lustre/ldlm/ldlm_lockd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #ifndef EXPORT_SYMTAB
  25 # define EXPORT_SYMTAB
  26 #endif
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28
  29 #ifdef __KERNEL__
  30 # include <linux/module.h>
  31 # include <linux/slab.h>
  32 # include <linux/init.h>
  33 # include <linux/wait.h>
  34 #else
  35 # include <liblustre.h>
  36 #endif
  37
  38 #include <linux/lustre_dlm.h>
  39 #include <linux/obd_class.h>
  40 #include <portals/list.h>
  41 #include "ldlm_internal.h"
  42
  43 extern kmem_cache_t *ldlm_resource_slab;
  44 extern kmem_cache_t *ldlm_lock_slab;
  45 extern struct lustre_lock ldlm_handle_lock;
  46 extern struct list_head ldlm_namespace_list;
  47 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
  48 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
  49
  50 static DECLARE_MUTEX(ldlm_ref_sem);
  51 static int ldlm_refcount = 0;
  52
  53 /* LDLM state */
  54
  55 static struct ldlm_state *ldlm_state;
  56
  57 inline unsigned long round_timeout(unsigned long timeout)
  58 {
  59         return ((timeout / HZ) + 1) * HZ;
  60 }
  61
  62 #ifdef __KERNEL__
  63 /* XXX should this be per-ldlm? */
  64 static struct list_head waiting_locks_list;
  65 static spinlock_t waiting_locks_spinlock;
  66 static struct timer_list waiting_locks_timer;
  67
  68 static struct expired_lock_thread {
  69         wait_queue_head_t         elt_waitq;
  70         int                       elt_state;
  71         struct list_head          elt_expired_locks;
  72         spinlock_t                elt_lock;
  73 } expired_lock_thread;
  74 #endif
  75
  76 #define ELT_STOPPED   0
  77 #define ELT_READY     1
  78 #define ELT_TERMINATE 2
  79
  80 struct ldlm_bl_pool {
  81         spinlock_t              blp_lock;
  82         struct list_head        blp_list;
  83         wait_queue_head_t       blp_waitq;
  84         atomic_t                blp_num_threads;
  85         struct completion       blp_comp;
  86 };
  87
  88 struct ldlm_bl_work_item {
  89         struct list_head        blwi_entry;
  90         struct ldlm_namespace   *blwi_ns;
  91         struct ldlm_lock_desc   blwi_ld;
  92         struct ldlm_lock        *blwi_lock;
  93 };
  94
  95 #ifdef __KERNEL__
  96
  97 static inline int have_expired_locks(void)
  98 {
  99         int need_to_run;
 100
 101         spin_lock_bh(&expired_lock_thread.elt_lock);
 102         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 103         spin_unlock_bh(&expired_lock_thread.elt_lock);
 104
 105         RETURN(need_to_run);
 106 }
 107
 108 static int expired_lock_main(void *arg)
 109 {
 110         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 111         struct l_wait_info lwi = { 0 };
 112         unsigned long flags;
 113
 114         ENTRY;
 115         lock_kernel();
 116         kportal_daemonize("ldlm_elt");
 117
 118         SIGNAL_MASK_LOCK(current, flags);
 119         sigfillset(&current->blocked);
 120         RECALC_SIGPENDING;
 121         SIGNAL_MASK_UNLOCK(current, flags);
 122
 123         unlock_kernel();
 124
 125         expired_lock_thread.elt_state = ELT_READY;
 126         wake_up(&expired_lock_thread.elt_waitq);
 127
 128         while (1) {
 129                 l_wait_event(expired_lock_thread.elt_waitq,
 130                              have_expired_locks() ||
 131                              expired_lock_thread.elt_state == ELT_TERMINATE,
 132                              &lwi);
 133
 134                 spin_lock_bh(&expired_lock_thread.elt_lock);
 135                 while (!list_empty(expired)) {
 136                         struct obd_export *export;
 137                         struct ldlm_lock *lock;
 138
 139                         lock = list_entry(expired->next, struct ldlm_lock,
 140                                           l_pending_chain);
 141                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
 142                             (void *)lock >= LP_POISON) {
 143                                 CERROR("free lock on elt list %p\n", lock);
 144                                 LBUG();
 145                         }
 146                         list_del_init(&lock->l_pending_chain);
 147                         if ((void *)lock->l_export < LP_POISON + PAGE_SIZE &&
 148                             (void *)lock->l_export >= LP_POISON + PAGE_SIZE) {
 149                                 CERROR("lock with free export on elt list %p\n",
 150                                        export);
 151                                 lock->l_export = NULL;
 152                                 LDLM_ERROR(lock, "free export\n");
 153                                 continue;
 154                         }
 155                         export = class_export_get(lock->l_export);
 156                         spin_unlock_bh(&expired_lock_thread.elt_lock);
 157
 158                         ptlrpc_fail_export(export);
 159                         class_export_put(export);
 160                         spin_lock_bh(&expired_lock_thread.elt_lock);
 161                 }
 162                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 163
 164                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 165                         break;
 166         }
 167
 168         expired_lock_thread.elt_state = ELT_STOPPED;
 169         wake_up(&expired_lock_thread.elt_waitq);
 170         RETURN(0);
 171 }
 172
 173 static void waiting_locks_callback(unsigned long unused)
 174 {
 175         struct ldlm_lock *lock;
 176         char str[PTL_NALFMT_SIZE];
 177
 178         spin_lock_bh(&waiting_locks_spinlock);
 179         while (!list_empty(&waiting_locks_list)) {
 180                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 181                                   l_pending_chain);
 182
 183                 if ((lock->l_callback_timeout > jiffies) ||
 184                     (lock->l_req_mode == LCK_GROUP))
 185                         break;
 186
 187                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 188                            "%s@%s nid %s ",
 189                            lock->l_export->exp_client_uuid.uuid,
 190                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 191                            ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str));
 192
 193                 spin_lock_bh(&expired_lock_thread.elt_lock);
 194                 list_del(&lock->l_pending_chain);
 195                 list_add(&lock->l_pending_chain,
 196                          &expired_lock_thread.elt_expired_locks);
 197                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 198                 wake_up(&expired_lock_thread.elt_waitq);
 199         }
 200
 201         /*
 202          * Make sure the timer will fire again if we have any locks
 203          * left.
 204          */
 205         if (!list_empty(&waiting_locks_list)) {
 206                 unsigned long timeout_rounded;
 207                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 208                                   l_pending_chain);
 209                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 210                 mod_timer(&waiting_locks_timer, timeout_rounded);
 211         }
 212         spin_unlock_bh(&waiting_locks_spinlock);
 213 }
 214
 215 /*
 216  * Indicate that we're waiting for a client to call us back cancelling a given
 217  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 218  * timer to fire appropriately.  (We round up to the next second, to avoid
 219  * floods of timer firings during periods of high lock contention and traffic).
 220  */
 221 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 222 {
 223         unsigned long timeout_rounded;
 224
 225         spin_lock_bh(&waiting_locks_spinlock);
 226         if (!list_empty(&lock->l_pending_chain)) {
 227                 LDLM_DEBUG(lock, "not re-adding to wait list");
 228                 spin_unlock_bh(&waiting_locks_spinlock);
 229                 return 0;
 230         }
 231         LDLM_DEBUG(lock, "adding to wait list");
 232
 233         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 234
 235         timeout_rounded = round_timeout(lock->l_callback_timeout);
 236
 237         if (timeout_rounded < waiting_locks_timer.expires ||
 238             !timer_pending(&waiting_locks_timer)) {
 239                 mod_timer(&waiting_locks_timer, timeout_rounded);
 240         }
 241         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 242         spin_unlock_bh(&waiting_locks_spinlock);
 243         return 1;
 244 }
 245
 246 /*
 247  * Remove a lock from the pending list, likely because it had its cancellation
 248  * callback arrive without incident.  This adjusts the lock-timeout timer if
 249  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 250  */
 251 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 252 {
 253         struct list_head *list_next;
 254
 255         if (lock->l_export == NULL) {
 256                 /* We don't have a "waiting locks list" on clients. */
 257                 LDLM_DEBUG(lock, "client lock: no-op");
 258                 return 0;
 259         }
 260
 261         spin_lock_bh(&waiting_locks_spinlock);
 262
 263         if (list_empty(&lock->l_pending_chain)) {
 264                 spin_unlock_bh(&waiting_locks_spinlock);
 265                 LDLM_DEBUG(lock, "wasn't waiting");
 266                 return 0;
 267         }
 268
 269         list_next = lock->l_pending_chain.next;
 270         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 271                 /* Removing the head of the list, adjust timer. */
 272                 if (list_next == &waiting_locks_list) {
 273                         /* No more, just cancel. */
 274                         del_timer(&waiting_locks_timer);
 275                 } else {
 276                         struct ldlm_lock *next;
 277                         next = list_entry(list_next, struct ldlm_lock,
 278                                           l_pending_chain);
 279                         mod_timer(&waiting_locks_timer,
 280                                   round_timeout(next->l_callback_timeout));
 281                 }
 282         }
 283
 284         spin_lock_bh(&expired_lock_thread.elt_lock);
 285         list_del_init(&lock->l_pending_chain);
 286         spin_unlock_bh(&expired_lock_thread.elt_lock);
 287
 288         spin_unlock_bh(&waiting_locks_spinlock);
 289         LDLM_DEBUG(lock, "removed");
 290         return 1;
 291 }
 292
 293 #else /* !__KERNEL__ */
 294
 295 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 296 {
 297         RETURN(1);
 298 }
 299
 300 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 301 {
 302         RETURN(0);
 303 }
 304
 305 #endif /* __KERNEL__ */
 306
 307 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,const char *ast_type)
 308 {
 309         struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 310         char str[PTL_NALFMT_SIZE];
 311
 312         LDLM_ERROR(lock, "%s AST failed (%d): evicting client %s@%s NID "LPX64
 313                    " (%s)", ast_type, rc, lock->l_export->exp_client_uuid.uuid,
 314                    conn->c_remote_uuid.uuid, conn->c_peer.peer_nid,
 315                    ptlrpc_peernid2str(&conn->c_peer, str));
 316         ptlrpc_fail_export(lock->l_export);
 317 }
 318
 319 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 320                                  struct ptlrpc_request *req, int rc,
 321                                  const char *ast_type)
 322 {
 323         struct ptlrpc_peer *peer = &req->rq_import->imp_connection->c_peer;
 324         char str[PTL_NALFMT_SIZE];
 325
 326         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 327                 LASSERT(lock->l_export);
 328                 if (lock->l_export->exp_libclient) {
 329                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 330                                    " timeout, just cancelling lock", ast_type,
 331                                    ptlrpc_peernid2str(peer, str));
 332                         ldlm_lock_cancel(lock);
 333                         rc = -ERESTART;
 334                 } else {
 335                         ldlm_del_waiting_lock(lock);
 336                         ldlm_failed_ast(lock, rc, ast_type);
 337                 }
 338         } else if (rc) {
 339                 if (rc == -EINVAL)
 340                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 341                                    " from %s AST - normal race",
 342                                    ptlrpc_peernid2str(peer, str),
 343                                    req->rq_repmsg->status, ast_type);
 344                 else
 345                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 346                                    "from %s AST", ptlrpc_peernid2str(peer, str),
 347                                    (req->rq_repmsg != NULL) ?
 348                                    req->rq_repmsg->status : 0, ast_type);
 349                 ldlm_lock_cancel(lock);
 350                 /* Server-side AST functions are called from ldlm_reprocess_all,
 351                  * which needs to be told to please restart its reprocessing. */
 352                 rc = -ERESTART;
 353         }
 354
 355         return rc;
 356 }
 357
 358 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 359                              struct ldlm_lock_desc *desc,
 360                              void *data, int flag)
 361 {
 362         struct ldlm_request *body;
 363         struct ptlrpc_request *req;
 364         int rc = 0, size = sizeof(*body);
 365         ENTRY;
 366
 367         if (flag == LDLM_CB_CANCELING) {
 368                 /* Don't need to do anything here. */
 369                 RETURN(0);
 370         }
 371
 372         LASSERT(lock);
 373
 374         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 375         if (lock->l_granted_mode != lock->l_req_mode) {
 376                 /* this blocking AST will be communicated as part of the
 377                  * completion AST instead */
 378                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 379                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");                 RETURN(0);
 380         }
 381
 382         if (lock->l_destroyed) {
 383                 /* What's the point? */
 384                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 385                 RETURN(0);
 386         }
 387
 388 #if 0
 389         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 390                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 391                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 392                 RETURN(-ETIMEDOUT);
 393         }
 394 #endif
 395
 396         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 397                               LDLM_BL_CALLBACK, 1, &size, NULL);
 398         if (req == NULL) {
 399                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 400                 RETURN(-ENOMEM);
 401         }
 402
 403         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 404         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 405                sizeof(body->lock_handle1));
 406         memcpy(&body->lock_desc, desc, sizeof(*desc));
 407         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 408
 409         LDLM_DEBUG(lock, "server preparing blocking AST");
 410         req->rq_replen = lustre_msg_size(0, NULL);
 411
 412         if (lock->l_granted_mode == lock->l_req_mode)
 413                 ldlm_add_waiting_lock(lock);
 414         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 415
 416         req->rq_send_state = LUSTRE_IMP_FULL;
 417         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 418         rc = ptlrpc_queue_wait(req);
 419         if (rc != 0)
 420                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 421
 422         ptlrpc_req_finished(req);
 423
 424         RETURN(rc);
 425 }
 426
 427 /* XXX copied from ptlrpc/service.c */
 428 static long timeval_sub(struct timeval *large, struct timeval *small)
 429 {
 430         return (large->tv_sec - small->tv_sec) * 1000000 +
 431                 (large->tv_usec - small->tv_usec);
 432 }
 433
 434 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 435 {
 436         struct ldlm_request *body;
 437         struct ptlrpc_request *req;
 438         struct timeval granted_time;
 439         long total_enqueue_wait;
 440         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 441         ENTRY;
 442
 443         LASSERT(lock != NULL);
 444
 445         do_gettimeofday(&granted_time);
 446         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 447
 448         if (total_enqueue_wait / 1000000 > obd_timeout)
 449                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 450
 451         if (lock->l_resource->lr_lvb_len) {
 452                 buffers = 2;
 453                 size[1] = lock->l_resource->lr_lvb_len;
 454         }
 455
 456         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 457                               LDLM_CP_CALLBACK, buffers, size, NULL);
 458         if (req == NULL)
 459                 RETURN(-ENOMEM);
 460
 461         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 462         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 463                sizeof(body->lock_handle1));
 464         body->lock_flags = flags;
 465         ldlm_lock2desc(lock, &body->lock_desc);
 466
 467         if (buffers == 2) {
 468                 void *lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 469                                            lock->l_resource->lr_lvb_len);
 470                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 471                        lock->l_resource->lr_lvb_len);
 472         }
 473
 474         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 475                    total_enqueue_wait);
 476         req->rq_replen = lustre_msg_size(0, NULL);
 477
 478         req->rq_send_state = LUSTRE_IMP_FULL;
 479         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 480
 481         /* We only send real blocking ASTs after the lock is granted */
 482         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 483         if (lock->l_flags & LDLM_FL_AST_SENT) {
 484                 body->lock_flags |= LDLM_FL_AST_SENT;
 485                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 486         }
 487         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 488
 489         rc = ptlrpc_queue_wait(req);
 490         if (rc != 0)
 491                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 492
 493         ptlrpc_req_finished(req);
 494
 495         RETURN(rc);
 496 }
 497
 498 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 499 {
 500         struct ldlm_resource *res = lock->l_resource;
 501         struct ldlm_request *body;
 502         struct ptlrpc_request *req;
 503         int rc = 0, size = sizeof(*body);
 504         ENTRY;
 505
 506         LASSERT(lock != NULL);
 507
 508         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 509                               LDLM_GL_CALLBACK, 1, &size, NULL);
 510         if (req == NULL)
 511                 RETURN(-ENOMEM);
 512
 513         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 514         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 515                sizeof(body->lock_handle1));
 516         ldlm_lock2desc(lock, &body->lock_desc);
 517
 518         size = lock->l_resource->lr_lvb_len;
 519         req->rq_replen = lustre_msg_size(1, &size);
 520
 521         req->rq_send_state = LUSTRE_IMP_FULL;
 522         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 523
 524         rc = ptlrpc_queue_wait(req);
 525         if (rc == -ELDLM_NO_LOCK_DATA)
 526                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 527         else if (rc != 0)
 528                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 529         else
 530                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 531                         (res, req->rq_repmsg, 0, 1);
 532         ptlrpc_req_finished(req);
 533         RETURN(rc);
 534 }
 535
 536 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 537                         ldlm_completion_callback completion_callback,
 538                         ldlm_blocking_callback blocking_callback,
 539                         ldlm_glimpse_callback glimpse_callback)
 540 {
 541         struct obd_device *obddev = req->rq_export->exp_obd;
 542         struct ldlm_reply *dlm_rep;
 543         struct ldlm_request *dlm_req;
 544         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 545         __u32 flags;
 546         ldlm_error_t err = ELDLM_OK;
 547         struct ldlm_lock *lock = NULL;
 548         void *cookie = NULL;
 549         ENTRY;
 550
 551         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 552
 553         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 554                                       lustre_swab_ldlm_request);
 555         if (dlm_req == NULL) {
 556                 CERROR ("Can't unpack dlm_req\n");
 557                 GOTO(out, rc = -EFAULT);
 558         }
 559
 560         flags = dlm_req->lock_flags;
 561
 562         /* The lock's callback data might be set in the policy function */
 563         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 564                                 dlm_req->lock_desc.l_resource.lr_name,
 565                                 dlm_req->lock_desc.l_resource.lr_type,
 566                                 dlm_req->lock_desc.l_req_mode,
 567                                 blocking_callback, completion_callback,
 568                                 glimpse_callback, NULL, 0);
 569         if (!lock)
 570                 GOTO(out, rc = -ENOMEM);
 571
 572         do_gettimeofday(&lock->l_enqueued_time);
 573         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 574                sizeof(lock->l_remote_handle));
 575         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 576
 577         LASSERT(req->rq_export);
 578         lock->l_export = class_export_get(req->rq_export);
 579         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 580         list_add(&lock->l_export_chain,
 581                  &lock->l_export->exp_ldlm_data.led_held_locks);
 582         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 583
 584         if (flags & LDLM_FL_HAS_INTENT) {
 585                 /* In this case, the reply buffer is allocated deep in
 586                  * local_lock_enqueue by the policy function. */
 587                 cookie = req;
 588         } else {
 589                 int buffers = 1;
 590                 if (lock->l_resource->lr_lvb_len) {
 591                         size[1] = lock->l_resource->lr_lvb_len;
 592                         buffers = 2;
 593                 }
 594
 595                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 596                         GOTO(out, rc = -ENOMEM);
 597
 598                 rc = lustre_pack_reply(req, buffers, size, NULL);
 599                 if (rc)
 600                         GOTO(out, rc);
 601         }
 602
 603         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 604                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 605                        sizeof(ldlm_policy_data_t));
 606         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 607                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 608                        sizeof(lock->l_req_extent));
 609
 610         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 611         if (err)
 612                 GOTO(out, err);
 613
 614         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 615         dlm_rep->lock_flags = flags;
 616
 617         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 618         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 619
 620         /* We never send a blocking AST until the lock is granted, but
 621          * we can tell it right now */
 622         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 623         if (lock->l_flags & LDLM_FL_AST_SENT) {
 624                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 625                 if (lock->l_granted_mode == lock->l_req_mode)
 626                         ldlm_add_waiting_lock(lock);
 627         }
 628         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 629
 630         EXIT;
 631  out:
 632         req->rq_status = err;
 633         if (req->rq_reply_state == NULL) {
 634                 err = lustre_pack_reply(req, 0, NULL, NULL);
 635                 if (rc == 0)
 636                         rc = err;
 637                 req->rq_status = rc;
 638         }
 639
 640         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 641          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 642         if (lock) {
 643                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 644                            "(err=%d, rc=%d)", err, rc);
 645
 646                 if (lock->l_resource->lr_lvb_len > 0 && rc == 0) {
 647                         void *lvb = lustre_msg_buf(req->rq_repmsg, 1,
 648                                                   lock->l_resource->lr_lvb_len);
 649                         LASSERT(lvb != NULL);
 650                         memcpy(lvb, lock->l_resource->lr_lvb_data,
 651                                lock->l_resource->lr_lvb_len);
 652                 }
 653
 654                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 655                         ldlm_reprocess_all(lock->l_resource);
 656                 LDLM_LOCK_PUT(lock);
 657         }
 658         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 659                           lock, rc);
 660
 661         return rc;
 662 }
 663
 664 int ldlm_handle_convert(struct ptlrpc_request *req)
 665 {
 666         struct ldlm_request *dlm_req;
 667         struct ldlm_reply *dlm_rep;
 668         struct ldlm_lock *lock;
 669         int rc, size = sizeof(*dlm_rep);
 670         ENTRY;
 671
 672         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 673                                       lustre_swab_ldlm_request);
 674         if (dlm_req == NULL) {
 675                 CERROR ("Can't unpack dlm_req\n");
 676                 RETURN (-EFAULT);
 677         }
 678
 679         rc = lustre_pack_reply(req, 1, &size, NULL);
 680         if (rc) {
 681                 CERROR("out of memory\n");
 682                 RETURN(-ENOMEM);
 683         }
 684         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 685         dlm_rep->lock_flags = dlm_req->lock_flags;
 686
 687         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 688         if (!lock) {
 689                 req->rq_status = EINVAL;
 690         } else {
 691                 LDLM_DEBUG(lock, "server-side convert handler START");
 692                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 693                                   &dlm_rep->lock_flags);
 694                 if (ldlm_del_waiting_lock(lock))
 695                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 696                 req->rq_status = 0;
 697         }
 698
 699         if (lock) {
 700                 ldlm_reprocess_all(lock->l_resource);
 701                 LDLM_DEBUG(lock, "server-side convert handler END");
 702                 LDLM_LOCK_PUT(lock);
 703         } else
 704                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 705
 706         RETURN(0);
 707 }
 708
 709 int ldlm_handle_cancel(struct ptlrpc_request *req)
 710 {
 711         struct ldlm_request *dlm_req;
 712         struct ldlm_lock *lock;
 713         struct ldlm_resource *res;
 714         char str[PTL_NALFMT_SIZE];
 715         int rc;
 716         ENTRY;
 717
 718         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 719                                       lustre_swab_ldlm_request);
 720         if (dlm_req == NULL) {
 721                 CERROR("bad request buffer for cancel\n");
 722                 RETURN(-EFAULT);
 723         }
 724
 725         rc = lustre_pack_reply(req, 0, NULL, NULL);
 726         if (rc) {
 727                 CERROR("out of memory\n");
 728                 RETURN(-ENOMEM);
 729         }
 730
 731         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 732         if (!lock) {
 733                 CERROR("received cancel for unknown lock cookie "LPX64
 734                        " from client %s nid %s\n",
 735                        dlm_req->lock_handle1.cookie,
 736                        req->rq_export->exp_client_uuid.uuid,
 737                        ptlrpc_peernid2str(&req->rq_peer, str));
 738                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 739                                   "(cookie "LPU64")",
 740                                   dlm_req->lock_handle1.cookie);
 741                 req->rq_status = ESTALE;
 742         } else {
 743                 LDLM_DEBUG(lock, "server-side cancel handler START");
 744                 res = lock->l_resource;
 745                 if (res && res->lr_namespace->ns_lvbo &&
 746                     res->lr_namespace->ns_lvbo->lvbo_update) {
 747                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 748                                 (res, NULL, 0, 0);
 749                                 //(res, req->rq_reqmsg, 1);
 750                 }
 751
 752                 ldlm_lock_cancel(lock);
 753                 if (ldlm_del_waiting_lock(lock))
 754                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 755                 req->rq_status = rc;
 756         }
 757
 758         if (ptlrpc_reply(req) != 0)
 759                 LBUG();
 760
 761         if (lock) {
 762                 ldlm_reprocess_all(lock->l_resource);
 763                 LDLM_DEBUG(lock, "server-side cancel handler END");
 764                 LDLM_LOCK_PUT(lock);
 765         }
 766
 767         RETURN(0);
 768 }
 769
 770 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 771                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 772 {
 773         int do_ast;
 774         ENTRY;
 775
 776         l_lock(&ns->ns_lock);
 777         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 778
 779         lock->l_flags |= LDLM_FL_CBPENDING;
 780         do_ast = (!lock->l_readers && !lock->l_writers);
 781
 782         if (do_ast) {
 783                 LDLM_DEBUG(lock, "already unused, calling "
 784                            "callback (%p)", lock->l_blocking_ast);
 785                 if (lock->l_blocking_ast != NULL) {
 786                         l_unlock(&ns->ns_lock);
 787                         l_check_no_ns_lock(ns);
 788                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 789                                              LDLM_CB_BLOCKING);
 790                         l_lock(&ns->ns_lock);
 791                 }
 792         } else {
 793                 LDLM_DEBUG(lock, "Lock still has references, will be"
 794                            " cancelled later");
 795         }
 796
 797         LDLM_DEBUG(lock, "client blocking callback handler END");
 798         l_unlock(&ns->ns_lock);
 799         LDLM_LOCK_PUT(lock);
 800         EXIT;
 801 }
 802
 803 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 804                                     struct ldlm_namespace *ns,
 805                                     struct ldlm_request *dlm_req,
 806                                     struct ldlm_lock *lock)
 807 {
 808         LIST_HEAD(ast_list);
 809         ENTRY;
 810
 811         l_lock(&ns->ns_lock);
 812         LDLM_DEBUG(lock, "client completion callback handler START");
 813
 814         /* If we receive the completion AST before the actual enqueue returned,
 815          * then we might need to switch lock modes, resources, or extents. */
 816         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 817                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 818                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 819         }
 820
 821         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 822                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 823                        sizeof(lock->l_policy_data));
 824                 LDLM_DEBUG(lock, "completion AST, new policy data");
 825         }
 826
 827         ldlm_resource_unlink_lock(lock);
 828         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 829                    &lock->l_resource->lr_name,
 830                    sizeof(lock->l_resource->lr_name)) != 0) {
 831                 ldlm_lock_change_resource(ns, lock,
 832                                          dlm_req->lock_desc.l_resource.lr_name);
 833                 LDLM_DEBUG(lock, "completion AST, new resource");
 834         }
 835
 836         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 837                 lock->l_flags |= LDLM_FL_CBPENDING;
 838                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 839         }
 840
 841         if (lock->l_lvb_len) {
 842                 void *lvb;
 843                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 844                                          lock->l_lvb_swabber);
 845                 if (lvb == NULL) {
 846                         LDLM_ERROR(lock, "completion AST did not contain "
 847                                    "expected LVB!");
 848                 } else {
 849                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 850                 }
 851         }
 852
 853         lock->l_resource->lr_tmp = &ast_list;
 854         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 855         lock->l_resource->lr_tmp = NULL;
 856         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 857         l_unlock(&ns->ns_lock);
 858         LDLM_LOCK_PUT(lock);
 859
 860         ldlm_run_ast_work(ns, &ast_list);
 861
 862         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 863                           lock);
 864         EXIT;
 865 }
 866
 867 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 868                                     struct ldlm_namespace *ns,
 869                                     struct ldlm_request *dlm_req,
 870                                     struct ldlm_lock *lock)
 871 {
 872         int rc = -ENOSYS;
 873         ENTRY;
 874
 875         l_lock(&ns->ns_lock);
 876         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 877
 878         if (lock->l_glimpse_ast != NULL) {
 879                 l_unlock(&ns->ns_lock);
 880                 l_check_no_ns_lock(ns);
 881                 rc = lock->l_glimpse_ast(lock, req);
 882                 l_lock(&ns->ns_lock);
 883         }
 884
 885         if (req->rq_repmsg != NULL) {
 886                 ptlrpc_reply(req);
 887         } else {
 888                 req->rq_status = rc;
 889                 ptlrpc_error(req);
 890         }
 891
 892         if (lock->l_granted_mode == LCK_PW &&
 893             !lock->l_readers && !lock->l_writers &&
 894             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 895 #ifdef __KERNEL__
 896                 ldlm_bl_to_thread(ns, NULL, lock);
 897                 l_unlock(&ns->ns_lock);
 898 #else
 899                 l_unlock(&ns->ns_lock);
 900                 ldlm_handle_bl_callback(ns, NULL, lock);
 901 #endif
 902                 EXIT;
 903                 return;
 904         }
 905
 906         l_unlock(&ns->ns_lock);
 907         LDLM_LOCK_PUT(lock);
 908         EXIT;
 909 }
 910
 911 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 912 {
 913         req->rq_status = rc;
 914         if (req->rq_reply_state == NULL) {
 915                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 916                 if (rc)
 917                         return rc;
 918         }
 919         return ptlrpc_reply(req);
 920 }
 921
 922 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
 923                       struct ldlm_lock *lock)
 924 {
 925 #ifdef __KERNEL__
 926         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
 927         struct ldlm_bl_work_item *blwi;
 928         ENTRY;
 929
 930         OBD_ALLOC(blwi, sizeof(*blwi));
 931         if (blwi == NULL)
 932                 RETURN(-ENOMEM);
 933
 934         blwi->blwi_ns = ns;
 935         if (ld != NULL)
 936                 blwi->blwi_ld = *ld;
 937         blwi->blwi_lock = lock;
 938
 939         spin_lock(&blp->blp_lock);
 940         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
 941         wake_up(&blp->blp_waitq);
 942         spin_unlock(&blp->blp_lock);
 943 #else
 944         LBUG();
 945 #endif
 946
 947         RETURN(0);
 948 }
 949
 950 static int ldlm_callback_handler(struct ptlrpc_request *req)
 951 {
 952         struct ldlm_namespace *ns;
 953         struct ldlm_request *dlm_req;
 954         struct ldlm_lock *lock;
 955         char str[PTL_NALFMT_SIZE];
 956         int rc;
 957         ENTRY;
 958
 959         /* Requests arrive in sender's byte order.  The ptlrpc service
 960          * handler has already checked and, if necessary, byte-swapped the
 961          * incoming request message body, but I am responsible for the
 962          * message buffers. */
 963
 964         if (req->rq_export == NULL) {
 965                 struct ldlm_request *dlm_req;
 966
 967                 CDEBUG(D_RPCTRACE, "operation %d from nid %s with bad "
 968                        "export cookie "LPX64" (ptl req %d/rep %d); this is "
 969                        "normal if this node rebooted with a lock held\n",
 970                        req->rq_reqmsg->opc,
 971                        ptlrpc_peernid2str(&req->rq_peer, str),
 972                        req->rq_reqmsg->handle.cookie,
 973                        req->rq_request_portal, req->rq_reply_portal);
 974
 975                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 976                                              lustre_swab_ldlm_request);
 977                 if (dlm_req != NULL)
 978                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
 979                                dlm_req->lock_handle1.cookie);
 980
 981                 ldlm_callback_reply(req, -ENOTCONN);
 982                 RETURN(0);
 983         }
 984
 985         LASSERT(req->rq_export != NULL);
 986         LASSERT(req->rq_export->exp_obd != NULL);
 987
 988         switch(req->rq_reqmsg->opc) {
 989         case LDLM_BL_CALLBACK:
 990                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 991                 break;
 992         case LDLM_CP_CALLBACK:
 993                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
 994                 break;
 995         case LDLM_GL_CALLBACK:
 996                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
 997                 break;
 998         case OBD_LOG_CANCEL:
 999                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1000                 rc = llog_origin_handle_cancel(req);
1001                 ldlm_callback_reply(req, rc);
1002                 RETURN(0);
1003         case LLOG_ORIGIN_HANDLE_OPEN:
1004                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1005                 rc = llog_origin_handle_open(req);
1006                 ldlm_callback_reply(req, rc);
1007                 RETURN(0);
1008         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1009                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1010                 rc = llog_origin_handle_next_block(req);
1011                 ldlm_callback_reply(req, rc);
1012                 RETURN(0);
1013         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1014                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1015                 rc = llog_origin_handle_prev_block(req);
1016                 ldlm_callback_reply(req, rc);
1017                 RETURN(0);
1018         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1019                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1020                 rc = llog_origin_handle_read_header(req);
1021                 ldlm_callback_reply(req, rc);
1022                 RETURN(0);
1023         case LLOG_ORIGIN_HANDLE_CLOSE:
1024                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1025                 rc = llog_origin_handle_close(req);
1026                 ldlm_callback_reply(req, rc);
1027                 RETURN(0);
1028         default:
1029                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1030                 ldlm_callback_reply(req, -EPROTO);
1031                 RETURN(0);
1032         }
1033
1034         ns = req->rq_export->exp_obd->obd_namespace;
1035         LASSERT(ns != NULL);
1036
1037         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1038                                       lustre_swab_ldlm_request);
1039         if (dlm_req == NULL) {
1040                 CERROR ("can't unpack dlm_req\n");
1041                 ldlm_callback_reply (req, -EPROTO);
1042                 RETURN (0);
1043         }
1044
1045         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1046         if (!lock) {
1047                 LDLM_DEBUG_NOLOCK("callback on lock "LPX64" - lock "
1048                                   "disappeared\n",dlm_req->lock_handle1.cookie);
1049                 ldlm_callback_reply(req, -EINVAL);
1050                 RETURN(0);
1051         }
1052
1053         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1054         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1055
1056         /* We want the ost thread to get this reply so that it can respond
1057          * to ost requests (write cache writeback) that might be triggered
1058          * in the callback.
1059          *
1060          * But we'd also like to be able to indicate in the reply that we're
1061          * cancelling right now, because it's unused, or have an intent result
1062          * in the reply, so we might have to push the responsibility for sending
1063          * the reply down into the AST handlers, alas. */
1064
1065         switch (req->rq_reqmsg->opc) {
1066         case LDLM_BL_CALLBACK:
1067                 CDEBUG(D_INODE, "blocking ast\n");
1068 #ifdef __KERNEL__
1069                 rc = ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock);
1070                 ldlm_callback_reply(req, rc);
1071 #else
1072                 rc = 0;
1073                 ldlm_callback_reply(req, rc);
1074                 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1075 #endif
1076                 break;
1077         case LDLM_CP_CALLBACK:
1078                 CDEBUG(D_INODE, "completion ast\n");
1079                 ldlm_callback_reply(req, 0);
1080                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1081                 break;
1082         case LDLM_GL_CALLBACK:
1083                 CDEBUG(D_INODE, "glimpse ast\n");
1084                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1085                 break;
1086         default:
1087                 LBUG();                         /* checked above */
1088         }
1089
1090         RETURN(0);
1091 }
1092
1093 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1094 {
1095         int rc;
1096         ENTRY;
1097
1098         /* Requests arrive in sender's byte order.  The ptlrpc service
1099          * handler has already checked and, if necessary, byte-swapped the
1100          * incoming request message body, but I am responsible for the
1101          * message buffers. */
1102
1103         if (req->rq_export == NULL) {
1104                 struct ldlm_request *dlm_req;
1105                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
1106                        req->rq_reqmsg->opc, req->rq_request_portal,
1107                        req->rq_reply_portal);
1108                 CERROR("--> export cookie: "LPX64"\n",
1109                        req->rq_reqmsg->handle.cookie);
1110                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1111                                              lustre_swab_ldlm_request);
1112                 if (dlm_req != NULL)
1113                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1114                 RETURN(-ENOTCONN);
1115         }
1116
1117         switch (req->rq_reqmsg->opc) {
1118
1119         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1120         case LDLM_CANCEL:
1121                 CDEBUG(D_INODE, "cancel\n");
1122                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1123                 rc = ldlm_handle_cancel(req);
1124                 if (rc)
1125                         break;
1126                 RETURN(0);
1127
1128         default:
1129                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1130                 RETURN(-EINVAL);
1131         }
1132
1133         RETURN(0);
1134 }
1135
1136 #ifdef __KERNEL__
1137 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1138 {
1139         struct ldlm_bl_work_item *blwi = NULL;
1140
1141         spin_lock(&blp->blp_lock);
1142         if (!list_empty(&blp->blp_list)) {
1143                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1144                                   blwi_entry);
1145                 list_del(&blwi->blwi_entry);
1146         }
1147         spin_unlock(&blp->blp_lock);
1148
1149         return blwi;
1150 }
1151
1152 struct ldlm_bl_thread_data {
1153         int                     bltd_num;
1154         struct ldlm_bl_pool     *bltd_blp;
1155 };
1156
1157 static int ldlm_bl_thread_main(void *arg)
1158 {
1159         struct ldlm_bl_thread_data *bltd = arg;
1160         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1161         unsigned long flags;
1162         ENTRY;
1163
1164         /* XXX boiler-plate */
1165         {
1166                 char name[sizeof(current->comm)];
1167                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1168                          bltd->bltd_num);
1169                 kportal_daemonize(name);
1170         }
1171         SIGNAL_MASK_LOCK(current, flags);
1172         sigfillset(&current->blocked);
1173         RECALC_SIGPENDING;
1174         SIGNAL_MASK_UNLOCK(current, flags);
1175
1176         atomic_inc(&blp->blp_num_threads);
1177         complete(&blp->blp_comp);
1178
1179         while(1) {
1180                 struct l_wait_info lwi = { 0 };
1181                 struct ldlm_bl_work_item *blwi = NULL;
1182
1183                 l_wait_event_exclusive(blp->blp_waitq,
1184                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1185                                        &lwi);
1186
1187                 if (blwi->blwi_ns == NULL)
1188                         break;
1189
1190                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1191                                         blwi->blwi_lock);
1192                 OBD_FREE(blwi, sizeof(*blwi));
1193         }
1194
1195         atomic_dec(&blp->blp_num_threads);
1196         complete(&blp->blp_comp);
1197         RETURN(0);
1198 }
1199
1200 #endif
1201
1202 static int ldlm_setup(void);
1203 static int ldlm_cleanup(int force);
1204
1205 int ldlm_get_ref(void)
1206 {
1207         int rc = 0;
1208         down(&ldlm_ref_sem);
1209         if (++ldlm_refcount == 1) {
1210                 rc = ldlm_setup();
1211                 if (rc)
1212                         ldlm_refcount--;
1213         }
1214         up(&ldlm_ref_sem);
1215
1216         RETURN(rc);
1217 }
1218
1219 void ldlm_put_ref(int force)
1220 {
1221         down(&ldlm_ref_sem);
1222         if (ldlm_refcount == 1) {
1223                 int rc = ldlm_cleanup(force);
1224                 if (rc)
1225                         CERROR("ldlm_cleanup failed: %d\n", rc);
1226                 else
1227                         ldlm_refcount--;
1228         } else {
1229                 ldlm_refcount--;
1230         }
1231         up(&ldlm_ref_sem);
1232
1233         EXIT;
1234 }
1235
1236 static int ldlm_setup(void)
1237 {
1238         struct ldlm_bl_pool *blp;
1239         int rc = 0;
1240 #ifdef __KERNEL__
1241         int i;
1242 #endif
1243         ENTRY;
1244
1245         if (ldlm_state != NULL)
1246                 RETURN(-EALREADY);
1247
1248         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1249         if (ldlm_state == NULL)
1250                 RETURN(-ENOMEM);
1251
1252 #ifdef __KERNEL__
1253         rc = ldlm_proc_setup();
1254         if (rc != 0)
1255                 GOTO(out_free, rc);
1256 #endif
1257
1258         ldlm_state->ldlm_cb_service =
1259                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1260                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1261                                 ldlm_callback_handler, "ldlm_cbd",
1262                                 ldlm_svc_proc_dir);
1263
1264         if (!ldlm_state->ldlm_cb_service) {
1265                 CERROR("failed to start service\n");
1266                 GOTO(out_proc, rc = -ENOMEM);
1267         }
1268
1269         ldlm_state->ldlm_cancel_service =
1270                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1271                                 LDLM_CANCEL_REQUEST_PORTAL,
1272                                 LDLM_CANCEL_REPLY_PORTAL,
1273                                 ldlm_cancel_handler, "ldlm_canceld",
1274                                 ldlm_svc_proc_dir);
1275
1276         if (!ldlm_state->ldlm_cancel_service) {
1277                 CERROR("failed to start service\n");
1278                 GOTO(out_proc, rc = -ENOMEM);
1279         }
1280
1281         OBD_ALLOC(blp, sizeof(*blp));
1282         if (blp == NULL)
1283                 GOTO(out_proc, rc = -ENOMEM);
1284         ldlm_state->ldlm_bl_pool = blp;
1285
1286         atomic_set(&blp->blp_num_threads, 0);
1287         init_waitqueue_head(&blp->blp_waitq);
1288         spin_lock_init(&blp->blp_lock);
1289
1290         INIT_LIST_HEAD(&blp->blp_list);
1291
1292 #ifdef __KERNEL__
1293         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1294                 struct ldlm_bl_thread_data bltd = {
1295                         .bltd_num = i,
1296                         .bltd_blp = blp,
1297                 };
1298                 init_completion(&blp->blp_comp);
1299                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1300                 if (rc < 0) {
1301                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1302                         GOTO(out_thread, rc);
1303                 }
1304                 wait_for_completion(&blp->blp_comp);
1305         }
1306
1307         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1308                                     LDLM_NUM_THREADS, "ldlm_cn");
1309         if (rc)
1310                 GOTO(out_thread, rc);
1311
1312         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1313                                     LDLM_NUM_THREADS, "ldlm_cb");
1314         if (rc)
1315                 GOTO(out_thread, rc);
1316
1317         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1318         spin_lock_init(&expired_lock_thread.elt_lock);
1319         expired_lock_thread.elt_state = ELT_STOPPED;
1320         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1321
1322         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1323         if (rc < 0) {
1324                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1325                 GOTO(out_thread, rc);
1326         }
1327
1328         wait_event(expired_lock_thread.elt_waitq,
1329                    expired_lock_thread.elt_state == ELT_READY);
1330
1331         INIT_LIST_HEAD(&waiting_locks_list);
1332         spin_lock_init(&waiting_locks_spinlock);
1333         waiting_locks_timer.function = waiting_locks_callback;
1334         waiting_locks_timer.data = 0;
1335         init_timer(&waiting_locks_timer);
1336 #endif
1337
1338         RETURN(0);
1339
1340 #ifdef __KERNEL__
1341  out_thread:
1342         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1343         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1344 #endif
1345
1346  out_proc:
1347 #ifdef __KERNEL__
1348         ldlm_proc_cleanup();
1349  out_free:
1350 #endif
1351         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1352         ldlm_state = NULL;
1353         return rc;
1354 }
1355
1356 static int ldlm_cleanup(int force)
1357 {
1358 #ifdef __KERNEL__
1359         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1360 #endif
1361         ENTRY;
1362
1363         if (!list_empty(&ldlm_namespace_list)) {
1364                 CERROR("ldlm still has namespaces; clean these up first.\n");
1365                 ldlm_dump_all_namespaces();
1366                 RETURN(-EBUSY);
1367         }
1368
1369 #ifdef __KERNEL__
1370         while (atomic_read(&blp->blp_num_threads) > 0) {
1371                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1372
1373                 init_completion(&blp->blp_comp);
1374
1375                 spin_lock(&blp->blp_lock);
1376                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1377                 wake_up(&blp->blp_waitq);
1378                 spin_unlock(&blp->blp_lock);
1379
1380                 wait_for_completion(&blp->blp_comp);
1381         }
1382         OBD_FREE(blp, sizeof(*blp));
1383
1384         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1385         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1386         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1387         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1388         ldlm_proc_cleanup();
1389
1390         expired_lock_thread.elt_state = ELT_TERMINATE;
1391         wake_up(&expired_lock_thread.elt_waitq);
1392         wait_event(expired_lock_thread.elt_waitq,
1393                    expired_lock_thread.elt_state == ELT_STOPPED);
1394
1395 #endif
1396
1397         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1398         ldlm_state = NULL;
1399
1400         RETURN(0);
1401 }
1402
1403 int __init ldlm_init(void)
1404 {
1405         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1406                                                sizeof(struct ldlm_resource), 0,
1407                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1408         if (ldlm_resource_slab == NULL)
1409                 return -ENOMEM;
1410
1411         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1412                                            sizeof(struct ldlm_lock), 0,
1413                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1414         if (ldlm_lock_slab == NULL) {
1415                 kmem_cache_destroy(ldlm_resource_slab);
1416                 return -ENOMEM;
1417         }
1418
1419         l_lock_init(&ldlm_handle_lock);
1420
1421         return 0;
1422 }
1423
1424 void __exit ldlm_exit(void)
1425 {
1426         if ( ldlm_refcount )
1427                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1428         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
1429                 CERROR("couldn't free ldlm resource slab\n");
1430         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
1431                 CERROR("couldn't free ldlm lock slab\n");
1432 }
1433
1434 /* ldlm_flock.c */
1435 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1436
1437 /* ldlm_extent.c */
1438 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1439
1440 /* ldlm_lock.c */
1441 EXPORT_SYMBOL(ldlm_get_processing_policy);
1442 EXPORT_SYMBOL(ldlm_lock2desc);
1443 EXPORT_SYMBOL(ldlm_register_intent);
1444 EXPORT_SYMBOL(ldlm_lockname);
1445 EXPORT_SYMBOL(ldlm_typename);
1446 EXPORT_SYMBOL(ldlm_lock2handle);
1447 EXPORT_SYMBOL(__ldlm_handle2lock);
1448 EXPORT_SYMBOL(ldlm_lock_get);
1449 EXPORT_SYMBOL(ldlm_lock_put);
1450 EXPORT_SYMBOL(ldlm_lock_match);
1451 EXPORT_SYMBOL(ldlm_lock_cancel);
1452 EXPORT_SYMBOL(ldlm_lock_addref);
1453 EXPORT_SYMBOL(ldlm_lock_decref);
1454 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1455 EXPORT_SYMBOL(ldlm_lock_change_resource);
1456 EXPORT_SYMBOL(ldlm_lock_set_data);
1457 EXPORT_SYMBOL(ldlm_it2str);
1458 EXPORT_SYMBOL(ldlm_lock_dump);
1459 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1460 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1461 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1462 EXPORT_SYMBOL(ldlm_lock_allow_match);
1463
1464 /* ldlm_request.c */
1465 EXPORT_SYMBOL(ldlm_completion_ast);
1466 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1467 EXPORT_SYMBOL(ldlm_cli_convert);
1468 EXPORT_SYMBOL(ldlm_cli_enqueue);
1469 EXPORT_SYMBOL(ldlm_cli_cancel);
1470 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1471 EXPORT_SYMBOL(ldlm_replay_locks);
1472 EXPORT_SYMBOL(ldlm_resource_foreach);
1473 EXPORT_SYMBOL(ldlm_namespace_foreach);
1474 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1475 EXPORT_SYMBOL(ldlm_change_cbdata);
1476
1477 /* ldlm_lockd.c */
1478 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1479 EXPORT_SYMBOL(ldlm_server_completion_ast);
1480 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1481 EXPORT_SYMBOL(ldlm_handle_enqueue);
1482 EXPORT_SYMBOL(ldlm_handle_cancel);
1483 EXPORT_SYMBOL(ldlm_handle_convert);
1484 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1485 EXPORT_SYMBOL(ldlm_get_ref);
1486 EXPORT_SYMBOL(ldlm_put_ref);
1487
1488 #if 0
1489 /* ldlm_test.c */
1490 EXPORT_SYMBOL(ldlm_test);
1491 EXPORT_SYMBOL(ldlm_regression_start);
1492 EXPORT_SYMBOL(ldlm_regression_stop);
1493 #endif
1494
1495 /* ldlm_resource.c */
1496 EXPORT_SYMBOL(ldlm_namespace_new);
1497 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1498 EXPORT_SYMBOL(ldlm_namespace_free);
1499 EXPORT_SYMBOL(ldlm_namespace_dump);
1500 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1501 EXPORT_SYMBOL(ldlm_resource_get);
1502 EXPORT_SYMBOL(ldlm_resource_putref);
1503
1504 /* l_lock.c */
1505 EXPORT_SYMBOL(l_lock);
1506 EXPORT_SYMBOL(l_unlock);
1507
1508 /* ldlm_lib.c */
1509 EXPORT_SYMBOL(client_obd_setup);
1510 EXPORT_SYMBOL(client_obd_cleanup);
1511 EXPORT_SYMBOL(client_connect_import);
1512 EXPORT_SYMBOL(client_disconnect_export);
1513 EXPORT_SYMBOL(target_abort_recovery);
1514 EXPORT_SYMBOL(target_handle_connect);
1515 EXPORT_SYMBOL(target_destroy_export);
1516 EXPORT_SYMBOL(target_cancel_recovery_timer);
1517 EXPORT_SYMBOL(target_send_reply);
1518 EXPORT_SYMBOL(target_queue_recovery_request);
1519 EXPORT_SYMBOL(target_handle_ping);
1520 EXPORT_SYMBOL(target_handle_disconnect);
1521 EXPORT_SYMBOL(target_queue_final_reply);