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 "ldlm_internal.h"
  41
  42 extern kmem_cache_t *ldlm_resource_slab;
  43 extern kmem_cache_t *ldlm_lock_slab;
  44 extern struct lustre_lock ldlm_handle_lock;
  45 extern struct list_head ldlm_namespace_list;
  46 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
  47 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
  48
  49 static DECLARE_MUTEX(ldlm_ref_sem);
  50 static int ldlm_refcount = 0;
  51
  52 /* LDLM state */
  53
  54 static struct ldlm_state *ldlm ;
  55
  56 inline unsigned long round_timeout(unsigned long timeout)
  57 {
  58         return ((timeout / HZ) + 1) * HZ;
  59 }
  60
  61 #ifdef __KERNEL__
  62 /* XXX should this be per-ldlm? */
  63 static struct list_head waiting_locks_list;
  64 static spinlock_t waiting_locks_spinlock;
  65 static struct timer_list waiting_locks_timer;
  66
  67 static struct expired_lock_thread {
  68         wait_queue_head_t         elt_waitq;
  69         int                       elt_state;
  70         struct list_head          elt_expired_locks;
  71         spinlock_t                elt_lock;
  72 } expired_lock_thread;
  73 #endif
  74
  75 #define ELT_STOPPED   0
  76 #define ELT_READY     1
  77 #define ELT_TERMINATE 2
  78
  79 struct ldlm_bl_pool {
  80         spinlock_t              blp_lock;
  81         struct list_head        blp_list;
  82         wait_queue_head_t       blp_waitq;
  83         atomic_t                blp_num_threads;
  84         struct completion       blp_comp;
  85 };
  86
  87 struct ldlm_bl_work_item {
  88         struct list_head        blwi_entry;
  89         struct ldlm_namespace   *blwi_ns;
  90         struct ldlm_lock_desc   blwi_ld;
  91         struct ldlm_lock        *blwi_lock;
  92 };
  93
  94 #ifdef __KERNEL__
  95
  96 static inline int have_expired_locks(void)
  97 {
  98         int need_to_run;
  99
 100         spin_lock_bh(&expired_lock_thread.elt_lock);
 101         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 102         spin_unlock_bh(&expired_lock_thread.elt_lock);
 103
 104         RETURN(need_to_run);
 105 }
 106
 107 static int expired_lock_main(void *arg)
 108 {
 109         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 110         struct l_wait_info lwi = { 0 };
 111         unsigned long flags;
 112
 113         ENTRY;
 114         lock_kernel();
 115         kportal_daemonize("ldlm_elt");
 116
 117         SIGNAL_MASK_LOCK(current, flags);
 118         sigfillset(&current->blocked);
 119         RECALC_SIGPENDING;
 120         SIGNAL_MASK_UNLOCK(current, flags);
 121
 122         unlock_kernel();
 123
 124         expired_lock_thread.elt_state = ELT_READY;
 125         wake_up(&expired_lock_thread.elt_waitq);
 126
 127         while (1) {
 128                 struct list_head *tmp, *n, work_list;
 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 ldlm_lock *lock;
 137
 138                         list_add(&work_list, expired);
 139                         list_del_init(expired);
 140
 141                         list_for_each_entry(lock, &work_list, l_pending_chain) {
 142                                 LDLM_DEBUG(lock, "moving to work list");
 143                         }
 144
 145                         spin_unlock_bh(&expired_lock_thread.elt_lock);
 146
 147
 148                         list_for_each_safe(tmp, n, &work_list) {
 149                                  lock = list_entry(tmp, struct ldlm_lock,
 150                                                    l_pending_chain);
 151                                  ptlrpc_fail_export(lock->l_export);
 152                         }
 153
 154
 155                         if (!list_empty(&work_list)) {
 156                                 list_for_each_entry(lock, &work_list, l_pending_chain) {
 157                                         LDLM_ERROR(lock, "still on work list!");
 158                                 }
 159                         }
 160                         LASSERTF (list_empty(&work_list),
 161                                   "some exports not failed properly\n");
 162
 163                         spin_lock_bh(&expired_lock_thread.elt_lock);
 164                 }
 165                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 166
 167                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 168                         break;
 169         }
 170
 171         expired_lock_thread.elt_state = ELT_STOPPED;
 172         wake_up(&expired_lock_thread.elt_waitq);
 173         RETURN(0);
 174 }
 175
 176 static void waiting_locks_callback(unsigned long unused)
 177 {
 178         struct ldlm_lock *lock;
 179         char str[PTL_NALFMT_SIZE];
 180
 181         spin_lock_bh(&waiting_locks_spinlock);
 182         while (!list_empty(&waiting_locks_list)) {
 183                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 184                                   l_pending_chain);
 185
 186                 if (lock->l_callback_timeout > jiffies)
 187                         break;
 188
 189                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 190                            "%s@%s nid "LPX64" (%s) ",
 191                            lock->l_export->exp_client_uuid.uuid,
 192                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 193                            lock->l_export->exp_connection->c_peer.peer_nid,
 194                            portals_nid2str(lock->l_export->exp_connection->c_peer.peer_ni->pni_number,
 195                                            lock->l_export->exp_connection->c_peer.peer_nid,
 196                                            str));
 197
 198                 spin_lock_bh(&expired_lock_thread.elt_lock);
 199                 list_del(&lock->l_pending_chain);
 200                 list_add(&lock->l_pending_chain,
 201                          &expired_lock_thread.elt_expired_locks);
 202                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 203                 wake_up(&expired_lock_thread.elt_waitq);
 204         }
 205
 206         /*
 207          * Make sure the timer will fire again if we have any locks
 208          * left.
 209          */
 210         if (!list_empty(&waiting_locks_list)) {
 211                 unsigned long timeout_rounded;
 212                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 213                                   l_pending_chain);
 214                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 215                 mod_timer(&waiting_locks_timer, timeout_rounded);
 216         }
 217         spin_unlock_bh(&waiting_locks_spinlock);
 218 }
 219
 220 /*
 221  * Indicate that we're waiting for a client to call us back cancelling a given
 222  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 223  * timer to fire appropriately.  (We round up to the next second, to avoid
 224  * floods of timer firings during periods of high lock contention and traffic).
 225  */
 226 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 227 {
 228         unsigned long timeout_rounded;
 229
 230         spin_lock_bh(&waiting_locks_spinlock);
 231         if (!list_empty(&lock->l_pending_chain)) {
 232                 LDLM_DEBUG(lock, "not re-adding to wait list");
 233                 spin_unlock_bh(&waiting_locks_spinlock);
 234                 return 0;
 235         }
 236         LDLM_DEBUG(lock, "adding to wait list");
 237
 238         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 239
 240         timeout_rounded = round_timeout(lock->l_callback_timeout);
 241
 242         if (timeout_rounded < waiting_locks_timer.expires ||
 243             !timer_pending(&waiting_locks_timer)) {
 244                 mod_timer(&waiting_locks_timer, timeout_rounded);
 245         }
 246         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 247         spin_unlock_bh(&waiting_locks_spinlock);
 248         return 1;
 249 }
 250
 251 /*
 252  * Remove a lock from the pending list, likely because it had its cancellation
 253  * callback arrive without incident.  This adjusts the lock-timeout timer if
 254  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 255  */
 256 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 257 {
 258         struct list_head *list_next;
 259
 260         if (lock->l_export == NULL) {
 261                 /* We don't have a "waiting locks list" on clients. */
 262                 LDLM_DEBUG(lock, "client lock: no-op");
 263                 return 0;
 264         }
 265
 266         spin_lock_bh(&waiting_locks_spinlock);
 267
 268         if (list_empty(&lock->l_pending_chain)) {
 269                 spin_unlock_bh(&waiting_locks_spinlock);
 270                 LDLM_DEBUG(lock, "wasn't waiting");
 271                 return 0;
 272         }
 273
 274         list_next = lock->l_pending_chain.next;
 275         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 276                 /* Removing the head of the list, adjust timer. */
 277                 if (list_next == &waiting_locks_list) {
 278                         /* No more, just cancel. */
 279                         del_timer(&waiting_locks_timer);
 280                 } else {
 281                         struct ldlm_lock *next;
 282                         next = list_entry(list_next, struct ldlm_lock,
 283                                           l_pending_chain);
 284                         mod_timer(&waiting_locks_timer,
 285                                   round_timeout(next->l_callback_timeout));
 286                 }
 287         }
 288
 289         spin_lock_bh(&expired_lock_thread.elt_lock);
 290         list_del_init(&lock->l_pending_chain);
 291         spin_unlock_bh(&expired_lock_thread.elt_lock);
 292
 293         spin_unlock_bh(&waiting_locks_spinlock);
 294         LDLM_DEBUG(lock, "removed");
 295         return 1;
 296 }
 297
 298 #else /* !__KERNEL__ */
 299
 300 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 301 {
 302         RETURN(1);
 303 }
 304
 305 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 306 {
 307         RETURN(0);
 308 }
 309
 310 #endif /* __KERNEL__ */
 311
 312 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, char *ast_type)
 313 {
 314         const struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 315         char str[PTL_NALFMT_SIZE];
 316
 317         CERROR("%s AST failed (%d) for res "LPU64"/"LPU64
 318                ", mode %s: evicting client %s@%s NID "LPX64" (%s)\n",
 319                ast_type, rc,
 320                lock->l_resource->lr_name.name[0],
 321                lock->l_resource->lr_name.name[1],
 322                ldlm_lockname[lock->l_granted_mode],
 323                lock->l_export->exp_client_uuid.uuid,
 324                conn->c_remote_uuid.uuid, conn->c_peer.peer_nid,
 325                portals_nid2str(conn->c_peer.peer_ni->pni_number,
 326                                conn->c_peer.peer_nid, str));
 327         ptlrpc_fail_export(lock->l_export);
 328 }
 329
 330 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 331                              struct ldlm_lock_desc *desc,
 332                              void *data, int flag)
 333 {
 334         struct ldlm_request *body;
 335         struct ptlrpc_request *req;
 336         int rc = 0, size = sizeof(*body);
 337         ENTRY;
 338
 339         if (flag == LDLM_CB_CANCELING) {
 340                 /* Don't need to do anything here. */
 341                 RETURN(0);
 342         }
 343
 344         LASSERT(lock);
 345
 346         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 347         if (lock->l_granted_mode != lock->l_req_mode) {
 348                 /* this blocking AST will be communicated as part of the
 349                  * completion AST instead */
 350                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 351                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");                 RETURN(0);
 352         }
 353
 354         if (lock->l_destroyed) {
 355                 /* What's the point? */
 356                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 357                 RETURN(0);
 358         }
 359
 360 #if 0
 361         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 362                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 363                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 364                 RETURN(-ETIMEDOUT);
 365         }
 366 #endif
 367
 368         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 369                               LDLM_BL_CALLBACK, 1, &size, NULL);
 370         if (req == NULL) {
 371                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 372                 RETURN(-ENOMEM);
 373         }
 374
 375         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 376         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 377                sizeof(body->lock_handle1));
 378         memcpy(&body->lock_desc, desc, sizeof(*desc));
 379         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 380
 381         LDLM_DEBUG(lock, "server preparing blocking AST");
 382         req->rq_replen = lustre_msg_size(0, NULL);
 383
 384         if (lock->l_granted_mode == lock->l_req_mode)
 385                 ldlm_add_waiting_lock(lock);
 386         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 387
 388         req->rq_send_state = LUSTRE_IMP_FULL;
 389         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 390         rc = ptlrpc_queue_wait(req);
 391         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 392                 LASSERT(lock->l_export);
 393                 if (lock->l_export->exp_libclient) {
 394                         CDEBUG(D_HA, "BLOCKING AST to liblustre client (nid "
 395                                LPU64") timeout, simply cancel lock 0x%p\n",
 396                                req->rq_peer.peer_nid, lock);
 397                         ldlm_lock_cancel(lock);
 398                         rc = -ERESTART;
 399                 } else {
 400                         ldlm_del_waiting_lock(lock);
 401                         ldlm_failed_ast(lock, rc, "blocking");
 402                 }
 403         } else if (rc) {
 404                 if (rc == -EINVAL)
 405                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 406                                "from blocking AST for lock %p--normal race\n",
 407                                req->rq_peer.peer_nid,
 408                                req->rq_repmsg->status, lock);
 409                 else if (rc == -ENOTCONN)
 410                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 411                                "from blocking AST for lock %p--this client was "
 412                                "probably rebooted while it held a lock, nothing"
 413                                " serious\n",req->rq_peer.peer_nid,
 414                                req->rq_repmsg->status, lock);
 415                 else
 416                         CDEBUG(D_ERROR, "client (nid "LPU64") returned %d "
 417                                "from blocking AST for lock %p\n",
 418                                req->rq_peer.peer_nid,
 419                                (req->rq_repmsg != NULL)?
 420                                req->rq_repmsg->status : 0,
 421                                lock);
 422                 LDLM_DEBUG(lock, "client sent rc %d rq_status %d from blocking "
 423                            "AST", rc, req->rq_status);
 424                 ldlm_lock_cancel(lock);
 425                 /* Server-side AST functions are called from ldlm_reprocess_all,
 426                  * which needs to be told to please restart its reprocessing. */
 427                 rc = -ERESTART;
 428         }
 429
 430         ptlrpc_req_finished(req);
 431
 432         RETURN(rc);
 433 }
 434
 435 /* XXX copied from ptlrpc/service.c */
 436 static long timeval_sub(struct timeval *large, struct timeval *small)
 437 {
 438         return (large->tv_sec - small->tv_sec) * 1000000 +
 439                 (large->tv_usec - small->tv_usec);
 440 }
 441
 442 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 443 {
 444         struct ldlm_request *body;
 445         struct ptlrpc_request *req;
 446         struct timeval granted_time;
 447         long total_enqueue_wait;
 448         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 449         ENTRY;
 450
 451         LASSERT(lock != NULL);
 452
 453         do_gettimeofday(&granted_time);
 454         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 455
 456         if (total_enqueue_wait / 1000000 > obd_timeout)
 457                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 458
 459         if (lock->l_resource->lr_lvb_len) {
 460                 buffers = 2;
 461                 size[1] = lock->l_resource->lr_lvb_len;
 462         }
 463
 464         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 465                               LDLM_CP_CALLBACK, buffers, size, NULL);
 466         if (req == NULL)
 467                 RETURN(-ENOMEM);
 468
 469         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 470         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 471                sizeof(body->lock_handle1));
 472         body->lock_flags = flags;
 473         ldlm_lock2desc(lock, &body->lock_desc);
 474
 475         if (buffers == 2) {
 476                 void *lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 477                                            lock->l_resource->lr_lvb_len);
 478                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 479                        lock->l_resource->lr_lvb_len);
 480         }
 481
 482         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 483                    total_enqueue_wait);
 484         req->rq_replen = lustre_msg_size(0, NULL);
 485
 486         req->rq_send_state = LUSTRE_IMP_FULL;
 487         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 488
 489         /* We only send real blocking ASTs after the lock is granted */
 490         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 491         if (lock->l_flags & LDLM_FL_AST_SENT) {
 492                 body->lock_flags |= LDLM_FL_AST_SENT;
 493                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 494         }
 495         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 496
 497         rc = ptlrpc_queue_wait(req);
 498         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 499                 ldlm_del_waiting_lock(lock);
 500                 ldlm_failed_ast(lock, rc, "completion");
 501         } else if (rc) {
 502                 LDLM_ERROR(lock, "client sent rc %d rq_status %d from "
 503                            "completion AST\n", rc, req->rq_status);
 504                 ldlm_lock_cancel(lock);
 505                 /* Server-side AST functions are called from ldlm_reprocess_all,
 506                  * which needs to be told to please restart its reprocessing. */
 507                 rc = -ERESTART;
 508         }
 509         ptlrpc_req_finished(req);
 510
 511         RETURN(rc);
 512 }
 513
 514 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 515 {
 516         struct ldlm_resource *res = lock->l_resource;
 517         struct ldlm_request *body;
 518         struct ptlrpc_request *req;
 519         int rc = 0, size = sizeof(*body);
 520         ENTRY;
 521
 522         LASSERT(lock != NULL);
 523
 524         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 525                               LDLM_GL_CALLBACK, 1, &size, NULL);
 526         if (req == NULL)
 527                 RETURN(-ENOMEM);
 528
 529         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 530         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 531                sizeof(body->lock_handle1));
 532         ldlm_lock2desc(lock, &body->lock_desc);
 533
 534         size = lock->l_resource->lr_lvb_len;
 535         req->rq_replen = lustre_msg_size(1, &size);
 536
 537         req->rq_send_state = LUSTRE_IMP_FULL;
 538         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 539
 540         rc = ptlrpc_queue_wait(req);
 541         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 542                 ldlm_del_waiting_lock(lock);
 543                 ldlm_failed_ast(lock, rc, "glimpse");
 544         } else if (rc) {
 545                 LDLM_ERROR(lock, "client sent rc %d rq_status %d from "
 546                            "completion AST\n", rc, req->rq_status);
 547                 ldlm_lock_cancel(lock);
 548         } else {
 549                 rc = res->lr_namespace->ns_lvbo->lvbo_update(res,
 550                                                              req->rq_repmsg, 0);
 551         }
 552         ptlrpc_req_finished(req);
 553         RETURN(rc);
 554 }
 555
 556 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 557                         ldlm_completion_callback completion_callback,
 558                         ldlm_blocking_callback blocking_callback,
 559                         ldlm_glimpse_callback glimpse_callback)
 560 {
 561         struct obd_device *obddev = req->rq_export->exp_obd;
 562         struct ldlm_reply *dlm_rep;
 563         struct ldlm_request *dlm_req;
 564         int rc, size[2] = {sizeof(*dlm_rep)};
 565         __u32 flags;
 566         ldlm_error_t err;
 567         struct ldlm_lock *lock = NULL;
 568         void *cookie = NULL;
 569         ENTRY;
 570
 571         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 572
 573         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 574                                       lustre_swab_ldlm_request);
 575         if (dlm_req == NULL) {
 576                 CERROR ("Can't unpack dlm_req\n");
 577                 RETURN (-EFAULT);
 578         }
 579
 580         flags = dlm_req->lock_flags;
 581
 582         /* The lock's callback data might be set in the policy function */
 583         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 584                                 dlm_req->lock_desc.l_resource.lr_name,
 585                                 dlm_req->lock_desc.l_resource.lr_type,
 586                                 dlm_req->lock_desc.l_req_mode,
 587                                 blocking_callback, completion_callback,
 588                                 glimpse_callback, NULL, 0);
 589         if (!lock)
 590                 GOTO(out, err = -ENOMEM);
 591
 592         do_gettimeofday(&lock->l_enqueued_time);
 593         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 594                sizeof(lock->l_remote_handle));
 595         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 596
 597         LASSERT(req->rq_export);
 598         lock->l_export = class_export_get(req->rq_export);
 599         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 600         list_add(&lock->l_export_chain,
 601                  &lock->l_export->exp_ldlm_data.led_held_locks);
 602         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 603
 604         if (flags & LDLM_FL_HAS_INTENT) {
 605                 /* In this case, the reply buffer is allocated deep in
 606                  * local_lock_enqueue by the policy function. */
 607                 cookie = req;
 608         } else {
 609                 int buffers = 1;
 610                 if (lock->l_resource->lr_lvb_len) {
 611                         size[1] = lock->l_resource->lr_lvb_len;
 612                         buffers = 2;
 613                 }
 614
 615                 rc = lustre_pack_reply(req, buffers, size, NULL);
 616                 if (rc)
 617                         RETURN(rc);
 618         }
 619
 620         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 621                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 622                        sizeof(ldlm_policy_data_t));
 623
 624         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 625         if (err)
 626                 GOTO(out, err);
 627
 628         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 629         dlm_rep->lock_flags = flags;
 630
 631         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 632         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 633
 634         /* We never send a blocking AST until the lock is granted, but
 635          * we can tell it right now */
 636         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 637         if (lock->l_flags & LDLM_FL_AST_SENT) {
 638                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 639                 if (lock->l_granted_mode == lock->l_req_mode)
 640                         ldlm_add_waiting_lock(lock);
 641         }
 642         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 643
 644         EXIT;
 645  out:
 646         if (lock->l_resource->lr_lvb_len > 0) {
 647                 void *lvb = lustre_msg_buf(req->rq_repmsg, 1,
 648                                            lock->l_resource->lr_lvb_len);
 649                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 650                        lock->l_resource->lr_lvb_len);
 651         }
 652         req->rq_status = err;
 653
 654         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 655          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 656         if (lock) {
 657                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 658                            "(err=%d)", err);
 659                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 660                         ldlm_reprocess_all(lock->l_resource);
 661                 LDLM_LOCK_PUT(lock);
 662         }
 663         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p)", lock);
 664
 665         return 0;
 666 }
 667
 668 int ldlm_handle_convert(struct ptlrpc_request *req)
 669 {
 670         struct ldlm_request *dlm_req;
 671         struct ldlm_reply *dlm_rep;
 672         struct ldlm_lock *lock;
 673         int rc, size = sizeof(*dlm_rep);
 674         ENTRY;
 675
 676         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 677                                       lustre_swab_ldlm_request);
 678         if (dlm_req == NULL) {
 679                 CERROR ("Can't unpack dlm_req\n");
 680                 RETURN (-EFAULT);
 681         }
 682
 683         rc = lustre_pack_reply(req, 1, &size, NULL);
 684         if (rc) {
 685                 CERROR("out of memory\n");
 686                 RETURN(-ENOMEM);
 687         }
 688         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 689         dlm_rep->lock_flags = dlm_req->lock_flags;
 690
 691         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 692         if (!lock) {
 693                 req->rq_status = EINVAL;
 694         } else {
 695                 LDLM_DEBUG(lock, "server-side convert handler START");
 696                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 697                                   &dlm_rep->lock_flags);
 698                 if (ldlm_del_waiting_lock(lock))
 699                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 700                 req->rq_status = 0;
 701         }
 702
 703         if (lock) {
 704                 ldlm_reprocess_all(lock->l_resource);
 705                 LDLM_DEBUG(lock, "server-side convert handler END");
 706                 LDLM_LOCK_PUT(lock);
 707         } else
 708                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 709
 710         RETURN(0);
 711 }
 712
 713 int ldlm_handle_cancel(struct ptlrpc_request *req)
 714 {
 715         struct ldlm_request *dlm_req;
 716         struct ldlm_lock *lock;
 717         struct ldlm_resource *res;
 718         char str[PTL_NALFMT_SIZE];
 719         int rc;
 720         ENTRY;
 721
 722         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 723                                       lustre_swab_ldlm_request);
 724         if (dlm_req == NULL) {
 725                 CERROR("bad request buffer for cancel\n");
 726                 RETURN(-EFAULT);
 727         }
 728
 729         rc = lustre_pack_reply(req, 0, NULL, NULL);
 730         if (rc) {
 731                 CERROR("out of memory\n");
 732                 RETURN(-ENOMEM);
 733         }
 734
 735         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 736         if (!lock) {
 737                 CERROR("received cancel for unknown lock cookie "LPX64
 738                        " from client %s nid "LPX64" (%s)\n",
 739                        dlm_req->lock_handle1.cookie,
 740                        req->rq_export->exp_client_uuid.uuid,
 741                        req->rq_peer.peer_nid,
 742                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 743                                        req->rq_peer.peer_nid, str));
 744                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 745                                   "(cookie "LPU64")",
 746                                   dlm_req->lock_handle1.cookie);
 747                 req->rq_status = ESTALE;
 748         } else {
 749                 LDLM_DEBUG(lock, "server-side cancel handler START");
 750                 res = lock->l_resource;
 751                 if (res && res->lr_namespace->ns_lvbo &&
 752                     res->lr_namespace->ns_lvbo->lvbo_update) {
 753                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 754                                 (res, NULL, 0);
 755                                 //(res, req->rq_reqmsg, 1);
 756                 }
 757
 758                 ldlm_lock_cancel(lock);
 759                 if (ldlm_del_waiting_lock(lock))
 760                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 761                 req->rq_status = rc;
 762         }
 763
 764         if (ptlrpc_reply(req) != 0)
 765                 LBUG();
 766
 767         if (lock) {
 768                 ldlm_reprocess_all(lock->l_resource);
 769                 LDLM_DEBUG(lock, "server-side cancel handler END");
 770                 LDLM_LOCK_PUT(lock);
 771         }
 772
 773         RETURN(0);
 774 }
 775
 776 static void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 777                                     struct ldlm_lock_desc *ld,
 778                                     struct ldlm_lock *lock)
 779 {
 780         int do_ast;
 781         ENTRY;
 782
 783         l_lock(&ns->ns_lock);
 784         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 785
 786         lock->l_flags |= LDLM_FL_CBPENDING;
 787         do_ast = (!lock->l_readers && !lock->l_writers);
 788
 789         if (do_ast) {
 790                 LDLM_DEBUG(lock, "already unused, calling "
 791                            "callback (%p)", lock->l_blocking_ast);
 792                 if (lock->l_blocking_ast != NULL) {
 793                         l_unlock(&ns->ns_lock);
 794                         l_check_no_ns_lock(ns);
 795                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 796                                              LDLM_CB_BLOCKING);
 797                         l_lock(&ns->ns_lock);
 798                 }
 799         } else {
 800                 LDLM_DEBUG(lock, "Lock still has references, will be"
 801                            " cancelled later");
 802         }
 803
 804         LDLM_DEBUG(lock, "client blocking callback handler END");
 805         l_unlock(&ns->ns_lock);
 806         LDLM_LOCK_PUT(lock);
 807         EXIT;
 808 }
 809
 810 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 811                                     struct ldlm_namespace *ns,
 812                                     struct ldlm_request *dlm_req,
 813                                     struct ldlm_lock *lock)
 814 {
 815         LIST_HEAD(ast_list);
 816         ENTRY;
 817
 818         l_lock(&ns->ns_lock);
 819         LDLM_DEBUG(lock, "client completion callback handler START");
 820
 821         /* If we receive the completion AST before the actual enqueue returned,
 822          * then we might need to switch lock modes, resources, or extents. */
 823         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 824                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 825                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 826         }
 827         if (lock->l_resource->lr_type != LDLM_PLAIN)
 828                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 829                        sizeof(lock->l_policy_data));
 830
 831         ldlm_resource_unlink_lock(lock);
 832         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 833                    &lock->l_resource->lr_name,
 834                    sizeof(lock->l_resource->lr_name)) != 0) {
 835                 ldlm_lock_change_resource(ns, lock,
 836                                          dlm_req->lock_desc.l_resource.lr_name);
 837                 LDLM_DEBUG(lock, "completion AST, new resource");
 838         }
 839
 840         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 841                 lock->l_flags |= LDLM_FL_CBPENDING;
 842                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 843         }
 844
 845         if (lock->l_lvb_len) {
 846                 void *lvb;
 847                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 848                                          lock->l_lvb_swabber);
 849                 if (lvb == NULL) {
 850                         LDLM_ERROR(lock, "completion AST did not contain "
 851                                    "expected LVB!");
 852                 } else {
 853                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 854                 }
 855         }
 856
 857         lock->l_resource->lr_tmp = &ast_list;
 858         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 859         lock->l_resource->lr_tmp = NULL;
 860         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 861         l_unlock(&ns->ns_lock);
 862         LDLM_LOCK_PUT(lock);
 863
 864         ldlm_run_ast_work(ns, &ast_list);
 865
 866         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 867                           lock);
 868         EXIT;
 869 }
 870
 871 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 872                                     struct ldlm_namespace *ns,
 873                                     struct ldlm_request *dlm_req,
 874                                     struct ldlm_lock *lock)
 875 {
 876         ENTRY;
 877
 878         l_lock(&ns->ns_lock);
 879         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 880
 881         if (lock->l_glimpse_ast != NULL) {
 882                 l_unlock(&ns->ns_lock);
 883                 l_check_no_ns_lock(ns);
 884                 lock->l_glimpse_ast(lock, req);
 885                 l_lock(&ns->ns_lock);
 886         }
 887
 888         if (lock->l_granted_mode == LCK_PW &&
 889             !lock->l_readers && !lock->l_writers &&
 890             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 891                 l_unlock(&ns->ns_lock);
 892                 ldlm_handle_bl_callback(ns, NULL, lock);
 893                 EXIT;
 894                 return;
 895         }
 896
 897         l_unlock(&ns->ns_lock);
 898         LDLM_LOCK_PUT(lock);
 899         EXIT;
 900 }
 901
 902 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 903 {
 904         req->rq_status = rc;
 905         rc = lustre_pack_reply(req, 0, NULL, NULL);
 906         if (rc)
 907                 return rc;
 908         return ptlrpc_reply(req);
 909 }
 910
 911 #ifdef __KERNEL__
 912 static int ldlm_bl_to_thread(struct ldlm_state *ldlm, struct ldlm_namespace *ns,
 913                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 914 {
 915         struct ldlm_bl_pool *blp = ldlm->ldlm_bl_pool;
 916         struct ldlm_bl_work_item *blwi;
 917         ENTRY;
 918
 919         OBD_ALLOC(blwi, sizeof(*blwi));
 920         if (blwi == NULL)
 921                 RETURN(-ENOMEM);
 922
 923         blwi->blwi_ns = ns;
 924         blwi->blwi_ld = *ld;
 925         blwi->blwi_lock = lock;
 926
 927         spin_lock(&blp->blp_lock);
 928         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
 929         wake_up(&blp->blp_waitq);
 930         spin_unlock(&blp->blp_lock);
 931
 932         RETURN(0);
 933 }
 934 #endif
 935
 936 static int ldlm_callback_handler(struct ptlrpc_request *req)
 937 {
 938         struct ldlm_namespace *ns;
 939         struct ldlm_request *dlm_req;
 940         struct ldlm_lock *lock;
 941         char str[PTL_NALFMT_SIZE];
 942         int rc;
 943         ENTRY;
 944
 945         /* Requests arrive in sender's byte order.  The ptlrpc service
 946          * handler has already checked and, if necessary, byte-swapped the
 947          * incoming request message body, but I am responsible for the
 948          * message buffers. */
 949
 950         if (req->rq_export == NULL) {
 951                 struct ldlm_request *dlm_req;
 952
 953                 CDEBUG(D_RPCTRACE, "operation %d from nid "LPX64" (%s) with bad "
 954                        "export cookie "LPX64" (ptl req %d/rep %d); this is "
 955                        "normal if this node rebooted with a lock held\n",
 956                        req->rq_reqmsg->opc, req->rq_peer.peer_nid,
 957                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 958                                        req->rq_peer.peer_nid, str),
 959                        req->rq_reqmsg->handle.cookie,
 960                        req->rq_request_portal, req->rq_reply_portal);
 961
 962                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 963                                              lustre_swab_ldlm_request);
 964                 if (dlm_req != NULL)
 965                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
 966                                dlm_req->lock_handle1.cookie);
 967
 968                 ldlm_callback_reply(req, -ENOTCONN);
 969                 RETURN(0);
 970         }
 971
 972         if (req->rq_reqmsg->opc == LDLM_BL_CALLBACK) {
 973                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 974         } else if (req->rq_reqmsg->opc == LDLM_CP_CALLBACK) {
 975                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
 976         } else if (req->rq_reqmsg->opc == LDLM_GL_CALLBACK) {
 977                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
 978         } else if (req->rq_reqmsg->opc == OBD_LOG_CANCEL) {
 979                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
 980         } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CREATE) {
 981                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
 982         } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_NEXT_BLOCK) {
 983                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
 984         } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_READ_HEADER) {
 985                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
 986         } else if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CLOSE) {
 987                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
 988         } else {
 989                 ldlm_callback_reply(req, -EPROTO);
 990                 RETURN(0);
 991         }
 992
 993         LASSERT(req->rq_export != NULL);
 994         LASSERT(req->rq_export->exp_obd != NULL);
 995
 996         /* FIXME - how to send reply */
 997         if (req->rq_reqmsg->opc == OBD_LOG_CANCEL) {
 998                 int rc = llog_origin_handle_cancel(req);
 999                 ldlm_callback_reply(req, rc);
1000                 RETURN(0);
1001         }
1002         if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CREATE) {
1003                 int rc = llog_origin_handle_create(req);
1004                 req->rq_status = rc;
1005                 ptlrpc_reply(req);
1006                 RETURN(0);
1007         }
1008         if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_NEXT_BLOCK) {
1009                 int rc = llog_origin_handle_next_block(req);
1010                 req->rq_status = rc;
1011                 ptlrpc_reply(req);
1012                 RETURN(0);
1013         }
1014         if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_READ_HEADER) {
1015                 int rc = llog_origin_handle_read_header(req);
1016                 req->rq_status = rc;
1017                 ptlrpc_reply(req);
1018                 RETURN(0);
1019         }
1020         if (req->rq_reqmsg->opc == LLOG_ORIGIN_HANDLE_CLOSE) {
1021                 int rc = llog_origin_handle_close(req);
1022                 ldlm_callback_reply(req, rc);
1023                 RETURN(0);
1024         }
1025
1026         ns = req->rq_export->exp_obd->obd_namespace;
1027         LASSERT(ns != NULL);
1028
1029         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1030                                       lustre_swab_ldlm_request);
1031         if (dlm_req == NULL) {
1032                 CERROR ("can't unpack dlm_req\n");
1033                 ldlm_callback_reply (req, -EPROTO);
1034                 RETURN (0);
1035         }
1036
1037         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1038         if (!lock) {
1039                 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
1040                        dlm_req->lock_handle1.cookie);
1041                 ldlm_callback_reply(req, -EINVAL);
1042                 RETURN(0);
1043         }
1044
1045         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1046         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1047
1048         /* We want the ost thread to get this reply so that it can respond
1049          * to ost requests (write cache writeback) that might be triggered
1050          * in the callback.
1051          *
1052          * But we'd also like to be able to indicate in the reply that we're
1053          * cancelling right now, because it's unused, or have an intent result
1054          * in the reply, so we might have to push the responsibility for sending
1055          * the reply down into the AST handlers, alas. */
1056         if (req->rq_reqmsg->opc == LDLM_CP_CALLBACK)
1057                 ldlm_callback_reply(req, 0);
1058
1059         switch (req->rq_reqmsg->opc) {
1060         case LDLM_BL_CALLBACK:
1061                 CDEBUG(D_INODE, "blocking ast\n");
1062 #ifdef __KERNEL__
1063                 rc = ldlm_bl_to_thread(ldlm, ns, &dlm_req->lock_desc, lock);
1064                 ldlm_callback_reply(req, rc);
1065 #else
1066                 rc = 0;
1067                 ldlm_callback_reply(req, rc);
1068                 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1069 #endif
1070                 break;
1071         case LDLM_CP_CALLBACK:
1072                 CDEBUG(D_INODE, "completion ast\n");
1073                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1074                 break;
1075         case LDLM_GL_CALLBACK:
1076                 CDEBUG(D_INODE, "glimpse ast\n");
1077                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1078                 break;
1079         default:
1080                 LBUG();                         /* checked above */
1081         }
1082
1083         RETURN(0);
1084 }
1085
1086 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1087 {
1088         int rc;
1089         ENTRY;
1090
1091         /* Requests arrive in sender's byte order.  The ptlrpc service
1092          * handler has already checked and, if necessary, byte-swapped the
1093          * incoming request message body, but I am responsible for the
1094          * message buffers. */
1095
1096         if (req->rq_export == NULL) {
1097                 struct ldlm_request *dlm_req;
1098                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
1099                        req->rq_reqmsg->opc, req->rq_request_portal,
1100                        req->rq_reply_portal);
1101                 CERROR("--> export cookie: "LPX64"\n",
1102                        req->rq_reqmsg->handle.cookie);
1103                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1104                                              lustre_swab_ldlm_request);
1105                 if (dlm_req != NULL)
1106                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1107                 RETURN(-ENOTCONN);
1108         }
1109
1110         switch (req->rq_reqmsg->opc) {
1111
1112         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1113         case LDLM_CANCEL:
1114                 CDEBUG(D_INODE, "cancel\n");
1115                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1116                 rc = ldlm_handle_cancel(req);
1117                 if (rc)
1118                         break;
1119                 RETURN(0);
1120
1121         default:
1122                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1123                 RETURN(-EINVAL);
1124         }
1125
1126         RETURN(0);
1127 }
1128
1129 #ifdef __KERNEL__
1130 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1131 {
1132         struct ldlm_bl_work_item *blwi = NULL;
1133
1134         spin_lock(&blp->blp_lock);
1135         if (!list_empty(&blp->blp_list)) {
1136                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1137                                   blwi_entry);
1138                 list_del(&blwi->blwi_entry);
1139         }
1140         spin_unlock(&blp->blp_lock);
1141
1142         return blwi;
1143 }
1144
1145 struct ldlm_bl_thread_data {
1146         int                     bltd_num;
1147         struct ldlm_bl_pool     *bltd_blp;
1148 };
1149
1150 static int ldlm_bl_thread_main(void *arg)
1151 {
1152         struct ldlm_bl_thread_data *bltd = arg;
1153         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1154         unsigned long flags;
1155         ENTRY;
1156
1157         /* XXX boiler-plate */
1158         {
1159                 char name[sizeof(current->comm)];
1160                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1161                          bltd->bltd_num);
1162                 kportal_daemonize(name);
1163         }
1164         SIGNAL_MASK_LOCK(current, flags);
1165         sigfillset(&current->blocked);
1166         RECALC_SIGPENDING;
1167         SIGNAL_MASK_UNLOCK(current, flags);
1168
1169         atomic_inc(&blp->blp_num_threads);
1170         complete(&blp->blp_comp);
1171
1172         while(1) {
1173                 struct l_wait_info lwi = { 0 };
1174                 struct ldlm_bl_work_item *blwi = NULL;
1175
1176                 l_wait_event_exclusive(blp->blp_waitq,
1177                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1178                                        &lwi);
1179
1180                 if (blwi->blwi_ns == NULL)
1181                         break;
1182
1183                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1184                                         blwi->blwi_lock);
1185                 OBD_FREE(blwi, sizeof(*blwi));
1186         }
1187
1188         atomic_dec(&blp->blp_num_threads);
1189         complete(&blp->blp_comp);
1190         RETURN(0);
1191 }
1192
1193 #endif
1194
1195 static int ldlm_setup(void);
1196 static int ldlm_cleanup(int force);
1197
1198 int ldlm_get_ref(void)
1199 {
1200         int rc = 0;
1201         down(&ldlm_ref_sem);
1202         if (++ldlm_refcount == 1) {
1203                 rc = ldlm_setup();
1204                 if (rc)
1205                         ldlm_refcount--;
1206         }
1207         up(&ldlm_ref_sem);
1208
1209         RETURN(rc);
1210 }
1211
1212 void ldlm_put_ref(int force)
1213 {
1214         down(&ldlm_ref_sem);
1215         if (ldlm_refcount == 1) {
1216                 int rc = ldlm_cleanup(force);
1217                 if (rc)
1218                         CERROR("ldlm_cleanup failed: %d\n", rc);
1219                 else
1220                         ldlm_refcount--;
1221         } else {
1222                 ldlm_refcount--;
1223         }
1224         up(&ldlm_ref_sem);
1225
1226         EXIT;
1227 }
1228
1229 static int ldlm_setup(void)
1230 {
1231         struct ldlm_bl_pool *blp;
1232         int rc = 0;
1233 #ifdef __KERNEL__
1234         int i;
1235 #endif
1236         ENTRY;
1237
1238         if (ldlm != NULL)
1239                 RETURN(-EALREADY);
1240
1241         OBD_ALLOC(ldlm, sizeof(*ldlm));
1242         if (ldlm == NULL)
1243                 RETURN(-ENOMEM);
1244
1245 #ifdef __KERNEL__
1246         rc = ldlm_proc_setup();
1247         if (rc != 0)
1248                 GOTO(out_free, rc);
1249 #endif
1250
1251         ldlm->ldlm_cb_service =
1252                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1253                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1254                                 ldlm_callback_handler, "ldlm_cbd",
1255                                 ldlm_svc_proc_dir);
1256
1257         if (!ldlm->ldlm_cb_service) {
1258                 CERROR("failed to start service\n");
1259                 GOTO(out_proc, rc = -ENOMEM);
1260         }
1261
1262         ldlm->ldlm_cancel_service =
1263                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1264                                 LDLM_CANCEL_REQUEST_PORTAL,
1265                                 LDLM_CANCEL_REPLY_PORTAL,
1266                                 ldlm_cancel_handler, "ldlm_canceld",
1267                                 ldlm_svc_proc_dir);
1268
1269         if (!ldlm->ldlm_cancel_service) {
1270                 CERROR("failed to start service\n");
1271                 GOTO(out_proc, rc = -ENOMEM);
1272         }
1273
1274         OBD_ALLOC(blp, sizeof(*blp));
1275         if (blp == NULL)
1276                 GOTO(out_proc, rc = -ENOMEM);
1277         ldlm->ldlm_bl_pool = blp;
1278
1279         atomic_set(&blp->blp_num_threads, 0);
1280         init_waitqueue_head(&blp->blp_waitq);
1281         spin_lock_init(&blp->blp_lock);
1282
1283         INIT_LIST_HEAD(&blp->blp_list);
1284
1285 #ifdef __KERNEL__
1286         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1287                 struct ldlm_bl_thread_data bltd = {
1288                         .bltd_num = i,
1289                         .bltd_blp = blp,
1290                 };
1291                 init_completion(&blp->blp_comp);
1292                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1293                 if (rc < 0) {
1294                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1295                         LBUG();
1296                         GOTO(out_thread, rc);
1297                 }
1298                 wait_for_completion(&blp->blp_comp);
1299         }
1300
1301         rc = ptlrpc_start_n_threads(NULL, ldlm->ldlm_cancel_service,
1302                                     LDLM_NUM_THREADS, "ldlm_cn");
1303         if (rc) {
1304                 LBUG();
1305                 GOTO(out_thread, rc);
1306         }
1307
1308         rc = ptlrpc_start_n_threads(NULL, ldlm->ldlm_cb_service,
1309                                     LDLM_NUM_THREADS, "ldlm_cb");
1310         if (rc) {
1311                 LBUG();
1312                 GOTO(out_thread, rc);
1313         }
1314
1315         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1316         spin_lock_init(&expired_lock_thread.elt_lock);
1317         expired_lock_thread.elt_state = ELT_STOPPED;
1318         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1319
1320         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1321         if (rc < 0) {
1322                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1323                 GOTO(out_thread, rc);
1324         }
1325
1326         wait_event(expired_lock_thread.elt_waitq,
1327                    expired_lock_thread.elt_state == ELT_READY);
1328
1329         INIT_LIST_HEAD(&waiting_locks_list);
1330         spin_lock_init(&waiting_locks_spinlock);
1331         waiting_locks_timer.function = waiting_locks_callback;
1332         waiting_locks_timer.data = 0;
1333         init_timer(&waiting_locks_timer);
1334 #endif
1335
1336         RETURN(0);
1337
1338 #ifdef __KERNEL__
1339  out_thread:
1340         ptlrpc_unregister_service(ldlm->ldlm_cancel_service);
1341         ptlrpc_unregister_service(ldlm->ldlm_cb_service);
1342 #endif
1343
1344  out_proc:
1345 #ifdef __KERNEL__
1346         ldlm_proc_cleanup();
1347  out_free:
1348 #endif
1349         OBD_FREE(ldlm, sizeof(*ldlm));
1350         ldlm = NULL;
1351         return rc;
1352 }
1353
1354 static int ldlm_cleanup(int force)
1355 {
1356 #ifdef __KERNEL__
1357         struct ldlm_bl_pool *blp = ldlm->ldlm_bl_pool;
1358 #endif
1359         ENTRY;
1360
1361         if (!list_empty(&ldlm_namespace_list)) {
1362                 CERROR("ldlm still has namespaces; clean these up first.\n");
1363                 ldlm_dump_all_namespaces();
1364                 RETURN(-EBUSY);
1365         }
1366
1367 #ifdef __KERNEL__
1368         while (atomic_read(&blp->blp_num_threads) > 0) {
1369                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1370
1371                 init_completion(&blp->blp_comp);
1372
1373                 spin_lock(&blp->blp_lock);
1374                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1375                 wake_up(&blp->blp_waitq);
1376                 spin_unlock(&blp->blp_lock);
1377
1378                 wait_for_completion(&blp->blp_comp);
1379         }
1380         OBD_FREE(blp, sizeof(*blp));
1381
1382         ptlrpc_stop_all_threads(ldlm->ldlm_cb_service);
1383         ptlrpc_unregister_service(ldlm->ldlm_cb_service);
1384         ptlrpc_stop_all_threads(ldlm->ldlm_cancel_service);
1385         ptlrpc_unregister_service(ldlm->ldlm_cancel_service);
1386         ldlm_proc_cleanup();
1387
1388         expired_lock_thread.elt_state = ELT_TERMINATE;
1389         wake_up(&expired_lock_thread.elt_waitq);
1390         wait_event(expired_lock_thread.elt_waitq,
1391                    expired_lock_thread.elt_state == ELT_STOPPED);
1392
1393 #endif
1394
1395         OBD_FREE(ldlm, sizeof(*ldlm));
1396         ldlm = NULL;
1397
1398         RETURN(0);
1399 }
1400
1401 int __init ldlm_init(void)
1402 {
1403         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1404                                                sizeof(struct ldlm_resource), 0,
1405                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1406         if (ldlm_resource_slab == NULL)
1407                 return -ENOMEM;
1408
1409         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1410                                            sizeof(struct ldlm_lock), 0,
1411                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1412         if (ldlm_lock_slab == NULL) {
1413                 kmem_cache_destroy(ldlm_resource_slab);
1414                 return -ENOMEM;
1415         }
1416
1417         l_lock_init(&ldlm_handle_lock);
1418
1419         return 0;
1420 }
1421
1422 void __exit ldlm_exit(void)
1423 {
1424         if ( ldlm_refcount )
1425                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1426         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
1427                 CERROR("couldn't free ldlm resource slab\n");
1428         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
1429                 CERROR("couldn't free ldlm lock slab\n");
1430 }
1431
1432 /* ldlm_flock.c */
1433 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1434
1435 /* ldlm_extent.c */
1436 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1437
1438 /* ldlm_lock.c */
1439 EXPORT_SYMBOL(ldlm_get_processing_policy);
1440 EXPORT_SYMBOL(ldlm_lock2desc);
1441 EXPORT_SYMBOL(ldlm_register_intent);
1442 EXPORT_SYMBOL(ldlm_lockname);
1443 EXPORT_SYMBOL(ldlm_typename);
1444 EXPORT_SYMBOL(ldlm_lock2handle);
1445 EXPORT_SYMBOL(__ldlm_handle2lock);
1446 EXPORT_SYMBOL(ldlm_lock_get);
1447 EXPORT_SYMBOL(ldlm_lock_put);
1448 EXPORT_SYMBOL(ldlm_lock_match);
1449 EXPORT_SYMBOL(ldlm_lock_cancel);
1450 EXPORT_SYMBOL(ldlm_lock_addref);
1451 EXPORT_SYMBOL(ldlm_lock_decref);
1452 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1453 EXPORT_SYMBOL(ldlm_lock_change_resource);
1454 EXPORT_SYMBOL(ldlm_lock_set_data);
1455 EXPORT_SYMBOL(ldlm_it2str);
1456 EXPORT_SYMBOL(ldlm_lock_dump);
1457 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1458 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1459 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1460 EXPORT_SYMBOL(ldlm_lock_allow_match);
1461
1462 /* ldlm_request.c */
1463 EXPORT_SYMBOL(ldlm_completion_ast);
1464 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1465 EXPORT_SYMBOL(ldlm_cli_convert);
1466 EXPORT_SYMBOL(ldlm_cli_enqueue);
1467 EXPORT_SYMBOL(ldlm_cli_cancel);
1468 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1469 EXPORT_SYMBOL(ldlm_replay_locks);
1470 EXPORT_SYMBOL(ldlm_resource_foreach);
1471 EXPORT_SYMBOL(ldlm_namespace_foreach);
1472 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1473 EXPORT_SYMBOL(ldlm_change_cbdata);
1474
1475 /* ldlm_lockd.c */
1476 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1477 EXPORT_SYMBOL(ldlm_server_completion_ast);
1478 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1479 EXPORT_SYMBOL(ldlm_handle_enqueue);
1480 EXPORT_SYMBOL(ldlm_handle_cancel);
1481 EXPORT_SYMBOL(ldlm_handle_convert);
1482 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1483 EXPORT_SYMBOL(ldlm_get_ref);
1484 EXPORT_SYMBOL(ldlm_put_ref);
1485
1486 #if 0
1487 /* ldlm_test.c */
1488 EXPORT_SYMBOL(ldlm_test);
1489 EXPORT_SYMBOL(ldlm_regression_start);
1490 EXPORT_SYMBOL(ldlm_regression_stop);
1491 #endif
1492
1493 /* ldlm_resource.c */
1494 EXPORT_SYMBOL(ldlm_namespace_new);
1495 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1496 EXPORT_SYMBOL(ldlm_namespace_free);
1497 EXPORT_SYMBOL(ldlm_resource_get);
1498 EXPORT_SYMBOL(ldlm_resource_putref);
1499
1500 /* l_lock.c */
1501 EXPORT_SYMBOL(l_lock);
1502 EXPORT_SYMBOL(l_unlock);
1503
1504 /* ldlm_lib.c */
1505 EXPORT_SYMBOL(client_obd_setup);
1506 EXPORT_SYMBOL(client_obd_cleanup);
1507 EXPORT_SYMBOL(client_connect_import);
1508 EXPORT_SYMBOL(client_disconnect_export);
1509 EXPORT_SYMBOL(target_abort_recovery);
1510 EXPORT_SYMBOL(target_handle_connect);
1511 EXPORT_SYMBOL(target_destroy_export);
1512 EXPORT_SYMBOL(target_cancel_recovery_timer);
1513 EXPORT_SYMBOL(target_send_reply);
1514 EXPORT_SYMBOL(target_queue_recovery_request);
1515 EXPORT_SYMBOL(target_handle_ping);
1516 EXPORT_SYMBOL(target_handle_disconnect);
1517 EXPORT_SYMBOL(target_queue_final_reply);