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) 2002-2004 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #ifndef EXPORT_SYMTAB
  25 # define EXPORT_SYMTAB
  26 #endif
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28
  29 #ifdef __KERNEL__
  30 # include <linux/module.h>
  31 # include <linux/slab.h>
  32 # include <linux/init.h>
  33 # include <linux/wait.h>
  34 #else
  35 # include <liblustre.h>
  36 #endif
  37
  38 #include <linux/lustre_dlm.h>
  39 #include <linux/obd_class.h>
  40 #include <libcfs/list.h>
  41 #include "ldlm_internal.h"
  42
  43 extern kmem_cache_t *ldlm_resource_slab;
  44 extern kmem_cache_t *ldlm_lock_slab;
  45 extern struct lustre_lock ldlm_handle_lock;
  46 extern struct list_head ldlm_namespace_list;
  47
  48 static DECLARE_MUTEX(ldlm_ref_sem);
  49 static int ldlm_refcount;
  50
  51 /* LDLM state */
  52
  53 static struct ldlm_state *ldlm_state;
  54
  55 inline unsigned long round_timeout(unsigned long timeout)
  56 {
  57         return ((timeout / HZ) + 1) * HZ;
  58 }
  59
  60 #ifdef __KERNEL__
  61 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
  62 static spinlock_t waiting_locks_spinlock;
  63 static struct list_head waiting_locks_list;
  64 static struct timer_list waiting_locks_timer;
  65
  66 static struct expired_lock_thread {
  67         wait_queue_head_t         elt_waitq;
  68         int                       elt_state;
  69         int                       elt_dump;
  70         struct list_head          elt_expired_locks;
  71 } expired_lock_thread;
  72 #endif
  73
  74 #define ELT_STOPPED   0
  75 #define ELT_READY     1
  76 #define ELT_TERMINATE 2
  77
  78 struct ldlm_bl_pool {
  79         spinlock_t              blp_lock;
  80         struct list_head        blp_list;
  81         wait_queue_head_t       blp_waitq;
  82         atomic_t                blp_num_threads;
  83         struct completion       blp_comp;
  84 };
  85
  86 struct ldlm_bl_work_item {
  87         struct list_head        blwi_entry;
  88         struct ldlm_namespace   *blwi_ns;
  89         struct ldlm_lock_desc   blwi_ld;
  90         struct ldlm_lock        *blwi_lock;
  91 };
  92
  93 #ifdef __KERNEL__
  94
  95 static inline int have_expired_locks(void)
  96 {
  97         int need_to_run;
  98
  99         spin_lock_bh(&waiting_locks_spinlock);
 100         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 101         spin_unlock_bh(&waiting_locks_spinlock);
 102
 103         RETURN(need_to_run);
 104 }
 105
 106 static int expired_lock_main(void *arg)
 107 {
 108         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 109         struct l_wait_info lwi = { 0 };
 110         unsigned long flags;
 111
 112         ENTRY;
 113         lock_kernel();
 114         kportal_daemonize("ldlm_elt");
 115
 116         SIGNAL_MASK_LOCK(current, flags);
 117         sigfillset(&current->blocked);
 118         RECALC_SIGPENDING;
 119         SIGNAL_MASK_UNLOCK(current, flags);
 120
 121         unlock_kernel();
 122
 123         expired_lock_thread.elt_state = ELT_READY;
 124         wake_up(&expired_lock_thread.elt_waitq);
 125
 126         while (1) {
 127                 l_wait_event(expired_lock_thread.elt_waitq,
 128                              have_expired_locks() ||
 129                              expired_lock_thread.elt_state == ELT_TERMINATE,
 130                              &lwi);
 131
 132                 spin_lock_bh(&waiting_locks_spinlock);
 133                 if (expired_lock_thread.elt_dump) {
 134                         spin_unlock_bh(&waiting_locks_spinlock);
 135
 136                         /* from waiting_locks_callback, but not in timer */
 137                         portals_debug_dumplog();
 138                         portals_run_lbug_upcall(__FILE__,
 139                                                 "waiting_locks_callback",
 140                                                 expired_lock_thread.elt_dump);
 141
 142                         spin_lock_bh(&waiting_locks_spinlock);
 143                         expired_lock_thread.elt_dump = 0;
 144                 }
 145
 146                 while (!list_empty(expired)) {
 147                         struct obd_export *export;
 148                         struct ldlm_lock *lock;
 149
 150                         lock = list_entry(expired->next, struct ldlm_lock,
 151                                           l_pending_chain);
 152                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
 153                             (void *)lock >= LP_POISON) {
 154                                 spin_unlock_bh(&waiting_locks_spinlock);
 155                                 CERROR("free lock on elt list %p\n", lock);
 156                                 LBUG();
 157                         }
 158                         list_del_init(&lock->l_pending_chain);
 159                         if ((void *)lock->l_export < LP_POISON + PAGE_SIZE &&
 160                             (void *)lock->l_export >= LP_POISON) {
 161                                 CERROR("lock with free export on elt list %p\n",
 162                                        lock->l_export);
 163                                 lock->l_export = NULL;
 164                                 LDLM_ERROR(lock, "free export");
 165                                 continue;
 166                         }
 167                         export = class_export_get(lock->l_export);
 168                         spin_unlock_bh(&waiting_locks_spinlock);
 169
 170                         class_fail_export(export);
 171                         class_export_put(export);
 172                         spin_lock_bh(&waiting_locks_spinlock);
 173                 }
 174                 spin_unlock_bh(&waiting_locks_spinlock);
 175
 176                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 177                         break;
 178         }
 179
 180         expired_lock_thread.elt_state = ELT_STOPPED;
 181         wake_up(&expired_lock_thread.elt_waitq);
 182         RETURN(0);
 183 }
 184
 185 /* This is called from within a timer interrupt and cannot schedule */
 186 static void waiting_locks_callback(unsigned long unused)
 187 {
 188         struct ldlm_lock *lock, *last = NULL;
 189         char str[PTL_NALFMT_SIZE];
 190
 191         if (obd_dump_on_timeout)
 192                 portals_debug_dumplog();
 193
 194         spin_lock_bh(&waiting_locks_spinlock);
 195         while (!list_empty(&waiting_locks_list)) {
 196                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 197                                   l_pending_chain);
 198
 199                 if (time_after(lock->l_callback_timeout, jiffies))
 200                         break;
 201
 202                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 203                            "%s@%s nid %s ",lock->l_export->exp_client_uuid.uuid,
 204                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 205                            ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer,str));
 206
 207                 if (lock == last) {
 208                         LDLM_ERROR(lock, "waiting on lock multiple times");
 209                         CERROR("wll %p n/p %p/%p, l_pending %p n/p %p/%p\n",
 210                                &waiting_locks_list,
 211                                waiting_locks_list.next, waiting_locks_list.prev,
 212                                &lock->l_pending_chain,
 213                                lock->l_pending_chain.next,
 214                                lock->l_pending_chain.prev);
 215
 216                         INIT_LIST_HEAD(&waiting_locks_list);    /* HACK */
 217                         expired_lock_thread.elt_dump = __LINE__;
 218                         spin_unlock_bh(&waiting_locks_spinlock);
 219
 220                         /* LBUG(); */
 221                         CEMERG("would be an LBUG, but isn't (bug 5653)\n");
 222                         portals_debug_dumpstack(NULL);
 223                         /*blocks* portals_debug_dumplog(); */
 224                         /*blocks* portals_run_lbug_upcall(file, func, line); */
 225                         break;
 226                 }
 227                 last = lock;
 228
 229                 list_del(&lock->l_pending_chain);
 230                 list_add(&lock->l_pending_chain,
 231                          &expired_lock_thread.elt_expired_locks);
 232
 233                 wake_up(&expired_lock_thread.elt_waitq);
 234         }
 235
 236         /*
 237          * Make sure the timer will fire again if we have any locks
 238          * left.
 239          */
 240         if (!list_empty(&waiting_locks_list)) {
 241                 unsigned long timeout_rounded;
 242                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 243                                   l_pending_chain);
 244                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 245                 mod_timer(&waiting_locks_timer, timeout_rounded);
 246         }
 247         spin_unlock_bh(&waiting_locks_spinlock);
 248 }
 249
 250 /*
 251  * Indicate that we're waiting for a client to call us back cancelling a given
 252  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 253  * timer to fire appropriately.  (We round up to the next second, to avoid
 254  * floods of timer firings during periods of high lock contention and traffic).
 255  *
 256  * Called with the namespace lock held.
 257  */
 258 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 259 {
 260         unsigned long timeout_rounded;
 261
 262         l_check_ns_lock(lock->l_resource->lr_namespace);
 263
 264         spin_lock_bh(&waiting_locks_spinlock);
 265         if (lock->l_destroyed) {
 266                 static unsigned long next;
 267                 spin_unlock_bh(&waiting_locks_spinlock);
 268                 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
 269                 if (time_after(jiffies, next)) {
 270                         next = jiffies + 14400 * HZ;
 271                         portals_debug_dumpstack(NULL);
 272                 }
 273                 return 0;
 274         }
 275
 276         if (!list_empty(&lock->l_pending_chain)) {
 277                 spin_unlock_bh(&waiting_locks_spinlock);
 278                 LDLM_DEBUG(lock, "not re-adding to wait list");
 279                 return 0;
 280         }
 281
 282         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 283
 284         timeout_rounded = round_timeout(lock->l_callback_timeout);
 285
 286         if (time_before(timeout_rounded, waiting_locks_timer.expires) ||
 287             !timer_pending(&waiting_locks_timer)) {
 288                 mod_timer(&waiting_locks_timer, timeout_rounded);
 289         }
 290         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 291         spin_unlock_bh(&waiting_locks_spinlock);
 292         LDLM_DEBUG(lock, "adding to wait list");
 293         return 1;
 294 }
 295
 296 /*
 297  * Remove a lock from the pending list, likely because it had its cancellation
 298  * callback arrive without incident.  This adjusts the lock-timeout timer if
 299  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 300  *
 301  * Called with namespace lock held.
 302  */
 303 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 304 {
 305         struct list_head *list_next;
 306
 307         l_check_ns_lock(lock->l_resource->lr_namespace);
 308
 309         if (lock->l_export == NULL) {
 310                 /* We don't have a "waiting locks list" on clients. */
 311                 LDLM_DEBUG(lock, "client lock: no-op");
 312                 return 0;
 313         }
 314
 315         spin_lock_bh(&waiting_locks_spinlock);
 316
 317         if (list_empty(&lock->l_pending_chain)) {
 318                 spin_unlock_bh(&waiting_locks_spinlock);
 319                 LDLM_DEBUG(lock, "wasn't waiting");
 320                 return 0;
 321         }
 322
 323         list_next = lock->l_pending_chain.next;
 324         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 325                 /* Removing the head of the list, adjust timer. */
 326                 if (list_next == &waiting_locks_list) {
 327                         /* No more, just cancel. */
 328                         del_timer(&waiting_locks_timer);
 329                 } else {
 330                         struct ldlm_lock *next;
 331                         next = list_entry(list_next, struct ldlm_lock,
 332                                           l_pending_chain);
 333                         mod_timer(&waiting_locks_timer,
 334                                   round_timeout(next->l_callback_timeout));
 335                 }
 336         }
 337         list_del_init(&lock->l_pending_chain);
 338
 339         spin_unlock_bh(&waiting_locks_spinlock);
 340         LDLM_DEBUG(lock, "removed");
 341         return 1;
 342 }
 343
 344 #else /* !__KERNEL__ */
 345
 346 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 347 {
 348         RETURN(1);
 349 }
 350
 351 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 352 {
 353         RETURN(0);
 354 }
 355
 356 #endif /* __KERNEL__ */
 357
 358 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
 359                             const char *ast_type)
 360 {
 361         struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 362         char str[PTL_NALFMT_SIZE];
 363
 364         ptlrpc_peernid2str(&conn->c_peer, str);
 365
 366         LCONSOLE_ERROR("A client on nid %s was evicted from service %s.\n",
 367                        str, lock->l_export->exp_obd->obd_name);
 368
 369         LDLM_ERROR(lock, "%s AST failed (%d): evicting client %s@%s NID "LPX64
 370                    " (%s)", ast_type, rc, lock->l_export->exp_client_uuid.uuid,
 371                    conn->c_remote_uuid.uuid, conn->c_peer.peer_id.nid, str);
 372
 373         if (obd_dump_on_timeout)
 374                 portals_debug_dumplog();
 375         class_fail_export(lock->l_export);
 376 }
 377
 378 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 379                                  struct ptlrpc_request *req, int rc,
 380                                  const char *ast_type)
 381 {
 382         struct ptlrpc_peer *peer = &req->rq_import->imp_connection->c_peer;
 383         char str[PTL_NALFMT_SIZE];
 384
 385         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 386                 LASSERT(lock->l_export);
 387                 if (lock->l_export->exp_libclient) {
 388                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 389                                    " timeout, just cancelling lock", ast_type,
 390                                    ptlrpc_peernid2str(peer, str));
 391                         ldlm_lock_cancel(lock);
 392                         rc = -ERESTART;
 393                 } else {
 394                         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 395                         ldlm_del_waiting_lock(lock);
 396                         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 397                         ldlm_failed_ast(lock, rc, ast_type);
 398                 }
 399         } else if (rc) {
 400                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 401                 if (rc == -EINVAL)
 402                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 403                                    " from %s AST - normal race",
 404                                    ptlrpc_peernid2str(peer, str),
 405                                    req->rq_repmsg->status, ast_type);
 406                 else
 407                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 408                                    "from %s AST", ptlrpc_peernid2str(peer, str),
 409                                    (req->rq_repmsg != NULL) ?
 410                                    req->rq_repmsg->status : 0, ast_type);
 411                 ldlm_lock_cancel(lock);
 412                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 413                 /* Server-side AST functions are called from ldlm_reprocess_all,
 414                  * which needs to be told to please restart its reprocessing. */
 415                 rc = -ERESTART;
 416         }
 417
 418         return rc;
 419 }
 420
 421 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 422                              struct ldlm_lock_desc *desc,
 423                              void *data, int flag)
 424 {
 425         struct ldlm_request *body;
 426         struct ptlrpc_request *req;
 427         int rc = 0, size = sizeof(*body);
 428         ENTRY;
 429
 430         if (flag == LDLM_CB_CANCELING) {
 431                 /* Don't need to do anything here. */
 432                 RETURN(0);
 433         }
 434
 435         LASSERT(lock);
 436
 437         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 438         if (lock->l_granted_mode != lock->l_req_mode) {
 439                 /* this blocking AST will be communicated as part of the
 440                  * completion AST instead */
 441                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 442                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 443                 RETURN(0);
 444         }
 445
 446         if (lock->l_destroyed) {
 447                 /* What's the point? */
 448                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 449                 RETURN(0);
 450         }
 451
 452 #if 0
 453         if (CURRENT_SECONDS - lock->l_export->exp_last_request_time > 30){
 454                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 455                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 456                 RETURN(-ETIMEDOUT);
 457         }
 458 #endif
 459
 460         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 461                               LDLM_BL_CALLBACK, 1, &size, NULL);
 462         if (req == NULL) {
 463                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 464                 RETURN(-ENOMEM);
 465         }
 466
 467         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 468         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 469                sizeof(body->lock_handle1));
 470         memcpy(&body->lock_desc, desc, sizeof(*desc));
 471         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 472
 473         LDLM_DEBUG(lock, "server preparing blocking AST");
 474         req->rq_replen = lustre_msg_size(0, NULL);
 475
 476         if (lock->l_granted_mode == lock->l_req_mode)
 477                 ldlm_add_waiting_lock(lock);
 478         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 479
 480         req->rq_send_state = LUSTRE_IMP_FULL;
 481         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 482         rc = ptlrpc_queue_wait(req);
 483         if (rc != 0)
 484                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 485
 486         ptlrpc_req_finished(req);
 487
 488         RETURN(rc);
 489 }
 490
 491 /* XXX copied from ptlrpc/service.c */
 492 static long timeval_sub(struct timeval *large, struct timeval *small)
 493 {
 494         return (large->tv_sec - small->tv_sec) * 1000000 +
 495                 (large->tv_usec - small->tv_usec);
 496 }
 497
 498 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 499 {
 500         struct ldlm_request *body;
 501         struct ptlrpc_request *req;
 502         struct timeval granted_time;
 503         long total_enqueue_wait;
 504         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 505         ENTRY;
 506
 507         LASSERT(lock != NULL);
 508
 509         do_gettimeofday(&granted_time);
 510         total_enqueue_wait = timeval_sub(&granted_time,&lock->l_enqueued_time);
 511
 512         if (total_enqueue_wait / 1000000 > obd_timeout)
 513                 LDLM_ERROR(lock, "enqueue wait took %luus from %lu",
 514                            total_enqueue_wait, lock->l_enqueued_time.tv_sec);
 515
 516         down(&lock->l_resource->lr_lvb_sem);
 517         if (lock->l_resource->lr_lvb_len) {
 518                 buffers = 2;
 519                 size[1] = lock->l_resource->lr_lvb_len;
 520         }
 521         up(&lock->l_resource->lr_lvb_sem);
 522
 523         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 524                               LDLM_CP_CALLBACK, buffers, size, NULL);
 525         if (req == NULL)
 526                 RETURN(-ENOMEM);
 527
 528         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 529         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 530                sizeof(body->lock_handle1));
 531         body->lock_flags = flags;
 532         ldlm_lock2desc(lock, &body->lock_desc);
 533
 534         if (buffers == 2) {
 535                 void *lvb;
 536
 537                 down(&lock->l_resource->lr_lvb_sem);
 538                 lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 539                                      lock->l_resource->lr_lvb_len);
 540                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 541                        lock->l_resource->lr_lvb_len);
 542                 up(&lock->l_resource->lr_lvb_sem);
 543         }
 544
 545         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 546                    total_enqueue_wait);
 547         req->rq_replen = lustre_msg_size(0, NULL);
 548
 549         req->rq_send_state = LUSTRE_IMP_FULL;
 550         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 551
 552         /* We only send real blocking ASTs after the lock is granted */
 553         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 554         if (lock->l_flags & LDLM_FL_AST_SENT) {
 555                 body->lock_flags |= LDLM_FL_AST_SENT;
 556                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 557         }
 558         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 559
 560         rc = ptlrpc_queue_wait(req);
 561         if (rc != 0)
 562                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 563
 564         ptlrpc_req_finished(req);
 565
 566         RETURN(rc);
 567 }
 568
 569 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 570 {
 571         struct ldlm_resource *res = lock->l_resource;
 572         struct ldlm_request *body;
 573         struct ptlrpc_request *req;
 574         int rc = 0, size = sizeof(*body);
 575         ENTRY;
 576
 577         LASSERT(lock != NULL);
 578
 579         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 580                               LDLM_GL_CALLBACK, 1, &size, NULL);
 581         if (req == NULL)
 582                 RETURN(-ENOMEM);
 583
 584         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 585         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 586                sizeof(body->lock_handle1));
 587         ldlm_lock2desc(lock, &body->lock_desc);
 588
 589         down(&lock->l_resource->lr_lvb_sem);
 590         size = lock->l_resource->lr_lvb_len;
 591         up(&lock->l_resource->lr_lvb_sem);
 592         req->rq_replen = lustre_msg_size(1, &size);
 593
 594         req->rq_send_state = LUSTRE_IMP_FULL;
 595         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 596
 597         rc = ptlrpc_queue_wait(req);
 598         if (rc == -ELDLM_NO_LOCK_DATA)
 599                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 600         else if (rc != 0)
 601                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 602         else
 603                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 604                         (res, req->rq_repmsg, 0, 1);
 605         ptlrpc_req_finished(req);
 606         RETURN(rc);
 607 }
 608
 609 static struct ldlm_lock *
 610 find_existing_lock(struct obd_export *exp, struct lustre_handle *remote_hdl)
 611 {
 612         struct obd_device *obd = exp->exp_obd;
 613         struct list_head *iter;
 614
 615         l_lock(&obd->obd_namespace->ns_lock);
 616         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
 617                 struct ldlm_lock *lock;
 618                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 619                 if (lock->l_remote_handle.cookie == remote_hdl->cookie) {
 620                         LDLM_LOCK_GET(lock);
 621                         l_unlock(&obd->obd_namespace->ns_lock);
 622                         return lock;
 623                 }
 624         }
 625         l_unlock(&obd->obd_namespace->ns_lock);
 626         return NULL;
 627 }
 628
 629
 630 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 631                         ldlm_completion_callback completion_callback,
 632                         ldlm_blocking_callback blocking_callback,
 633                         ldlm_glimpse_callback glimpse_callback)
 634 {
 635         struct obd_device *obddev = req->rq_export->exp_obd;
 636         struct ldlm_reply *dlm_rep;
 637         struct ldlm_request *dlm_req;
 638         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 639         __u32 flags;
 640         ldlm_error_t err = ELDLM_OK;
 641         struct ldlm_lock *lock = NULL;
 642         void *cookie = NULL;
 643         ENTRY;
 644
 645         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 646
 647         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 648                                       lustre_swab_ldlm_request);
 649         if (dlm_req == NULL) {
 650                 CERROR ("Can't unpack dlm_req\n");
 651                 GOTO(out, rc = -EFAULT);
 652         }
 653
 654         flags = dlm_req->lock_flags;
 655
 656         LASSERT(req->rq_export);
 657
 658         if (flags & LDLM_FL_REPLAY) {
 659                 lock = find_existing_lock(req->rq_export,
 660                                           &dlm_req->lock_handle1);
 661                 if (lock != NULL) {
 662                         DEBUG_REQ(D_HA, req, "found existing lock cookie "LPX64,
 663                                   lock->l_handle.h_cookie);
 664                         GOTO(existing_lock, rc = 0);
 665                 }
 666         }
 667
 668         if (dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
 669             dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE) {
 670                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d\n",
 671                           dlm_req->lock_desc.l_resource.lr_type);
 672                 GOTO(out, rc = -EFAULT);
 673         }
 674
 675         if (dlm_req->lock_desc.l_req_mode < LCK_EX ||
 676             dlm_req->lock_desc.l_req_mode > LCK_NL ||
 677             dlm_req->lock_desc.l_req_mode & (dlm_req->lock_desc.l_req_mode-1)) {
 678                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d\n",
 679                           dlm_req->lock_desc.l_req_mode);
 680                 GOTO(out, rc = -EFAULT);
 681         }
 682
 683         /* The lock's callback data might be set in the policy function */
 684         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 685                                 dlm_req->lock_desc.l_resource.lr_name,
 686                                 dlm_req->lock_desc.l_resource.lr_type,
 687                                 dlm_req->lock_desc.l_req_mode,
 688                                 blocking_callback, completion_callback,
 689                                 glimpse_callback, NULL, 0);
 690         if (!lock)
 691                 GOTO(out, rc = -ENOMEM);
 692
 693         do_gettimeofday(&lock->l_enqueued_time);
 694         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 695                sizeof(lock->l_remote_handle));
 696         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 697
 698         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
 699         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 700         /* Don't enqueue a lock onto the export if it has already
 701          * been evicted.  Cancel it now instead. (bug 3822) */
 702         if (req->rq_export->exp_failed) {
 703                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
 704                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 705                 GOTO(out, rc = -ENOTCONN);
 706         }
 707         lock->l_export = class_export_get(req->rq_export);
 708         list_add(&lock->l_export_chain,
 709                  &lock->l_export->exp_ldlm_data.led_held_locks);
 710         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 711
 712 existing_lock:
 713
 714         if (flags & LDLM_FL_HAS_INTENT) {
 715                 /* In this case, the reply buffer is allocated deep in
 716                  * local_lock_enqueue by the policy function. */
 717                 cookie = req;
 718         } else {
 719                 int buffers = 1;
 720
 721                 down(&lock->l_resource->lr_lvb_sem);
 722                 if (lock->l_resource->lr_lvb_len) {
 723                         size[1] = lock->l_resource->lr_lvb_len;
 724                         buffers = 2;
 725                 }
 726                 up(&lock->l_resource->lr_lvb_sem);
 727
 728                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 729                         GOTO(out, rc = -ENOMEM);
 730
 731                 rc = lustre_pack_reply(req, buffers, size, NULL);
 732                 if (rc)
 733                         GOTO(out, rc);
 734         }
 735
 736         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 737                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 738                        sizeof(ldlm_policy_data_t));
 739         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 740                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 741                        sizeof(lock->l_req_extent));
 742
 743         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 744         if (err)
 745                 GOTO(out, err);
 746
 747         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 748         dlm_rep->lock_flags = flags;
 749
 750         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 751         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 752
 753         /* We never send a blocking AST until the lock is granted, but
 754          * we can tell it right now */
 755         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 756         /* Don't move a pending lock onto the export if it has already
 757          * been evicted.  Cancel it now instead. (bug 5683) */
 758         if (req->rq_export->exp_failed ||
 759             OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT)) {
 760                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
 761                 rc = -ENOTCONN;
 762         } else if (lock->l_flags & LDLM_FL_AST_SENT) {
 763                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 764                 if (lock->l_granted_mode == lock->l_req_mode)
 765                         ldlm_add_waiting_lock(lock);
 766         }
 767         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 768
 769         EXIT;
 770  out:
 771         req->rq_status = err;
 772         if (req->rq_reply_state == NULL) {
 773                 err = lustre_pack_reply(req, 0, NULL, NULL);
 774                 if (rc == 0)
 775                         rc = err;
 776                 req->rq_status = rc;
 777         }
 778
 779         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 780          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 781         if (lock) {
 782                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 783                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 784                            "(err=%d, rc=%d)", err, rc);
 785                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 786
 787                 if (rc == 0) {
 788                         down(&lock->l_resource->lr_lvb_sem);
 789                         size[1] = lock->l_resource->lr_lvb_len;
 790                         if (size[1] > 0) {
 791                                 void *lvb = lustre_msg_buf(req->rq_repmsg,
 792                                                            1, size[1]);
 793                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
 794                                          req, lock);
 795
 796                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 797                                        size[1]);
 798                         }
 799                         up(&lock->l_resource->lr_lvb_sem);
 800                 } else {
 801                         ldlm_resource_unlink_lock(lock);
 802                         ldlm_lock_destroy(lock);
 803                 }
 804
 805                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 806                         ldlm_reprocess_all(lock->l_resource);
 807                 LDLM_LOCK_PUT(lock);
 808         }
 809         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 810                           lock, rc);
 811
 812         return rc;
 813 }
 814
 815 int ldlm_handle_convert(struct ptlrpc_request *req)
 816 {
 817         struct ldlm_request *dlm_req;
 818         struct ldlm_reply *dlm_rep;
 819         struct ldlm_lock *lock;
 820         int rc, size = sizeof(*dlm_rep);
 821         ENTRY;
 822
 823         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 824                                       lustre_swab_ldlm_request);
 825         if (dlm_req == NULL) {
 826                 CERROR ("Can't unpack dlm_req\n");
 827                 RETURN (-EFAULT);
 828         }
 829
 830         rc = lustre_pack_reply(req, 1, &size, NULL);
 831         if (rc) {
 832                 CERROR("out of memory\n");
 833                 RETURN(-ENOMEM);
 834         }
 835         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 836         dlm_rep->lock_flags = dlm_req->lock_flags;
 837
 838         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 839         if (!lock) {
 840                 req->rq_status = EINVAL;
 841         } else {
 842                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 843                 LDLM_DEBUG(lock, "server-side convert handler START");
 844                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 845                                   &dlm_rep->lock_flags);
 846                 if (ldlm_del_waiting_lock(lock))
 847                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 848                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 849                 req->rq_status = 0;
 850         }
 851
 852         if (lock) {
 853                 if (!req->rq_status)
 854                         ldlm_reprocess_all(lock->l_resource);
 855                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 856                 LDLM_DEBUG(lock, "server-side convert handler END");
 857                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 858                 LDLM_LOCK_PUT(lock);
 859         } else
 860                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 861
 862         RETURN(0);
 863 }
 864
 865 int ldlm_handle_cancel(struct ptlrpc_request *req)
 866 {
 867         struct ldlm_request *dlm_req;
 868         struct ldlm_lock *lock;
 869         struct ldlm_resource *res;
 870         int rc;
 871         ENTRY;
 872
 873         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 874                                       lustre_swab_ldlm_request);
 875         if (dlm_req == NULL) {
 876                 CERROR("bad request buffer for cancel\n");
 877                 RETURN(-EFAULT);
 878         }
 879
 880         rc = lustre_pack_reply(req, 0, NULL, NULL);
 881         if (rc) {
 882                 CERROR("out of memory\n");
 883                 RETURN(-ENOMEM);
 884         }
 885
 886         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 887         if (!lock) {
 888                 CERROR("received cancel for unknown lock cookie "LPX64
 889                        " from client %s id %s\n",
 890                        dlm_req->lock_handle1.cookie,
 891                        req->rq_export->exp_client_uuid.uuid,
 892                        req->rq_peerstr);
 893                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 894                                   "(cookie "LPU64")",
 895                                   dlm_req->lock_handle1.cookie);
 896                 req->rq_status = ESTALE;
 897         } else {
 898                 res = lock->l_resource;
 899                 l_lock(&res->lr_namespace->ns_lock);
 900                 LDLM_DEBUG(lock, "server-side cancel handler START");
 901                 l_unlock(&res->lr_namespace->ns_lock);
 902                 if (res && res->lr_namespace->ns_lvbo &&
 903                     res->lr_namespace->ns_lvbo->lvbo_update) {
 904                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 905                                 (res, NULL, 0, 0);
 906                                 //(res, req->rq_reqmsg, 1);
 907                 }
 908
 909                 l_lock(&res->lr_namespace->ns_lock);
 910                 ldlm_lock_cancel(lock);
 911                 if (ldlm_del_waiting_lock(lock))
 912                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 913                 l_unlock(&res->lr_namespace->ns_lock);
 914                 req->rq_status = rc;
 915         }
 916
 917         if (ptlrpc_reply(req) != 0)
 918                 LBUG();
 919
 920         if (lock) {
 921                 ldlm_reprocess_all(lock->l_resource);
 922                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 923                 LDLM_DEBUG(lock, "server-side cancel handler END");
 924                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 925                 LDLM_LOCK_PUT(lock);
 926         }
 927
 928         RETURN(0);
 929 }
 930
 931 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 932                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 933 {
 934         int do_ast;
 935         ENTRY;
 936
 937         l_lock(&ns->ns_lock);
 938         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 939
 940         lock->l_flags |= LDLM_FL_CBPENDING;
 941         do_ast = (!lock->l_readers && !lock->l_writers);
 942
 943         if (do_ast) {
 944                 LDLM_DEBUG(lock, "already unused, calling "
 945                            "callback (%p)", lock->l_blocking_ast);
 946                 if (lock->l_blocking_ast != NULL) {
 947                         l_unlock(&ns->ns_lock);
 948                         l_check_no_ns_lock(ns);
 949                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 950                                              LDLM_CB_BLOCKING);
 951                         l_lock(&ns->ns_lock);
 952                 }
 953         } else {
 954                 LDLM_DEBUG(lock, "Lock still has references, will be"
 955                            " cancelled later");
 956         }
 957
 958         LDLM_DEBUG(lock, "client blocking callback handler END");
 959         l_unlock(&ns->ns_lock);
 960         LDLM_LOCK_PUT(lock);
 961         EXIT;
 962 }
 963
 964 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 965                                     struct ldlm_namespace *ns,
 966                                     struct ldlm_request *dlm_req,
 967                                     struct ldlm_lock *lock)
 968 {
 969         LIST_HEAD(ast_list);
 970         ENTRY;
 971
 972         l_lock(&ns->ns_lock);
 973         LDLM_DEBUG(lock, "client completion callback handler START");
 974
 975         /* If we receive the completion AST before the actual enqueue returned,
 976          * then we might need to switch lock modes, resources, or extents. */
 977         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 978                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 979                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 980         }
 981
 982         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 983                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 984                        sizeof(lock->l_policy_data));
 985                 LDLM_DEBUG(lock, "completion AST, new policy data");
 986         }
 987
 988         ldlm_resource_unlink_lock(lock);
 989         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 990                    &lock->l_resource->lr_name,
 991                    sizeof(lock->l_resource->lr_name)) != 0) {
 992                 ldlm_lock_change_resource(ns, lock,
 993                                          dlm_req->lock_desc.l_resource.lr_name);
 994                 LDLM_DEBUG(lock, "completion AST, new resource");
 995         }
 996
 997         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 998                 lock->l_flags |= LDLM_FL_CBPENDING;
 999                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1000         }
1001
1002         if (lock->l_lvb_len) {
1003                 void *lvb;
1004                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
1005                                          lock->l_lvb_swabber);
1006                 if (lvb == NULL) {
1007                         LDLM_ERROR(lock, "completion AST did not contain "
1008                                    "expected LVB!");
1009                 } else {
1010                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1011                 }
1012         }
1013
1014         lock->l_resource->lr_tmp = &ast_list;
1015         ldlm_grant_lock(lock, req, sizeof(*req), 1);
1016         lock->l_resource->lr_tmp = NULL;
1017         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1018         l_unlock(&ns->ns_lock);
1019         LDLM_LOCK_PUT(lock);
1020
1021         ldlm_run_ast_work(ns, &ast_list);
1022
1023         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1024                           lock);
1025         EXIT;
1026 }
1027
1028 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1029                                     struct ldlm_namespace *ns,
1030                                     struct ldlm_request *dlm_req,
1031                                     struct ldlm_lock *lock)
1032 {
1033         int rc = -ENOSYS;
1034         ENTRY;
1035
1036         l_lock(&ns->ns_lock);
1037         LDLM_DEBUG(lock, "client glimpse AST callback handler");
1038
1039         if (lock->l_glimpse_ast != NULL) {
1040                 l_unlock(&ns->ns_lock);
1041                 l_check_no_ns_lock(ns);
1042                 rc = lock->l_glimpse_ast(lock, req);
1043                 l_lock(&ns->ns_lock);
1044         }
1045
1046         if (req->rq_repmsg != NULL) {
1047                 ptlrpc_reply(req);
1048         } else {
1049                 req->rq_status = rc;
1050                 ptlrpc_error(req);
1051         }
1052
1053         l_unlock(&ns->ns_lock);
1054         if (lock->l_granted_mode == LCK_PW &&
1055             !lock->l_readers && !lock->l_writers &&
1056             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
1057                 if (ldlm_bl_to_thread(ns, NULL, lock))
1058                         ldlm_handle_bl_callback(ns, NULL, lock);
1059                 EXIT;
1060                 return;
1061         }
1062
1063         LDLM_LOCK_PUT(lock);
1064         EXIT;
1065 }
1066
1067 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1068 {
1069         req->rq_status = rc;
1070         if (req->rq_reply_state == NULL) {
1071                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1072                 if (rc)
1073                         return rc;
1074         }
1075         return ptlrpc_reply(req);
1076 }
1077
1078 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1079                       struct ldlm_lock *lock)
1080 {
1081 #ifdef __KERNEL__
1082         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1083         struct ldlm_bl_work_item *blwi;
1084         ENTRY;
1085
1086         OBD_ALLOC(blwi, sizeof(*blwi));
1087         if (blwi == NULL)
1088                 RETURN(-ENOMEM);
1089
1090         blwi->blwi_ns = ns;
1091         if (ld != NULL)
1092                 blwi->blwi_ld = *ld;
1093         blwi->blwi_lock = lock;
1094
1095         spin_lock(&blp->blp_lock);
1096         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1097         wake_up(&blp->blp_waitq);
1098         spin_unlock(&blp->blp_lock);
1099
1100         RETURN(0);
1101 #else
1102         RETURN(-ENOSYS);
1103 #endif
1104 }
1105
1106 static int ldlm_callback_handler(struct ptlrpc_request *req)
1107 {
1108         struct ldlm_namespace *ns;
1109         struct ldlm_request *dlm_req;
1110         struct ldlm_lock *lock;
1111         int rc;
1112         ENTRY;
1113
1114         /* Requests arrive in sender's byte order.  The ptlrpc service
1115          * handler has already checked and, if necessary, byte-swapped the
1116          * incoming request message body, but I am responsible for the
1117          * message buffers. */
1118
1119         if (req->rq_export == NULL) {
1120                 struct ldlm_request *dlm_req;
1121
1122                 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1123                        "export cookie "LPX64"; this is "
1124                        "normal if this node rebooted with a lock held\n",
1125                        req->rq_reqmsg->opc,
1126                        req->rq_peerstr,
1127                        req->rq_reqmsg->handle.cookie);
1128
1129                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1130                                              lustre_swab_ldlm_request);
1131                 if (dlm_req != NULL)
1132                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1133                                dlm_req->lock_handle1.cookie);
1134
1135                 ldlm_callback_reply(req, -ENOTCONN);
1136                 RETURN(0);
1137         }
1138
1139         LASSERT(req->rq_export != NULL);
1140         LASSERT(req->rq_export->exp_obd != NULL);
1141
1142         switch(req->rq_reqmsg->opc) {
1143         case LDLM_BL_CALLBACK:
1144                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1145                 break;
1146         case LDLM_CP_CALLBACK:
1147                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1148                 break;
1149         case LDLM_GL_CALLBACK:
1150                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1151                 break;
1152         case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1153                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1154                 rc = llog_origin_handle_cancel(req);
1155                 ldlm_callback_reply(req, rc);
1156                 RETURN(0);
1157         case OBD_QC_CALLBACK:
1158                 OBD_FAIL_RETURN(OBD_FAIL_OBD_QC_CALLBACK_NET, 0);
1159                 rc = target_handle_qc_callback(req);
1160                 ldlm_callback_reply(req, rc);
1161                 RETURN(0);
1162         case QUOTA_DQACQ:
1163         case QUOTA_DQREL:
1164                 /* reply in handler */
1165                 rc = target_handle_dqacq_callback(req);
1166                 RETURN(0);
1167         case LLOG_ORIGIN_HANDLE_CREATE:
1168                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1169                 rc = llog_origin_handle_create(req);
1170                 ldlm_callback_reply(req, rc);
1171                 RETURN(0);
1172         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1173                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1174                 rc = llog_origin_handle_next_block(req);
1175                 ldlm_callback_reply(req, rc);
1176                 RETURN(0);
1177         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1178                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1179                 rc = llog_origin_handle_read_header(req);
1180                 ldlm_callback_reply(req, rc);
1181                 RETURN(0);
1182         case LLOG_ORIGIN_HANDLE_CLOSE:
1183                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1184                 rc = llog_origin_handle_close(req);
1185                 ldlm_callback_reply(req, rc);
1186                 RETURN(0);
1187         default:
1188                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1189                 ldlm_callback_reply(req, -EPROTO);
1190                 RETURN(0);
1191         }
1192
1193         ns = req->rq_export->exp_obd->obd_namespace;
1194         LASSERT(ns != NULL);
1195
1196         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1197                                       lustre_swab_ldlm_request);
1198         if (dlm_req == NULL) {
1199                 CERROR ("can't unpack dlm_req\n");
1200                 ldlm_callback_reply (req, -EPROTO);
1201                 RETURN (0);
1202         }
1203
1204         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1205         if (!lock) {
1206                 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
1207                        dlm_req->lock_handle1.cookie);
1208                 ldlm_callback_reply(req, -EINVAL);
1209                 RETURN(0);
1210         }
1211
1212         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1213         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1214
1215         /* We want the ost thread to get this reply so that it can respond
1216          * to ost requests (write cache writeback) that might be triggered
1217          * in the callback.
1218          *
1219          * But we'd also like to be able to indicate in the reply that we're
1220          * cancelling right now, because it's unused, or have an intent result
1221          * in the reply, so we might have to push the responsibility for sending
1222          * the reply down into the AST handlers, alas. */
1223
1224         switch (req->rq_reqmsg->opc) {
1225         case LDLM_BL_CALLBACK:
1226                 CDEBUG(D_INODE, "blocking ast\n");
1227                 ldlm_callback_reply(req, 0);
1228                 if (ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock))
1229                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1230                 break;
1231         case LDLM_CP_CALLBACK:
1232                 CDEBUG(D_INODE, "completion ast\n");
1233                 ldlm_callback_reply(req, 0);
1234                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1235                 break;
1236         case LDLM_GL_CALLBACK:
1237                 CDEBUG(D_INODE, "glimpse ast\n");
1238                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1239                 break;
1240         default:
1241                 LBUG();                         /* checked above */
1242         }
1243
1244         RETURN(0);
1245 }
1246
1247 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1248 {
1249         int rc;
1250         ENTRY;
1251
1252         /* Requests arrive in sender's byte order.  The ptlrpc service
1253          * handler has already checked and, if necessary, byte-swapped the
1254          * incoming request message body, but I am responsible for the
1255          * message buffers. */
1256
1257         if (req->rq_export == NULL) {
1258                 struct ldlm_request *dlm_req;
1259
1260                 CERROR("operation %d from %s with bad export cookie "LPU64"\n",
1261                        req->rq_reqmsg->opc, req->rq_peerstr,
1262                        req->rq_reqmsg->handle.cookie);
1263
1264                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1265                                              lustre_swab_ldlm_request);
1266                 if (dlm_req != NULL)
1267                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1268
1269                 ldlm_callback_reply(req, -ENOTCONN);
1270                 RETURN(0);
1271         }
1272
1273         switch (req->rq_reqmsg->opc) {
1274
1275         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1276         case LDLM_CANCEL:
1277                 CDEBUG(D_INODE, "cancel\n");
1278                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1279                 rc = ldlm_handle_cancel(req);
1280                 if (rc)
1281                         break;
1282                 RETURN(0);
1283         case OBD_LOG_CANCEL:
1284                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1285                 rc = llog_origin_handle_cancel(req);
1286                 ldlm_callback_reply(req, rc);
1287                 RETURN(0);
1288         default:
1289                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1290                 ldlm_callback_reply(req, -EINVAL);
1291         }
1292
1293         RETURN(0);
1294 }
1295
1296 #ifdef __KERNEL__
1297 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1298 {
1299         struct ldlm_bl_work_item *blwi = NULL;
1300
1301         spin_lock(&blp->blp_lock);
1302         if (!list_empty(&blp->blp_list)) {
1303                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1304                                   blwi_entry);
1305                 list_del(&blwi->blwi_entry);
1306         }
1307         spin_unlock(&blp->blp_lock);
1308
1309         return blwi;
1310 }
1311
1312 struct ldlm_bl_thread_data {
1313         int                     bltd_num;
1314         struct ldlm_bl_pool     *bltd_blp;
1315 };
1316
1317 static int ldlm_bl_thread_main(void *arg)
1318 {
1319         struct ldlm_bl_thread_data *bltd = arg;
1320         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1321         unsigned long flags;
1322         ENTRY;
1323
1324         /* XXX boiler-plate */
1325         {
1326                 char name[sizeof(current->comm)];
1327                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1328                          bltd->bltd_num);
1329                 kportal_daemonize(name);
1330         }
1331         SIGNAL_MASK_LOCK(current, flags);
1332         sigfillset(&current->blocked);
1333         RECALC_SIGPENDING;
1334         SIGNAL_MASK_UNLOCK(current, flags);
1335
1336         atomic_inc(&blp->blp_num_threads);
1337         complete(&blp->blp_comp);
1338
1339         while(1) {
1340                 struct l_wait_info lwi = { 0 };
1341                 struct ldlm_bl_work_item *blwi = NULL;
1342
1343                 l_wait_event_exclusive(blp->blp_waitq,
1344                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1345                                        &lwi);
1346
1347                 if (blwi->blwi_ns == NULL)
1348                         break;
1349
1350                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1351                                         blwi->blwi_lock);
1352                 OBD_FREE(blwi, sizeof(*blwi));
1353         }
1354
1355         atomic_dec(&blp->blp_num_threads);
1356         complete(&blp->blp_comp);
1357         RETURN(0);
1358 }
1359
1360 #endif
1361
1362 static int ldlm_setup(void);
1363 static int ldlm_cleanup(int force);
1364
1365 int ldlm_get_ref(void)
1366 {
1367         int rc = 0;
1368         down(&ldlm_ref_sem);
1369         if (++ldlm_refcount == 1) {
1370                 rc = ldlm_setup();
1371                 if (rc)
1372                         ldlm_refcount--;
1373         }
1374         up(&ldlm_ref_sem);
1375
1376         RETURN(rc);
1377 }
1378
1379 void ldlm_put_ref(int force)
1380 {
1381         down(&ldlm_ref_sem);
1382         if (ldlm_refcount == 1) {
1383                 int rc = ldlm_cleanup(force);
1384                 if (rc)
1385                         CERROR("ldlm_cleanup failed: %d\n", rc);
1386                 else
1387                         ldlm_refcount--;
1388         } else {
1389                 ldlm_refcount--;
1390         }
1391         up(&ldlm_ref_sem);
1392
1393         EXIT;
1394 }
1395
1396 static int ldlm_setup(void)
1397 {
1398         struct ldlm_bl_pool *blp;
1399         int rc = 0;
1400 #ifdef __KERNEL__
1401         int i;
1402 #endif
1403         ENTRY;
1404
1405         if (ldlm_state != NULL)
1406                 RETURN(-EALREADY);
1407
1408         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1409         if (ldlm_state == NULL)
1410                 RETURN(-ENOMEM);
1411
1412 #ifdef LPROCFS
1413         rc = ldlm_proc_setup();
1414         if (rc != 0)
1415                 GOTO(out_free, rc);
1416 #endif
1417
1418         ldlm_state->ldlm_cb_service =
1419                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1420                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1421                                 1500, ldlm_callback_handler, "ldlm_cbd",
1422                                 ldlm_svc_proc_dir, NULL);
1423
1424         if (!ldlm_state->ldlm_cb_service) {
1425                 CERROR("failed to start service\n");
1426                 GOTO(out_proc, rc = -ENOMEM);
1427         }
1428
1429         ldlm_state->ldlm_cancel_service =
1430                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1431                                 LDLM_CANCEL_REQUEST_PORTAL,
1432                                 LDLM_CANCEL_REPLY_PORTAL, 30000,
1433                                 ldlm_cancel_handler, "ldlm_canceld",
1434                                 ldlm_svc_proc_dir, NULL);
1435
1436         if (!ldlm_state->ldlm_cancel_service) {
1437                 CERROR("failed to start service\n");
1438                 GOTO(out_proc, rc = -ENOMEM);
1439         }
1440
1441         OBD_ALLOC(blp, sizeof(*blp));
1442         if (blp == NULL)
1443                 GOTO(out_proc, rc = -ENOMEM);
1444         ldlm_state->ldlm_bl_pool = blp;
1445
1446         atomic_set(&blp->blp_num_threads, 0);
1447         init_waitqueue_head(&blp->blp_waitq);
1448         spin_lock_init(&blp->blp_lock);
1449
1450         INIT_LIST_HEAD(&blp->blp_list);
1451
1452 #ifdef __KERNEL__
1453         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1454                 struct ldlm_bl_thread_data bltd = {
1455                         .bltd_num = i,
1456                         .bltd_blp = blp,
1457                 };
1458                 init_completion(&blp->blp_comp);
1459                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1460                 if (rc < 0) {
1461                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1462                         GOTO(out_thread, rc);
1463                 }
1464                 wait_for_completion(&blp->blp_comp);
1465         }
1466
1467         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1468                                     LDLM_NUM_THREADS, "ldlm_cn");
1469         if (rc)
1470                 GOTO(out_thread, rc);
1471
1472         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1473                                     LDLM_NUM_THREADS, "ldlm_cb");
1474         if (rc)
1475                 GOTO(out_thread, rc);
1476
1477         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1478         expired_lock_thread.elt_state = ELT_STOPPED;
1479         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1480
1481         INIT_LIST_HEAD(&waiting_locks_list);
1482         spin_lock_init(&waiting_locks_spinlock);
1483         waiting_locks_timer.function = waiting_locks_callback;
1484         waiting_locks_timer.data = 0;
1485         init_timer(&waiting_locks_timer);
1486
1487         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1488         if (rc < 0) {
1489                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1490                 GOTO(out_thread, rc);
1491         }
1492
1493         wait_event(expired_lock_thread.elt_waitq,
1494                    expired_lock_thread.elt_state == ELT_READY);
1495 #endif
1496
1497         RETURN(0);
1498
1499 #ifdef __KERNEL__
1500  out_thread:
1501         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1502         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1503 #endif
1504
1505  out_proc:
1506 #ifdef LPROCFS
1507         ldlm_proc_cleanup();
1508  out_free:
1509 #endif
1510         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1511         ldlm_state = NULL;
1512         return rc;
1513 }
1514
1515 static int ldlm_cleanup(int force)
1516 {
1517 #ifdef __KERNEL__
1518         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1519 #endif
1520         ENTRY;
1521
1522         if (!list_empty(&ldlm_namespace_list)) {
1523                 CERROR("ldlm still has namespaces; clean these up first.\n");
1524                 ldlm_dump_all_namespaces(D_DLMTRACE);
1525                 RETURN(-EBUSY);
1526         }
1527
1528 #ifdef __KERNEL__
1529         while (atomic_read(&blp->blp_num_threads) > 0) {
1530                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1531
1532                 init_completion(&blp->blp_comp);
1533
1534                 spin_lock(&blp->blp_lock);
1535                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1536                 wake_up(&blp->blp_waitq);
1537                 spin_unlock(&blp->blp_lock);
1538
1539                 wait_for_completion(&blp->blp_comp);
1540         }
1541         OBD_FREE(blp, sizeof(*blp));
1542
1543         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1544         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1545         ldlm_proc_cleanup();
1546
1547         expired_lock_thread.elt_state = ELT_TERMINATE;
1548         wake_up(&expired_lock_thread.elt_waitq);
1549         wait_event(expired_lock_thread.elt_waitq,
1550                    expired_lock_thread.elt_state == ELT_STOPPED);
1551
1552 #endif
1553
1554         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1555         ldlm_state = NULL;
1556
1557         RETURN(0);
1558 }
1559
1560 int __init ldlm_init(void)
1561 {
1562         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1563                                                sizeof(struct ldlm_resource), 0,
1564                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1565         if (ldlm_resource_slab == NULL)
1566                 return -ENOMEM;
1567
1568         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1569                                            sizeof(struct ldlm_lock), 0,
1570                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1571         if (ldlm_lock_slab == NULL) {
1572                 kmem_cache_destroy(ldlm_resource_slab);
1573                 return -ENOMEM;
1574         }
1575
1576         l_lock_init(&ldlm_handle_lock);
1577
1578         return 0;
1579 }
1580
1581 void __exit ldlm_exit(void)
1582 {
1583         if ( ldlm_refcount )
1584                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1585         LASSERTF(kmem_cache_destroy(ldlm_resource_slab) == 0,
1586                  "couldn't free ldlm resource slab\n");
1587         LASSERTF(kmem_cache_destroy(ldlm_lock_slab) == 0,
1588                  "couldn't free ldlm lock slab\n");
1589 }
1590
1591 /* ldlm_flock.c */
1592 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1593
1594 /* ldlm_extent.c */
1595 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1596
1597 /* ldlm_lock.c */
1598 EXPORT_SYMBOL(ldlm_get_processing_policy);
1599 EXPORT_SYMBOL(ldlm_lock2desc);
1600 EXPORT_SYMBOL(ldlm_register_intent);
1601 EXPORT_SYMBOL(ldlm_lockname);
1602 EXPORT_SYMBOL(ldlm_typename);
1603 EXPORT_SYMBOL(ldlm_lock2handle);
1604 EXPORT_SYMBOL(__ldlm_handle2lock);
1605 EXPORT_SYMBOL(ldlm_lock_get);
1606 EXPORT_SYMBOL(ldlm_lock_put);
1607 EXPORT_SYMBOL(ldlm_lock_match);
1608 EXPORT_SYMBOL(ldlm_lock_cancel);
1609 EXPORT_SYMBOL(ldlm_lock_addref);
1610 EXPORT_SYMBOL(ldlm_lock_decref);
1611 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1612 EXPORT_SYMBOL(ldlm_lock_change_resource);
1613 EXPORT_SYMBOL(ldlm_lock_set_data);
1614 EXPORT_SYMBOL(ldlm_it2str);
1615 EXPORT_SYMBOL(ldlm_lock_dump);
1616 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1617 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1618 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1619 EXPORT_SYMBOL(ldlm_lock_allow_match);
1620
1621 /* ldlm_request.c */
1622 EXPORT_SYMBOL(ldlm_completion_ast);
1623 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1624 EXPORT_SYMBOL(ldlm_cli_convert);
1625 EXPORT_SYMBOL(ldlm_cli_enqueue);
1626 EXPORT_SYMBOL(ldlm_cli_cancel);
1627 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1628 EXPORT_SYMBOL(ldlm_cli_join_lru);
1629 EXPORT_SYMBOL(ldlm_replay_locks);
1630 EXPORT_SYMBOL(ldlm_resource_foreach);
1631 EXPORT_SYMBOL(ldlm_namespace_foreach);
1632 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1633 EXPORT_SYMBOL(ldlm_change_cbdata);
1634
1635 /* ldlm_lockd.c */
1636 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1637 EXPORT_SYMBOL(ldlm_server_completion_ast);
1638 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1639 EXPORT_SYMBOL(ldlm_handle_enqueue);
1640 EXPORT_SYMBOL(ldlm_handle_cancel);
1641 EXPORT_SYMBOL(ldlm_handle_convert);
1642 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1643 EXPORT_SYMBOL(ldlm_get_ref);
1644 EXPORT_SYMBOL(ldlm_put_ref);
1645
1646 /* ldlm_resource.c */
1647 EXPORT_SYMBOL(ldlm_namespace_new);
1648 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1649 EXPORT_SYMBOL(ldlm_namespace_free);
1650 EXPORT_SYMBOL(ldlm_namespace_dump);
1651 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1652 EXPORT_SYMBOL(ldlm_resource_get);
1653 EXPORT_SYMBOL(ldlm_resource_putref);
1654
1655 /* l_lock.c */
1656 EXPORT_SYMBOL(l_lock);
1657 EXPORT_SYMBOL(l_unlock);
1658
1659 /* ldlm_lib.c */
1660 EXPORT_SYMBOL(client_import_add_conn);
1661 EXPORT_SYMBOL(client_import_del_conn);
1662 EXPORT_SYMBOL(client_obd_setup);
1663 EXPORT_SYMBOL(client_obd_cleanup);
1664 EXPORT_SYMBOL(client_connect_import);
1665 EXPORT_SYMBOL(client_disconnect_export);
1666 EXPORT_SYMBOL(target_abort_recovery);
1667 EXPORT_SYMBOL(target_cleanup_recovery);
1668 EXPORT_SYMBOL(target_handle_connect);
1669 EXPORT_SYMBOL(target_destroy_export);
1670 EXPORT_SYMBOL(target_cancel_recovery_timer);
1671 EXPORT_SYMBOL(target_send_reply);
1672 EXPORT_SYMBOL(target_queue_recovery_request);
1673 EXPORT_SYMBOL(target_handle_ping);
1674 EXPORT_SYMBOL(target_handle_disconnect);
1675 EXPORT_SYMBOL(target_queue_final_reply);