973 void liblustre_ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 974                                        struct ldlm_lock_desc *ld,
 975                                        struct ldlm_lock *lock)
 976 {
 977         ldlm_handle_bl_callback(ns, ld, lock);
 978 }
 979 #endif
 980
 981 static int ldlm_callback_handler(struct ptlrpc_request *req)
 982 {
 983         struct ldlm_namespace *ns;
 984         struct ldlm_request *dlm_req;
 985         struct ldlm_lock *lock;
 986         char str[PTL_NALFMT_SIZE];
 987         int rc;
 988         ENTRY;
 989
 990         /* Requests arrive in sender's byte order.  The ptlrpc service
 991          * handler has already checked and, if necessary, byte-swapped the
 992          * incoming request message body, but I am responsible for the
 993          * message buffers. */
 994
 995         if (req->rq_export == NULL) {
 996                 struct ldlm_request *dlm_req;
 997
 998                 CDEBUG(D_RPCTRACE, "operation %d from nid "LPX64" (%s) with bad "
 999                        "export cookie "LPX64" (ptl req %d/rep %d); this is "
1000                        "normal if this node rebooted with a lock held\n",
1001                        req->rq_reqmsg->opc, req->rq_peer.peer_nid,
1002                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
1003                                        req->rq_peer.peer_nid, str),
1004                        req->rq_reqmsg->handle.cookie,
1005                        req->rq_request_portal, req->rq_reply_portal);
1006
1007                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1008                                              lustre_swab_ldlm_request);
1009                 if (dlm_req != NULL)
1010                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1011                                dlm_req->lock_handle1.cookie);
1012
1013                 ldlm_callback_reply(req, -ENOTCONN);
1014                 RETURN(0);
1015         }
1016
1017         LASSERT(req->rq_export != NULL);
1018         LASSERT(req->rq_export->exp_obd != NULL);
1019
1020         switch(req->rq_reqmsg->opc) {
1021         case LDLM_BL_CALLBACK:
1022                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1023                 break;
1024         case LDLM_CP_CALLBACK:
1025                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1026                 break;
1027         case LDLM_GL_CALLBACK:
1028                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1029                 break;
1030         case OBD_LOG_CANCEL:
1031                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1032                 rc = llog_origin_handle_cancel(req);
1033                 ldlm_callback_reply(req, rc);
1034                 RETURN(0);
1035         case LLOG_ORIGIN_HANDLE_CREATE:
1036                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1037                 rc = llog_origin_handle_create(req);
1038                 ldlm_callback_reply(req, rc);
1039                 RETURN(0);
1040         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1041                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1042                 rc = llog_origin_handle_next_block(req);
1043                 ldlm_callback_reply(req, rc);
1044                 RETURN(0);
1045         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1046                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1047                 rc = llog_origin_handle_read_header(req);
1048                 ldlm_callback_reply(req, rc);
1049                 RETURN(0);
1050         case LLOG_ORIGIN_HANDLE_CLOSE:
1051                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1052                 rc = llog_origin_handle_close(req);
1053                 ldlm_callback_reply(req, rc);
1054                 RETURN(0);
1055         default:
1056                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1057                 ldlm_callback_reply(req, -EPROTO);
1058                 RETURN(0);
1059         }
1060
1061         ns = req->rq_export->exp_obd->obd_namespace;
1062         LASSERT(ns != NULL);
1063
1064         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1065                                       lustre_swab_ldlm_request);
1066         if (dlm_req == NULL) {
1067                 CERROR ("can't unpack dlm_req\n");
1068                 ldlm_callback_reply (req, -EPROTO);
1069                 RETURN (0);
1070         }
1071
1072         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1073         if (!lock) {
1074                 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
1075                        dlm_req->lock_handle1.cookie);
1076                 ldlm_callback_reply(req, -EINVAL);
1077                 RETURN(0);
1078         }
1079
1080         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1081         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1082
1083         /* We want the ost thread to get this reply so that it can respond
1084          * to ost requests (write cache writeback) that might be triggered
1085          * in the callback.
1086          *
1087          * But we'd also like to be able to indicate in the reply that we're
1088          * cancelling right now, because it's unused, or have an intent result
1089          * in the reply, so we might have to push the responsibility for sending
1090          * the reply down into the AST handlers, alas. */
1091
1092         switch (req->rq_reqmsg->opc) {
1093         case LDLM_BL_CALLBACK:
1094                 CDEBUG(D_INODE, "blocking ast\n");
1095 #ifdef __KERNEL__
1096                 rc = ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock);
1097                 ldlm_callback_reply(req, rc);
1098 #else
1099                 rc = 0;
1100                 ldlm_callback_reply(req, rc);
1101                 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1102 #endif
1103                 break;
1104         case LDLM_CP_CALLBACK:
1105                 CDEBUG(D_INODE, "completion ast\n");
1106                 ldlm_callback_reply(req, 0);
1107                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1108                 break;
1109         case LDLM_GL_CALLBACK:
1110                 CDEBUG(D_INODE, "glimpse ast\n");
1111                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1112                 break;
1113         default:
1114                 LBUG();                         /* checked above */
1115         }
1116
1117         RETURN(0);
1118 }
1119
1120 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1121 {
1122         int rc;
1123         ENTRY;
1124
1125         /* Requests arrive in sender's byte order.  The ptlrpc service
1126          * handler has already checked and, if necessary, byte-swapped the
1127          * incoming request message body, but I am responsible for the
1128          * message buffers. */
1129
1130         if (req->rq_export == NULL) {
1131                 struct ldlm_request *dlm_req;
1132                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
1133                        req->rq_reqmsg->opc, req->rq_request_portal,
1134                        req->rq_reply_portal);
1135                 CERROR("--> export cookie: "LPX64"\n",
1136                        req->rq_reqmsg->handle.cookie);
1137                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1138                                              lustre_swab_ldlm_request);
1139                 if (dlm_req != NULL)
1140                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1141                 RETURN(-ENOTCONN);
1142         }
1143
1144         switch (req->rq_reqmsg->opc) {
1145
1146         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1147         case LDLM_CANCEL:
1148                 CDEBUG(D_INODE, "cancel\n");
1149                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1150                 rc = ldlm_handle_cancel(req);
1151                 if (rc)
1152                         break;
1153                 RETURN(0);
1154
1155         default:
1156                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1157                 RETURN(-EINVAL);
1158         }
1159
1160         RETURN(0);
1161 }
1162
1163 #ifdef __KERNEL__
1164 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1165 {
1166         struct ldlm_bl_work_item *blwi = NULL;
1167
1168         spin_lock(&blp->blp_lock);
1169         if (!list_empty(&blp->blp_list)) {
1170                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1171                                   blwi_entry);
1172                 list_del(&blwi->blwi_entry);
1173         }
1174         spin_unlock(&blp->blp_lock);
1175
1176         return blwi;
1177 }
1178
1179 struct ldlm_bl_thread_data {
1180         int                     bltd_num;
1181         struct ldlm_bl_pool     *bltd_blp;
1182 };
1183
1184 static int ldlm_bl_thread_main(void *arg)
1185 {
1186         struct ldlm_bl_thread_data *bltd = arg;
1187         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1188         unsigned long flags;
1189         ENTRY;
1190
1191         /* XXX boiler-plate */
1192         {
1193                 char name[sizeof(current->comm)];
1194                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1195                          bltd->bltd_num);
1196                 kportal_daemonize(name);
1197         }
1198         SIGNAL_MASK_LOCK(current, flags);
1199         sigfillset(&current->blocked);
1200         RECALC_SIGPENDING;
1201         SIGNAL_MASK_UNLOCK(current, flags);
1202
1203         atomic_inc(&blp->blp_num_threads);
1204         complete(&blp->blp_comp);
1205
1206         while(1) {
1207                 struct l_wait_info lwi = { 0 };
1208                 struct ldlm_bl_work_item *blwi = NULL;
1209
1210                 l_wait_event_exclusive(blp->blp_waitq,
1211                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1212                                        &lwi);
1213
1214                 if (blwi->blwi_ns == NULL)
1215                         break;
1216
1217                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1218                                         blwi->blwi_lock);
1219                 OBD_FREE(blwi, sizeof(*blwi));
1220         }
1221
1222         atomic_dec(&blp->blp_num_threads);
1223         complete(&blp->blp_comp);
1224         RETURN(0);
1225 }
1226
1227 #endif
1228
1229 static int ldlm_setup(void);
1230 static int ldlm_cleanup(int force);
1231
1232 int ldlm_get_ref(void)
1233 {
1234         int rc = 0;
1235         down(&ldlm_ref_sem);
1236         if (++ldlm_refcount == 1) {
1237                 rc = ldlm_setup();
1238                 if (rc)
1239                         ldlm_refcount--;
1240         }
1241         up(&ldlm_ref_sem);
1242
1243         RETURN(rc);
1244 }
1245
1246 void ldlm_put_ref(int force)
1247 {
1248         down(&ldlm_ref_sem);
1249         if (ldlm_refcount == 1) {
1250                 int rc = ldlm_cleanup(force);
1251                 if (rc)
1252                         CERROR("ldlm_cleanup failed: %d\n", rc);
1253                 else
1254                         ldlm_refcount--;
1255         } else {
1256                 ldlm_refcount--;
1257         }
1258         up(&ldlm_ref_sem);
1259
1260         EXIT;
1261 }
1262
1263 static int ldlm_setup(void)
1264 {
1265         struct ldlm_bl_pool *blp;
1266         int rc = 0;
1267 #ifdef __KERNEL__
1268         int i;
1269 #endif
1270         ENTRY;
1271
1272         if (ldlm_state != NULL)
1273                 RETURN(-EALREADY);
1274
1275         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1276         if (ldlm_state == NULL)
1277                 RETURN(-ENOMEM);
1278
1279 #ifdef __KERNEL__
1280         rc = ldlm_proc_setup();
1281         if (rc != 0)
1282                 GOTO(out_free, rc);
1283 #endif
1284
1285         ldlm_state->ldlm_cb_service =
1286                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1287                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1288                                 ldlm_callback_handler, "ldlm_cbd",
1289                                 ldlm_svc_proc_dir);
1290
1291         if (!ldlm_state->ldlm_cb_service) {
1292                 CERROR("failed to start service\n");
1293                 GOTO(out_proc, rc = -ENOMEM);
1294         }
1295
1296         ldlm_state->ldlm_cancel_service =
1297                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1298                                 LDLM_CANCEL_REQUEST_PORTAL,
1299                                 LDLM_CANCEL_REPLY_PORTAL,
1300                                 ldlm_cancel_handler, "ldlm_canceld",
1301                                 ldlm_svc_proc_dir);
1302
1303         if (!ldlm_state->ldlm_cancel_service) {
1304                 CERROR("failed to start service\n");
1305                 GOTO(out_proc, rc = -ENOMEM);
1306         }
1307
1308         OBD_ALLOC(blp, sizeof(*blp));
1309         if (blp == NULL)
1310                 GOTO(out_proc, rc = -ENOMEM);
1311         ldlm_state->ldlm_bl_pool = blp;
1312
1313         atomic_set(&blp->blp_num_threads, 0);
1314         init_waitqueue_head(&blp->blp_waitq);
1315         spin_lock_init(&blp->blp_lock);
1316
1317         INIT_LIST_HEAD(&blp->blp_list);
1318
1319 #ifdef __KERNEL__
1320         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1321                 struct ldlm_bl_thread_data bltd = {
1322                         .bltd_num = i,
1323                         .bltd_blp = blp,
1324                 };
1325                 init_completion(&blp->blp_comp);
1326                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1327                 if (rc < 0) {
1328                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1329                         LBUG();
1330                         GOTO(out_thread, rc);
1331                 }
1332                 wait_for_completion(&blp->blp_comp);
1333         }
1334
1335         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1336                                     LDLM_NUM_THREADS, "ldlm_cn");
1337         if (rc) {
1338                 LBUG();
1339                 GOTO(out_thread, rc);
1340         }
1341
1342         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1343                                     LDLM_NUM_THREADS, "ldlm_cb");
1344         if (rc) {
1345                 LBUG();
1346                 GOTO(out_thread, rc);
1347         }
1348
1349         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1350         spin_lock_init(&expired_lock_thread.elt_lock);
1351         expired_lock_thread.elt_state = ELT_STOPPED;
1352         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1353
1354         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1355         if (rc < 0) {
1356                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1357                 GOTO(out_thread, rc);
1358         }
1359
1360         wait_event(expired_lock_thread.elt_waitq,
1361                    expired_lock_thread.elt_state == ELT_READY);
1362
1363         INIT_LIST_HEAD(&waiting_locks_list);
1364         spin_lock_init(&waiting_locks_spinlock);
1365         waiting_locks_timer.function = waiting_locks_callback;
1366         waiting_locks_timer.data = 0;
1367         init_timer(&waiting_locks_timer);
1368 #endif
1369
1370         RETURN(0);
1371
1372 #ifdef __KERNEL__
1373  out_thread:
1374         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1375         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1376 #endif
1377
1378  out_proc:
1379 #ifdef __KERNEL__
1380         ldlm_proc_cleanup();
1381  out_free:
1382 #endif
1383         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1384         ldlm_state = NULL;
1385         return rc;
1386 }
1387
1388 static int ldlm_cleanup(int force)
1389 {
1390 #ifdef __KERNEL__
1391         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1392 #endif
1393         ENTRY;
1394
1395         if (!list_empty(&ldlm_namespace_list)) {
1396                 CERROR("ldlm still has namespaces; clean these up first.\n");
1397                 ldlm_dump_all_namespaces();
1398                 RETURN(-EBUSY);
1399         }
1400
1401 #ifdef __KERNEL__
1402         while (atomic_read(&blp->blp_num_threads) > 0) {
1403                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1404
1405                 init_completion(&blp->blp_comp);
1406
1407                 spin_lock(&blp->blp_lock);
1408                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1409                 wake_up(&blp->blp_waitq);
1410                 spin_unlock(&blp->blp_lock);
1411
1412                 wait_for_completion(&blp->blp_comp);
1413         }
1414         OBD_FREE(blp, sizeof(*blp));
1415
1416         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1417         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1418         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1419         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1420         ldlm_proc_cleanup();
1421
1422         expired_lock_thread.elt_state = ELT_TERMINATE;
1423         wake_up(&expired_lock_thread.elt_waitq);
1424         wait_event(expired_lock_thread.elt_waitq,
1425                    expired_lock_thread.elt_state == ELT_STOPPED);
1426
1427 #endif
1428
1429         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1430         ldlm_state = NULL;
1431
1432         RETURN(0);
1433 }
1434
1435 int __init ldlm_init(void)
1436 {
1437         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1438                                                sizeof(struct ldlm_resource), 0,
1439                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1440         if (ldlm_resource_slab == NULL)
1441                 return -ENOMEM;
1442
1443         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1444                                            sizeof(struct ldlm_lock), 0,
1445                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1446         if (ldlm_lock_slab == NULL) {
1447                 kmem_cache_destroy(ldlm_resource_slab);
1448                 return -ENOMEM;
1449         }
1450
1451         l_lock_init(&ldlm_handle_lock);
1452
1453         return 0;
1454 }
1455
1456 void __exit ldlm_exit(void)
1457 {
1458         if ( ldlm_refcount )
1459                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1460         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
1461                 CERROR("couldn't free ldlm resource slab\n");
1462         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
1463                 CERROR("couldn't free ldlm lock slab\n");
1464 }
1465
1466 /* ldlm_flock.c */
1467 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1468
1469 /* ldlm_extent.c */
1470 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1471
1472 /* ldlm_lock.c */
1473 EXPORT_SYMBOL(ldlm_get_processing_policy);
1474 EXPORT_SYMBOL(ldlm_lock2desc);
1475 EXPORT_SYMBOL(ldlm_register_intent);
1476 EXPORT_SYMBOL(ldlm_lockname);
1477 EXPORT_SYMBOL(ldlm_typename);
1478 EXPORT_SYMBOL(ldlm_lock2handle);
1479 EXPORT_SYMBOL(__ldlm_handle2lock);
1480 EXPORT_SYMBOL(ldlm_lock_get);
1481 EXPORT_SYMBOL(ldlm_lock_put);
1482 EXPORT_SYMBOL(ldlm_lock_match);
1483 EXPORT_SYMBOL(ldlm_lock_cancel);
1484 EXPORT_SYMBOL(ldlm_lock_addref);
1485 EXPORT_SYMBOL(ldlm_lock_decref);
1486 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1487 EXPORT_SYMBOL(ldlm_lock_change_resource);
1488 EXPORT_SYMBOL(ldlm_lock_set_data);
1489 EXPORT_SYMBOL(ldlm_it2str);
1490 EXPORT_SYMBOL(ldlm_lock_dump);
1491 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1492 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1493 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1494 EXPORT_SYMBOL(ldlm_lock_allow_match);
1495
1496 /* ldlm_request.c */
1497 EXPORT_SYMBOL(ldlm_completion_ast);
1498 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1499 EXPORT_SYMBOL(ldlm_cli_convert);
1500 EXPORT_SYMBOL(ldlm_cli_enqueue);
1501 EXPORT_SYMBOL(ldlm_cli_cancel);
1502 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1503 EXPORT_SYMBOL(ldlm_replay_locks);
1504 EXPORT_SYMBOL(ldlm_resource_foreach);
1505 EXPORT_SYMBOL(ldlm_namespace_foreach);
1506 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1507 EXPORT_SYMBOL(ldlm_change_cbdata);
1508
1509 /* ldlm_lockd.c */
1510 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1511 EXPORT_SYMBOL(ldlm_server_completion_ast);
1512 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1513 EXPORT_SYMBOL(ldlm_handle_enqueue);
1514 EXPORT_SYMBOL(ldlm_handle_cancel);
1515 EXPORT_SYMBOL(ldlm_handle_convert);
1516 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1517 EXPORT_SYMBOL(ldlm_get_ref);
1518 EXPORT_SYMBOL(ldlm_put_ref);
1519
1520 #if 0
1521 /* ldlm_test.c */
1522 EXPORT_SYMBOL(ldlm_test);
1523 EXPORT_SYMBOL(ldlm_regression_start);
1524 EXPORT_SYMBOL(ldlm_regression_stop);
1525 #endif
1526
1527 /* ldlm_resource.c */
1528 EXPORT_SYMBOL(ldlm_namespace_new);
1529 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1530 EXPORT_SYMBOL(ldlm_namespace_free);
1531 EXPORT_SYMBOL(ldlm_namespace_dump);
1532 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1533 EXPORT_SYMBOL(ldlm_resource_get);
1534 EXPORT_SYMBOL(ldlm_resource_putref);
1535
1536 /* l_lock.c */
1537 EXPORT_SYMBOL(l_lock);
1538 EXPORT_SYMBOL(l_unlock);
1539
1540 /* ldlm_lib.c */
1541 EXPORT_SYMBOL(client_obd_setup);
1542 EXPORT_SYMBOL(client_obd_cleanup);
1543 EXPORT_SYMBOL(client_connect_import);
1544 EXPORT_SYMBOL(client_disconnect_export);
1545 EXPORT_SYMBOL(target_abort_recovery);
1546 EXPORT_SYMBOL(target_handle_connect);
1547 EXPORT_SYMBOL(target_destroy_export);
1548 EXPORT_SYMBOL(target_cancel_recovery_timer);
1549 EXPORT_SYMBOL(target_send_reply);
1550 EXPORT_SYMBOL(target_queue_recovery_request);
1551 EXPORT_SYMBOL(target_handle_ping);
1552 EXPORT_SYMBOL(target_handle_disconnect);
1553 EXPORT_SYMBOL(target_queue_final_reply);