lustre/ldlm/ldlm_lockd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #ifndef EXPORT_SYMTAB
  25 # define EXPORT_SYMTAB
  26 #endif
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28
  29 #ifdef __KERNEL__
  30 # include <linux/module.h>
  31 # include <linux/slab.h>
  32 # include <linux/init.h>
  33 # include <linux/wait.h>
  34 #else
  35 # include <liblustre.h>
  36 #endif
  37
  38 #include <linux/lustre_dlm.h>
  39 #include <linux/obd_class.h>
  40 #include <portals/list.h>
  41 #include "ldlm_internal.h"
  42
  43 extern kmem_cache_t *ldlm_resource_slab;
  44 extern kmem_cache_t *ldlm_lock_slab;
  45 extern struct lustre_lock ldlm_handle_lock;
  46 extern struct list_head ldlm_namespace_list;
  47 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
  48 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
  49
  50 static DECLARE_MUTEX(ldlm_ref_sem);
  51 static int ldlm_refcount = 0;
  52
  53 /* LDLM state */
  54
  55 static struct ldlm_state *ldlm_state;
  56
  57 inline unsigned long round_timeout(unsigned long timeout)
  58 {
  59         return ((timeout / HZ) + 1) * HZ;
  60 }
  61
  62 #ifdef __KERNEL__
  63 /* XXX should this be per-ldlm? */
  64 static struct list_head waiting_locks_list;
  65 static spinlock_t waiting_locks_spinlock;
  66 static struct timer_list waiting_locks_timer;
  67
  68 static struct expired_lock_thread {
  69         wait_queue_head_t         elt_waitq;
  70         int                       elt_state;
  71         struct list_head          elt_expired_locks;
  72         spinlock_t                elt_lock;
  73 } expired_lock_thread;
  74 #endif
  75
  76 #define ELT_STOPPED   0
  77 #define ELT_READY     1
  78 #define ELT_TERMINATE 2
  79
  80 struct ldlm_bl_pool {
  81         spinlock_t              blp_lock;
  82         struct list_head        blp_list;
  83         wait_queue_head_t       blp_waitq;
  84         atomic_t                blp_num_threads;
  85         struct completion       blp_comp;
  86 };
  87
  88 struct ldlm_bl_work_item {
  89         struct list_head        blwi_entry;
  90         struct ldlm_namespace   *blwi_ns;
  91         struct ldlm_lock_desc   blwi_ld;
  92         struct ldlm_lock        *blwi_lock;
  93 };
  94
  95 #ifdef __KERNEL__
  96
  97 static inline int have_expired_locks(void)
  98 {
  99         int need_to_run;
 100
 101         spin_lock_bh(&expired_lock_thread.elt_lock);
 102         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 103         spin_unlock_bh(&expired_lock_thread.elt_lock);
 104
 105         RETURN(need_to_run);
 106 }
 107
 108 static int expired_lock_main(void *arg)
 109 {
 110         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 111         struct l_wait_info lwi = { 0 };
 112         unsigned long flags;
 113
 114         ENTRY;
 115         lock_kernel();
 116         kportal_daemonize("ldlm_elt");
 117
 118         SIGNAL_MASK_LOCK(current, flags);
 119         sigfillset(&current->blocked);
 120         RECALC_SIGPENDING;
 121         SIGNAL_MASK_UNLOCK(current, flags);
 122
 123         unlock_kernel();
 124
 125         expired_lock_thread.elt_state = ELT_READY;
 126         wake_up(&expired_lock_thread.elt_waitq);
 127
 128         while (1) {
 129                 struct list_head *tmp, *n, work_list;
 130                 l_wait_event(expired_lock_thread.elt_waitq,
 131                              have_expired_locks() ||
 132                              expired_lock_thread.elt_state == ELT_TERMINATE,
 133                              &lwi);
 134
 135                 spin_lock_bh(&expired_lock_thread.elt_lock);
 136                 while (!list_empty(expired)) {
 137                         struct ldlm_lock *lock;
 138
 139                         list_add(&work_list, expired);
 140                         list_del_init(expired);
 141
 142                         list_for_each_entry(lock, &work_list, l_pending_chain) {
 143                                 LDLM_DEBUG(lock, "moving to work list");
 144                         }
 145
 146                         spin_unlock_bh(&expired_lock_thread.elt_lock);
 147
 148
 149                         list_for_each_safe(tmp, n, &work_list) {
 150                                  lock = list_entry(tmp, struct ldlm_lock,
 151                                                    l_pending_chain);
 152                                  ptlrpc_fail_export(lock->l_export);
 153                         }
 154
 155
 156                         if (!list_empty(&work_list)) {
 157                                 list_for_each_entry(lock, &work_list, l_pending_chain) {
 158                                         LDLM_ERROR(lock, "still on work list!");
 159                                 }
 160                         }
 161                         LASSERTF (list_empty(&work_list),
 162                                   "some exports not failed properly\n");
 163
 164                         spin_lock_bh(&expired_lock_thread.elt_lock);
 165                 }
 166                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 167
 168                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 169                         break;
 170         }
 171
 172         expired_lock_thread.elt_state = ELT_STOPPED;
 173         wake_up(&expired_lock_thread.elt_waitq);
 174         RETURN(0);
 175 }
 176
 177 static void waiting_locks_callback(unsigned long unused)
 178 {
 179         struct ldlm_lock *lock;
 180         char str[PTL_NALFMT_SIZE];
 181
 182         spin_lock_bh(&waiting_locks_spinlock);
 183         while (!list_empty(&waiting_locks_list)) {
 184                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 185                                   l_pending_chain);
 186
 187                 if (lock->l_callback_timeout > jiffies)
 188                         break;
 189
 190                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 191                            "%s@%s nid "LPX64" (%s) ",
 192                            lock->l_export->exp_client_uuid.uuid,
 193                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 194                            lock->l_export->exp_connection->c_peer.peer_nid,
 195                            portals_nid2str(lock->l_export->exp_connection->c_peer.peer_ni->pni_number,
 196                                            lock->l_export->exp_connection->c_peer.peer_nid,
 197                                            str));
 198
 199                 spin_lock_bh(&expired_lock_thread.elt_lock);
 200                 list_del(&lock->l_pending_chain);
 201                 list_add(&lock->l_pending_chain,
 202                          &expired_lock_thread.elt_expired_locks);
 203                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 204                 wake_up(&expired_lock_thread.elt_waitq);
 205         }
 206
 207         /*
 208          * Make sure the timer will fire again if we have any locks
 209          * left.
 210          */
 211         if (!list_empty(&waiting_locks_list)) {
 212                 unsigned long timeout_rounded;
 213                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 214                                   l_pending_chain);
 215                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 216                 mod_timer(&waiting_locks_timer, timeout_rounded);
 217         }
 218         spin_unlock_bh(&waiting_locks_spinlock);
 219 }
 220
 221 /*
 222  * Indicate that we're waiting for a client to call us back cancelling a given
 223  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 224  * timer to fire appropriately.  (We round up to the next second, to avoid
 225  * floods of timer firings during periods of high lock contention and traffic).
 226  */
 227 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 228 {
 229         unsigned long timeout_rounded;
 230
 231         spin_lock_bh(&waiting_locks_spinlock);
 232         if (!list_empty(&lock->l_pending_chain)) {
 233                 LDLM_DEBUG(lock, "not re-adding to wait list");
 234                 spin_unlock_bh(&waiting_locks_spinlock);
 235                 return 0;
 236         }
 237         LDLM_DEBUG(lock, "adding to wait list");
 238
 239         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 240
 241         timeout_rounded = round_timeout(lock->l_callback_timeout);
 242
 243         if (timeout_rounded < waiting_locks_timer.expires ||
 244             !timer_pending(&waiting_locks_timer)) {
 245                 mod_timer(&waiting_locks_timer, timeout_rounded);
 246         }
 247         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 248         spin_unlock_bh(&waiting_locks_spinlock);
 249         return 1;
 250 }
 251
 252 /*
 253  * Remove a lock from the pending list, likely because it had its cancellation
 254  * callback arrive without incident.  This adjusts the lock-timeout timer if
 255  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 256  */
 257 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 258 {
 259         struct list_head *list_next;
 260
 261         if (lock->l_export == NULL) {
 262                 /* We don't have a "waiting locks list" on clients. */
 263                 LDLM_DEBUG(lock, "client lock: no-op");
 264                 return 0;
 265         }
 266
 267         spin_lock_bh(&waiting_locks_spinlock);
 268
 269         if (list_empty(&lock->l_pending_chain)) {
 270                 spin_unlock_bh(&waiting_locks_spinlock);
 271                 LDLM_DEBUG(lock, "wasn't waiting");
 272                 return 0;
 273         }
 274
 275         list_next = lock->l_pending_chain.next;
 276         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 277                 /* Removing the head of the list, adjust timer. */
 278                 if (list_next == &waiting_locks_list) {
 279                         /* No more, just cancel. */
 280                         del_timer(&waiting_locks_timer);
 281                 } else {
 282                         struct ldlm_lock *next;
 283                         next = list_entry(list_next, struct ldlm_lock,
 284                                           l_pending_chain);
 285                         mod_timer(&waiting_locks_timer,
 286                                   round_timeout(next->l_callback_timeout));
 287                 }
 288         }
 289
 290         spin_lock_bh(&expired_lock_thread.elt_lock);
 291         list_del_init(&lock->l_pending_chain);
 292         spin_unlock_bh(&expired_lock_thread.elt_lock);
 293
 294         spin_unlock_bh(&waiting_locks_spinlock);
 295         LDLM_DEBUG(lock, "removed");
 296         return 1;
 297 }
 298
 299 #else /* !__KERNEL__ */
 300
 301 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 302 {
 303         RETURN(1);
 304 }
 305
 306 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 307 {
 308         RETURN(0);
 309 }
 310
 311 #endif /* __KERNEL__ */
 312
 313 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, char *ast_type)
 314 {
 315         const struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 316         char str[PTL_NALFMT_SIZE];
 317
 318         CERROR("%s AST failed (%d) for res "LPU64"/"LPU64
 319                ", mode %s: evicting client %s@%s NID "LPX64" (%s)\n",
 320                ast_type, rc,
 321                lock->l_resource->lr_name.name[0],
 322                lock->l_resource->lr_name.name[1],
 323                ldlm_lockname[lock->l_granted_mode],
 324                lock->l_export->exp_client_uuid.uuid,
 325                conn->c_remote_uuid.uuid, conn->c_peer.peer_nid,
 326                portals_nid2str(conn->c_peer.peer_ni->pni_number,
 327                                conn->c_peer.peer_nid, str));
 328         ptlrpc_fail_export(lock->l_export);
 329 }
 330
 331 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 332                              struct ldlm_lock_desc *desc,
 333                              void *data, int flag)
 334 {
 335         struct ldlm_request *body;
 336         struct ptlrpc_request *req;
 337         int rc = 0, size = sizeof(*body);
 338         ENTRY;
 339
 340         if (flag == LDLM_CB_CANCELING) {
 341                 /* Don't need to do anything here. */
 342                 RETURN(0);
 343         }
 344
 345         LASSERT(lock);
 346
 347         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 348         if (lock->l_granted_mode != lock->l_req_mode) {
 349                 /* this blocking AST will be communicated as part of the
 350                  * completion AST instead */
 351                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 352                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");                 RETURN(0);
 353         }
 354
 355         if (lock->l_destroyed) {
 356                 /* What's the point? */
 357                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 358                 RETURN(0);
 359         }
 360
 361 #if 0
 362         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 363                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 364                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 365                 RETURN(-ETIMEDOUT);
 366         }
 367 #endif
 368
 369         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 370                               LDLM_BL_CALLBACK, 1, &size, NULL);
 371         if (req == NULL) {
 372                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 373                 RETURN(-ENOMEM);
 374         }
 375
 376         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 377         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 378                sizeof(body->lock_handle1));
 379         memcpy(&body->lock_desc, desc, sizeof(*desc));
 380         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 381
 382         LDLM_DEBUG(lock, "server preparing blocking AST");
 383         req->rq_replen = lustre_msg_size(0, NULL);
 384
 385         if (lock->l_granted_mode == lock->l_req_mode)
 386                 ldlm_add_waiting_lock(lock);
 387         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 388
 389         req->rq_send_state = LUSTRE_IMP_FULL;
 390         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 391         rc = ptlrpc_queue_wait(req);
 392         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 393                 LASSERT(lock->l_export);
 394                 if (lock->l_export->exp_libclient) {
 395                         CDEBUG(D_HA, "BLOCKING AST to liblustre client (nid "
 396                                LPU64") timeout, simply cancel lock 0x%p\n",
 397                                req->rq_peer.peer_nid, lock);
 398                         ldlm_lock_cancel(lock);
 399                         rc = -ERESTART;
 400                 } else {
 401                         ldlm_del_waiting_lock(lock);
 402                         ldlm_failed_ast(lock, rc, "blocking");
 403                 }
 404         } else if (rc) {
 405                 if (rc == -EINVAL)
 406                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 407                                "from blocking AST for lock %p--normal race\n",
 408                                req->rq_peer.peer_nid,
 409                                req->rq_repmsg->status, lock);
 410                 else if (rc == -ENOTCONN)
 411                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 412                                "from blocking AST for lock %p--this client was "
 413                                "probably rebooted while it held a lock, nothing"
 414                                " serious\n",req->rq_peer.peer_nid,
 415                                req->rq_repmsg->status, lock);
 416                 else
 417                         CDEBUG(D_ERROR, "client (nid "LPU64") returned %d "
 418                                "from blocking AST for lock %p\n",
 419                                req->rq_peer.peer_nid,
 420                                (req->rq_repmsg != NULL)?
 421                                req->rq_repmsg->status : 0,
 422                                lock);
 423                 LDLM_DEBUG(lock, "client sent rc %d rq_status %d from blocking "
 424                            "AST", rc, req->rq_status);
 425                 ldlm_lock_cancel(lock);
 426                 /* Server-side AST functions are called from ldlm_reprocess_all,
 427                  * which needs to be told to please restart its reprocessing. */
 428                 rc = -ERESTART;
 429         }
 430
 431         ptlrpc_req_finished(req);
 432
 433         RETURN(rc);
 434 }
 435
 436 /* XXX copied from ptlrpc/service.c */
 437 static long timeval_sub(struct timeval *large, struct timeval *small)
 438 {
 439         return (large->tv_sec - small->tv_sec) * 1000000 +
 440                 (large->tv_usec - small->tv_usec);
 441 }
 442
 443 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 444 {
 445         struct ldlm_request *body;
 446         struct ptlrpc_request *req;
 447         struct timeval granted_time;
 448         long total_enqueue_wait;
 449         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 450         ENTRY;
 451
 452         LASSERT(lock != NULL);
 453
 454         do_gettimeofday(&granted_time);
 455         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 456
 457         if (total_enqueue_wait / 1000000 > obd_timeout)
 458                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 459
 460         if (lock->l_resource->lr_lvb_len) {
 461                 buffers = 2;
 462                 size[1] = lock->l_resource->lr_lvb_len;
 463         }
 464
 465         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 466                               LDLM_CP_CALLBACK, buffers, size, NULL);
 467         if (req == NULL)
 468                 RETURN(-ENOMEM);
 469
 470         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 471         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 472                sizeof(body->lock_handle1));
 473         body->lock_flags = flags;
 474         ldlm_lock2desc(lock, &body->lock_desc);
 475
 476         if (buffers == 2) {
 477                 void *lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 478                                            lock->l_resource->lr_lvb_len);
 479                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 480                        lock->l_resource->lr_lvb_len);
 481         }
 482
 483         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 484                    total_enqueue_wait);
 485         req->rq_replen = lustre_msg_size(0, NULL);
 486
 487         req->rq_send_state = LUSTRE_IMP_FULL;
 488         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 489
 490         /* We only send real blocking ASTs after the lock is granted */
 491         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 492         if (lock->l_flags & LDLM_FL_AST_SENT) {
 493                 body->lock_flags |= LDLM_FL_AST_SENT;
 494                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 495         }
 496         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 497
 498         rc = ptlrpc_queue_wait(req);
 499         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 500                 ldlm_del_waiting_lock(lock);
 501                 ldlm_failed_ast(lock, rc, "completion");
 502         } else if (rc == -EINVAL) {
 503                 LDLM_DEBUG(lock, "lost the race -- client no longer has this "
 504                            "lock");
 505         } else if (rc) {
 506                 LDLM_ERROR(lock, "client sent rc %d rq_status %d from "
 507                            "completion AST", rc, req->rq_status);
 508                 ldlm_lock_cancel(lock);
 509                 /* Server-side AST functions are called from ldlm_reprocess_all,
 510                  * which needs to be told to please restart its reprocessing. */
 511                 rc = -ERESTART;
 512         }
 513         ptlrpc_req_finished(req);
 514
 515         RETURN(rc);
 516 }
 517
 518 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 519 {
 520         struct ldlm_resource *res = lock->l_resource;
 521         struct ldlm_request *body;
 522         struct ptlrpc_request *req;
 523         int rc = 0, size = sizeof(*body);
 524         ENTRY;
 525
 526         LASSERT(lock != NULL);
 527
 528         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 529                               LDLM_GL_CALLBACK, 1, &size, NULL);
 530         if (req == NULL)
 531                 RETURN(-ENOMEM);
 532
 533         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 534         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 535                sizeof(body->lock_handle1));
 536         ldlm_lock2desc(lock, &body->lock_desc);
 537
 538         size = lock->l_resource->lr_lvb_len;
 539         req->rq_replen = lustre_msg_size(1, &size);
 540
 541         req->rq_send_state = LUSTRE_IMP_FULL;
 542         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 543
 544         rc = ptlrpc_queue_wait(req);
 545         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 546                 ldlm_del_waiting_lock(lock);
 547                 ldlm_failed_ast(lock, rc, "glimpse");
 548         } else if (rc == -EINVAL) {
 549                 LDLM_DEBUG(lock, "lost the race -- client no longer has this "
 550                            "lock");
 551         } else if (rc == -ELDLM_NO_LOCK_DATA) {
 552                 LDLM_DEBUG(lock, "lost a race -- client has a lock, but no "
 553                            "inode");
 554         } else if (rc) {
 555                 LDLM_ERROR(lock, "client sent rc %d rq_status %d from "
 556                            "glimpse AST", rc, req->rq_status);
 557         } else {
 558                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 559                         (res, req->rq_repmsg, 0, 1);
 560         }
 561         ptlrpc_req_finished(req);
 562         RETURN(rc);
 563 }
 564
 565 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 566                         ldlm_completion_callback completion_callback,
 567                         ldlm_blocking_callback blocking_callback,
 568                         ldlm_glimpse_callback glimpse_callback)
 569 {
 570         struct obd_device *obddev = req->rq_export->exp_obd;
 571         struct ldlm_reply *dlm_rep;
 572         struct ldlm_request *dlm_req;
 573         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 574         __u32 flags;
 575         ldlm_error_t err = ELDLM_OK;
 576         struct ldlm_lock *lock = NULL;
 577         void *cookie = NULL;
 578         ENTRY;
 579
 580         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 581
 582         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 583                                       lustre_swab_ldlm_request);
 584         if (dlm_req == NULL) {
 585                 CERROR ("Can't unpack dlm_req\n");
 586                 GOTO(out, rc = -EFAULT);
 587         }
 588
 589         flags = dlm_req->lock_flags;
 590
 591         /* The lock's callback data might be set in the policy function */
 592         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 593                                 dlm_req->lock_desc.l_resource.lr_name,
 594                                 dlm_req->lock_desc.l_resource.lr_type,
 595                                 dlm_req->lock_desc.l_req_mode,
 596                                 blocking_callback, completion_callback,
 597                                 glimpse_callback, NULL, 0);
 598         if (!lock)
 599                 GOTO(out, rc = -ENOMEM);
 600
 601         do_gettimeofday(&lock->l_enqueued_time);
 602         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 603                sizeof(lock->l_remote_handle));
 604         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 605
 606         LASSERT(req->rq_export);
 607         lock->l_export = class_export_get(req->rq_export);
 608         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 609         list_add(&lock->l_export_chain,
 610                  &lock->l_export->exp_ldlm_data.led_held_locks);
 611         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 612
 613         if (flags & LDLM_FL_HAS_INTENT) {
 614                 /* In this case, the reply buffer is allocated deep in
 615                  * local_lock_enqueue by the policy function. */
 616                 cookie = req;
 617         } else {
 618                 int buffers = 1;
 619                 if (lock->l_resource->lr_lvb_len) {
 620                         size[1] = lock->l_resource->lr_lvb_len;
 621                         buffers = 2;
 622                 }
 623
 624                 rc = lustre_pack_reply(req, buffers, size, NULL);
 625                 if (rc)
 626                         GOTO(out, rc);
 627         }
 628
 629         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 630                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 631                        sizeof(ldlm_policy_data_t));
 632         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 633                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 634                        sizeof(lock->l_req_extent));
 635
 636         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 637         if (err)
 638                 GOTO(out, err);
 639
 640         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 641         dlm_rep->lock_flags = flags;
 642
 643         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 644         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 645
 646         /* We never send a blocking AST until the lock is granted, but
 647          * we can tell it right now */
 648         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 649         if (lock->l_flags & LDLM_FL_AST_SENT) {
 650                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 651                 if (lock->l_granted_mode == lock->l_req_mode)
 652                         ldlm_add_waiting_lock(lock);
 653         }
 654         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 655
 656         EXIT;
 657  out:
 658         req->rq_status = err;
 659         if (req->rq_reply_state == NULL) {
 660                 err = lustre_pack_reply(req, 0, NULL, NULL);
 661                 if (rc == 0)
 662                         rc = err;
 663         }
 664
 665         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 666          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 667         if (lock) {
 668                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 669                            "(err=%d, rc=%d)", err, rc);
 670
 671                 if (lock->l_resource->lr_lvb_len > 0) {
 672                         void *lvb = lustre_msg_buf(req->rq_repmsg, 1,
 673                                                   lock->l_resource->lr_lvb_len);
 674                         memcpy(lvb, lock->l_resource->lr_lvb_data,
 675                                lock->l_resource->lr_lvb_len);
 676                 }
 677
 678                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 679                         ldlm_reprocess_all(lock->l_resource);
 680                 LDLM_LOCK_PUT(lock);
 681         }
 682         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 683                           lock, rc);
 684
 685         return rc;
 686 }
 687
 688 int ldlm_handle_convert(struct ptlrpc_request *req)
 689 {
 690         struct ldlm_request *dlm_req;
 691         struct ldlm_reply *dlm_rep;
 692         struct ldlm_lock *lock;
 693         int rc, size = sizeof(*dlm_rep);
 694         ENTRY;
 695
 696         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 697                                       lustre_swab_ldlm_request);
 698         if (dlm_req == NULL) {
 699                 CERROR ("Can't unpack dlm_req\n");
 700                 RETURN (-EFAULT);
 701         }
 702
 703         rc = lustre_pack_reply(req, 1, &size, NULL);
 704         if (rc) {
 705                 CERROR("out of memory\n");
 706                 RETURN(-ENOMEM);
 707         }
 708         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 709         dlm_rep->lock_flags = dlm_req->lock_flags;
 710
 711         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 712         if (!lock) {
 713                 req->rq_status = EINVAL;
 714         } else {
 715                 LDLM_DEBUG(lock, "server-side convert handler START");
 716                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 717                                   &dlm_rep->lock_flags);
 718                 if (ldlm_del_waiting_lock(lock))
 719                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 720                 req->rq_status = 0;
 721         }
 722
 723         if (lock) {
 724                 ldlm_reprocess_all(lock->l_resource);
 725                 LDLM_DEBUG(lock, "server-side convert handler END");
 726                 LDLM_LOCK_PUT(lock);
 727         } else
 728                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 729
 730         RETURN(0);
 731 }
 732
 733 int ldlm_handle_cancel(struct ptlrpc_request *req)
 734 {
 735         struct ldlm_request *dlm_req;
 736         struct ldlm_lock *lock;
 737         struct ldlm_resource *res;
 738         char str[PTL_NALFMT_SIZE];
 739         int rc;
 740         ENTRY;
 741
 742         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 743                                       lustre_swab_ldlm_request);
 744         if (dlm_req == NULL) {
 745                 CERROR("bad request buffer for cancel\n");
 746                 RETURN(-EFAULT);
 747         }
 748
 749         rc = lustre_pack_reply(req, 0, NULL, NULL);
 750         if (rc) {
 751                 CERROR("out of memory\n");
 752                 RETURN(-ENOMEM);
 753         }
 754
 755         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 756         if (!lock) {
 757                 CERROR("received cancel for unknown lock cookie "LPX64
 758                        " from client %s nid "LPX64" (%s)\n",
 759                        dlm_req->lock_handle1.cookie,
 760                        req->rq_export->exp_client_uuid.uuid,
 761                        req->rq_peer.peer_nid,
 762                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 763                                        req->rq_peer.peer_nid, str));
 764                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 765                                   "(cookie "LPU64")",
 766                                   dlm_req->lock_handle1.cookie);
 767                 req->rq_status = ESTALE;
 768         } else {
 769                 LDLM_DEBUG(lock, "server-side cancel handler START");
 770                 res = lock->l_resource;
 771                 if (res && res->lr_namespace->ns_lvbo &&
 772                     res->lr_namespace->ns_lvbo->lvbo_update) {
 773                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 774                                 (res, NULL, 0, 0);
 775                                 //(res, req->rq_reqmsg, 1);
 776                 }
 777
 778                 ldlm_lock_cancel(lock);
 779                 if (ldlm_del_waiting_lock(lock))
 780                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 781                 req->rq_status = rc;
 782         }
 783
 784         if (ptlrpc_reply(req) != 0)
 785                 LBUG();
 786
 787         if (lock) {
 788                 ldlm_reprocess_all(lock->l_resource);
 789                 LDLM_DEBUG(lock, "server-side cancel handler END");
 790                 LDLM_LOCK_PUT(lock);
 791         }
 792
 793         RETURN(0);
 794 }
 795
 796 static void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 797                                     struct ldlm_lock_desc *ld,
 798                                     struct ldlm_lock *lock)
 799 {
 800         int do_ast;
 801         ENTRY;
 802
 803         l_lock(&ns->ns_lock);
 804         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 805
 806         lock->l_flags |= LDLM_FL_CBPENDING;
 807         do_ast = (!lock->l_readers && !lock->l_writers);
 808
 809         if (do_ast) {
 810                 LDLM_DEBUG(lock, "already unused, calling "
 811                            "callback (%p)", lock->l_blocking_ast);
 812                 if (lock->l_blocking_ast != NULL) {
 813                         l_unlock(&ns->ns_lock);
 814                         l_check_no_ns_lock(ns);
 815                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 816                                              LDLM_CB_BLOCKING);
 817                         l_lock(&ns->ns_lock);
 818                 }
 819         } else {
 820                 LDLM_DEBUG(lock, "Lock still has references, will be"
 821                            " cancelled later");
 822         }
 823
 824         LDLM_DEBUG(lock, "client blocking callback handler END");
 825         l_unlock(&ns->ns_lock);
 826         LDLM_LOCK_PUT(lock);
 827         EXIT;
 828 }
 829
 830 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 831                                     struct ldlm_namespace *ns,
 832                                     struct ldlm_request *dlm_req,
 833                                     struct ldlm_lock *lock)
 834 {
 835         LIST_HEAD(ast_list);
 836         ENTRY;
 837
 838         l_lock(&ns->ns_lock);
 839         LDLM_DEBUG(lock, "client completion callback handler START");
 840
 841         /* If we receive the completion AST before the actual enqueue returned,
 842          * then we might need to switch lock modes, resources, or extents. */
 843         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 844                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 845                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 846         }
 847
 848         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 849                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 850                        sizeof(lock->l_policy_data));
 851                 LDLM_DEBUG(lock, "completion AST, new policy data");
 852         }
 853
 854         ldlm_resource_unlink_lock(lock);
 855         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 856                    &lock->l_resource->lr_name,
 857                    sizeof(lock->l_resource->lr_name)) != 0) {
 858                 ldlm_lock_change_resource(ns, lock,
 859                                          dlm_req->lock_desc.l_resource.lr_name);
 860                 LDLM_DEBUG(lock, "completion AST, new resource");
 861         }
 862
 863         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 864                 lock->l_flags |= LDLM_FL_CBPENDING;
 865                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 866         }
 867
 868         if (lock->l_lvb_len) {
 869                 void *lvb;
 870                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 871                                          lock->l_lvb_swabber);
 872                 if (lvb == NULL) {
 873                         LDLM_ERROR(lock, "completion AST did not contain "
 874                                    "expected LVB!");
 875                 } else {
 876                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 877                 }
 878         }
 879
 880         lock->l_resource->lr_tmp = &ast_list;
 881         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 882         lock->l_resource->lr_tmp = NULL;
 883         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 884         l_unlock(&ns->ns_lock);
 885         LDLM_LOCK_PUT(lock);
 886
 887         ldlm_run_ast_work(ns, &ast_list);
 888
 889         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 890                           lock);
 891         EXIT;
 892 }
 893
 894 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 895                                     struct ldlm_namespace *ns,
 896                                     struct ldlm_request *dlm_req,
 897                                     struct ldlm_lock *lock)
 898 {
 899         int rc = -ENOSYS;
 900         ENTRY;
 901
 902         l_lock(&ns->ns_lock);
 903         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 904
 905         if (lock->l_glimpse_ast != NULL) {
 906                 l_unlock(&ns->ns_lock);
 907                 l_check_no_ns_lock(ns);
 908                 rc = lock->l_glimpse_ast(lock, req);
 909                 l_lock(&ns->ns_lock);
 910         }
 911
 912         if (req->rq_repmsg != NULL) {
 913                 ptlrpc_reply(req);
 914         } else {
 915                 req->rq_status = rc;
 916                 ptlrpc_error(req);
 917         }
 918
 919         if (lock->l_granted_mode == LCK_PW &&
 920             !lock->l_readers && !lock->l_writers &&
 921             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 922                 l_unlock(&ns->ns_lock);
 923                 ldlm_handle_bl_callback(ns, NULL, lock);
 924                 EXIT;
 925                 return;
 926         }
 927
 928         l_unlock(&ns->ns_lock);
 929         LDLM_LOCK_PUT(lock);
 930         EXIT;
 931 }
 932
 933 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 934 {
 935         req->rq_status = rc;
 936         if (req->rq_reply_state == NULL) {
 937                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 938                 if (rc)
 939                         return rc;
 940         }
 941         return ptlrpc_reply(req);
 942 }
 943
 944 #ifdef __KERNEL__
 945 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
 946                       struct ldlm_lock *lock)
 947 {
 948         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
 949         struct ldlm_bl_work_item *blwi;
 950         ENTRY;
 951
 952         OBD_ALLOC(blwi, sizeof(*blwi));
 953         if (blwi == NULL)
 954                 RETURN(-ENOMEM);
 955
 956         blwi->blwi_ns = ns;
 957         if (ld != NULL)
 958                 blwi->blwi_ld = *ld;
 959         blwi->blwi_lock = lock;
 960
 961         spin_lock(&blp->blp_lock);
 962         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
 963         wake_up(&blp->blp_waitq);
 964         spin_unlock(&blp->blp_lock);
 965
 966         RETURN(0);
 967 }
 968 #else
 969 /* XXX */
 970 void liblustre_ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 971                                        struct ldlm_lock_desc *ld,
 972                                        struct ldlm_lock *lock)
 973 {
 974         ldlm_handle_bl_callback(ns, ld, lock);
 975 }
 976 #endif
 977
 978 static int ldlm_callback_handler(struct ptlrpc_request *req)
 979 {
 980         struct ldlm_namespace *ns;
 981         struct ldlm_request *dlm_req;
 982         struct ldlm_lock *lock;
 983         char str[PTL_NALFMT_SIZE];
 984         int rc;
 985         ENTRY;
 986
 987         /* Requests arrive in sender's byte order.  The ptlrpc service
 988          * handler has already checked and, if necessary, byte-swapped the
 989          * incoming request message body, but I am responsible for the
 990          * message buffers. */
 991
 992         if (req->rq_export == NULL) {
 993                 struct ldlm_request *dlm_req;
 994
 995                 CDEBUG(D_RPCTRACE, "operation %d from nid "LPX64" (%s) with bad "
 996                        "export cookie "LPX64" (ptl req %d/rep %d); this is "
 997                        "normal if this node rebooted with a lock held\n",
 998                        req->rq_reqmsg->opc, req->rq_peer.peer_nid,
 999                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
1000                                        req->rq_peer.peer_nid, str),
1001                        req->rq_reqmsg->handle.cookie,
1002                        req->rq_request_portal, req->rq_reply_portal);
1003
1004                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1005                                              lustre_swab_ldlm_request);
1006                 if (dlm_req != NULL)
1007                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1008                                dlm_req->lock_handle1.cookie);
1009
1010                 ldlm_callback_reply(req, -ENOTCONN);
1011                 RETURN(0);
1012         }
1013
1014         LASSERT(req->rq_export != NULL);
1015         LASSERT(req->rq_export->exp_obd != NULL);
1016
1017         switch(req->rq_reqmsg->opc) {
1018         case LDLM_BL_CALLBACK:
1019                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1020                 break;
1021         case LDLM_CP_CALLBACK:
1022                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1023                 break;
1024         case LDLM_GL_CALLBACK:
1025                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1026                 break;
1027         case OBD_LOG_CANCEL:
1028                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1029                 rc = llog_origin_handle_cancel(req);
1030                 ldlm_callback_reply(req, rc);
1031                 RETURN(0);
1032         case LLOG_ORIGIN_HANDLE_CREATE:
1033                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1034                 rc = llog_origin_handle_create(req);
1035                 ldlm_callback_reply(req, rc);
1036                 RETURN(0);
1037         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1038                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1039                 rc = llog_origin_handle_next_block(req);
1040                 ldlm_callback_reply(req, rc);
1041                 RETURN(0);
1042         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1043                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1044                 rc = llog_origin_handle_read_header(req);
1045                 ldlm_callback_reply(req, rc);
1046                 RETURN(0);
1047         case LLOG_ORIGIN_HANDLE_CLOSE:
1048                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1049                 rc = llog_origin_handle_close(req);
1050                 ldlm_callback_reply(req, rc);
1051                 RETURN(0);
1052         default:
1053                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1054                 ldlm_callback_reply(req, -EPROTO);
1055                 RETURN(0);
1056         }
1057
1058         ns = req->rq_export->exp_obd->obd_namespace;
1059         LASSERT(ns != NULL);
1060
1061         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1062                                       lustre_swab_ldlm_request);
1063         if (dlm_req == NULL) {
1064                 CERROR ("can't unpack dlm_req\n");
1065                 ldlm_callback_reply (req, -EPROTO);
1066                 RETURN (0);
1067         }
1068
1069         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1070         if (!lock) {
1071                 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
1072                        dlm_req->lock_handle1.cookie);
1073                 ldlm_callback_reply(req, -EINVAL);
1074                 RETURN(0);
1075         }
1076
1077         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1078         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1079
1080         /* We want the ost thread to get this reply so that it can respond
1081          * to ost requests (write cache writeback) that might be triggered
1082          * in the callback.
1083          *
1084          * But we'd also like to be able to indicate in the reply that we're
1085          * cancelling right now, because it's unused, or have an intent result
1086          * in the reply, so we might have to push the responsibility for sending
1087          * the reply down into the AST handlers, alas. */
1088
1089         switch (req->rq_reqmsg->opc) {
1090         case LDLM_BL_CALLBACK:
1091                 CDEBUG(D_INODE, "blocking ast\n");
1092 #ifdef __KERNEL__
1093                 rc = ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock);
1094                 ldlm_callback_reply(req, rc);
1095 #else
1096                 rc = 0;
1097                 ldlm_callback_reply(req, rc);
1098                 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1099 #endif
1100                 break;
1101         case LDLM_CP_CALLBACK:
1102                 CDEBUG(D_INODE, "completion ast\n");
1103                 ldlm_callback_reply(req, 0);
1104                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1105                 break;
1106         case LDLM_GL_CALLBACK:
1107                 CDEBUG(D_INODE, "glimpse ast\n");
1108                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1109                 break;
1110         default:
1111                 LBUG();                         /* checked above */
1112         }
1113
1114         RETURN(0);
1115 }
1116
1117 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1118 {
1119         int rc;
1120         ENTRY;
1121
1122         /* Requests arrive in sender's byte order.  The ptlrpc service
1123          * handler has already checked and, if necessary, byte-swapped the
1124          * incoming request message body, but I am responsible for the
1125          * message buffers. */
1126
1127         if (req->rq_export == NULL) {
1128                 struct ldlm_request *dlm_req;
1129                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
1130                        req->rq_reqmsg->opc, req->rq_request_portal,
1131                        req->rq_reply_portal);
1132                 CERROR("--> export cookie: "LPX64"\n",
1133                        req->rq_reqmsg->handle.cookie);
1134                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1135                                              lustre_swab_ldlm_request);
1136                 if (dlm_req != NULL)
1137                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1138                 RETURN(-ENOTCONN);
1139         }
1140
1141         switch (req->rq_reqmsg->opc) {
1142
1143         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1144         case LDLM_CANCEL:
1145                 CDEBUG(D_INODE, "cancel\n");
1146                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1147                 rc = ldlm_handle_cancel(req);
1148                 if (rc)
1149                         break;
1150                 RETURN(0);
1151
1152         default:
1153                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1154                 RETURN(-EINVAL);
1155         }
1156
1157         RETURN(0);
1158 }
1159
1160 #ifdef __KERNEL__
1161 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1162 {
1163         struct ldlm_bl_work_item *blwi = NULL;
1164
1165         spin_lock(&blp->blp_lock);
1166         if (!list_empty(&blp->blp_list)) {
1167                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1168                                   blwi_entry);
1169                 list_del(&blwi->blwi_entry);
1170         }
1171         spin_unlock(&blp->blp_lock);
1172
1173         return blwi;
1174 }
1175
1176 struct ldlm_bl_thread_data {
1177         int                     bltd_num;
1178         struct ldlm_bl_pool     *bltd_blp;
1179 };
1180
1181 static int ldlm_bl_thread_main(void *arg)
1182 {
1183         struct ldlm_bl_thread_data *bltd = arg;
1184         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1185         unsigned long flags;
1186         ENTRY;
1187
1188         /* XXX boiler-plate */
1189         {
1190                 char name[sizeof(current->comm)];
1191                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1192                          bltd->bltd_num);
1193                 kportal_daemonize(name);
1194         }
1195         SIGNAL_MASK_LOCK(current, flags);
1196         sigfillset(&current->blocked);
1197         RECALC_SIGPENDING;
1198         SIGNAL_MASK_UNLOCK(current, flags);
1199
1200         atomic_inc(&blp->blp_num_threads);
1201         complete(&blp->blp_comp);
1202
1203         while(1) {
1204                 struct l_wait_info lwi = { 0 };
1205                 struct ldlm_bl_work_item *blwi = NULL;
1206
1207                 l_wait_event_exclusive(blp->blp_waitq,
1208                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1209                                        &lwi);
1210
1211                 if (blwi->blwi_ns == NULL)
1212                         break;
1213
1214                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1215                                         blwi->blwi_lock);
1216                 OBD_FREE(blwi, sizeof(*blwi));
1217         }
1218
1219         atomic_dec(&blp->blp_num_threads);
1220         complete(&blp->blp_comp);
1221         RETURN(0);
1222 }
1223
1224 #endif
1225
1226 static int ldlm_setup(void);
1227 static int ldlm_cleanup(int force);
1228
1229 int ldlm_get_ref(void)
1230 {
1231         int rc = 0;
1232         down(&ldlm_ref_sem);
1233         if (++ldlm_refcount == 1) {
1234                 rc = ldlm_setup();
1235                 if (rc)
1236                         ldlm_refcount--;
1237         }
1238         up(&ldlm_ref_sem);
1239
1240         RETURN(rc);
1241 }
1242
1243 void ldlm_put_ref(int force)
1244 {
1245         down(&ldlm_ref_sem);
1246         if (ldlm_refcount == 1) {
1247                 int rc = ldlm_cleanup(force);
1248                 if (rc)
1249                         CERROR("ldlm_cleanup failed: %d\n", rc);
1250                 else
1251                         ldlm_refcount--;
1252         } else {
1253                 ldlm_refcount--;
1254         }
1255         up(&ldlm_ref_sem);
1256
1257         EXIT;
1258 }
1259
1260 static int ldlm_setup(void)
1261 {
1262         struct ldlm_bl_pool *blp;
1263         int rc = 0;
1264 #ifdef __KERNEL__
1265         int i;
1266 #endif
1267         ENTRY;
1268
1269         if (ldlm_state != NULL)
1270                 RETURN(-EALREADY);
1271
1272         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1273         if (ldlm_state == NULL)
1274                 RETURN(-ENOMEM);
1275
1276 #ifdef __KERNEL__
1277         rc = ldlm_proc_setup();
1278         if (rc != 0)
1279                 GOTO(out_free, rc);
1280 #endif
1281
1282         ldlm_state->ldlm_cb_service =
1283                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1284                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1285                                 ldlm_callback_handler, "ldlm_cbd",
1286                                 ldlm_svc_proc_dir);
1287
1288         if (!ldlm_state->ldlm_cb_service) {
1289                 CERROR("failed to start service\n");
1290                 GOTO(out_proc, rc = -ENOMEM);
1291         }
1292
1293         ldlm_state->ldlm_cancel_service =
1294                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1295                                 LDLM_CANCEL_REQUEST_PORTAL,
1296                                 LDLM_CANCEL_REPLY_PORTAL,
1297                                 ldlm_cancel_handler, "ldlm_canceld",
1298                                 ldlm_svc_proc_dir);
1299
1300         if (!ldlm_state->ldlm_cancel_service) {
1301                 CERROR("failed to start service\n");
1302                 GOTO(out_proc, rc = -ENOMEM);
1303         }
1304
1305         OBD_ALLOC(blp, sizeof(*blp));
1306         if (blp == NULL)
1307                 GOTO(out_proc, rc = -ENOMEM);
1308         ldlm_state->ldlm_bl_pool = blp;
1309
1310         atomic_set(&blp->blp_num_threads, 0);
1311         init_waitqueue_head(&blp->blp_waitq);
1312         spin_lock_init(&blp->blp_lock);
1313
1314         INIT_LIST_HEAD(&blp->blp_list);
1315
1316 #ifdef __KERNEL__
1317         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1318                 struct ldlm_bl_thread_data bltd = {
1319                         .bltd_num = i,
1320                         .bltd_blp = blp,
1321                 };
1322                 init_completion(&blp->blp_comp);
1323                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1324                 if (rc < 0) {
1325                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1326                         LBUG();
1327                         GOTO(out_thread, rc);
1328                 }
1329                 wait_for_completion(&blp->blp_comp);
1330         }
1331
1332         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1333                                     LDLM_NUM_THREADS, "ldlm_cn");
1334         if (rc) {
1335                 LBUG();
1336                 GOTO(out_thread, rc);
1337         }
1338
1339         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1340                                     LDLM_NUM_THREADS, "ldlm_cb");
1341         if (rc) {
1342                 LBUG();
1343                 GOTO(out_thread, rc);
1344         }
1345
1346         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1347         spin_lock_init(&expired_lock_thread.elt_lock);
1348         expired_lock_thread.elt_state = ELT_STOPPED;
1349         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1350
1351         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1352         if (rc < 0) {
1353                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1354                 GOTO(out_thread, rc);
1355         }
1356
1357         wait_event(expired_lock_thread.elt_waitq,
1358                    expired_lock_thread.elt_state == ELT_READY);
1359
1360         INIT_LIST_HEAD(&waiting_locks_list);
1361         spin_lock_init(&waiting_locks_spinlock);
1362         waiting_locks_timer.function = waiting_locks_callback;
1363         waiting_locks_timer.data = 0;
1364         init_timer(&waiting_locks_timer);
1365 #endif
1366
1367         RETURN(0);
1368
1369 #ifdef __KERNEL__
1370  out_thread:
1371         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1372         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1373 #endif
1374
1375  out_proc:
1376 #ifdef __KERNEL__
1377         ldlm_proc_cleanup();
1378  out_free:
1379 #endif
1380         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1381         ldlm_state = NULL;
1382         return rc;
1383 }
1384
1385 static int ldlm_cleanup(int force)
1386 {
1387 #ifdef __KERNEL__
1388         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1389 #endif
1390         ENTRY;
1391
1392         if (!list_empty(&ldlm_namespace_list)) {
1393                 CERROR("ldlm still has namespaces; clean these up first.\n");
1394                 ldlm_dump_all_namespaces();
1395                 RETURN(-EBUSY);
1396         }
1397
1398 #ifdef __KERNEL__
1399         while (atomic_read(&blp->blp_num_threads) > 0) {
1400                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1401
1402                 init_completion(&blp->blp_comp);
1403
1404                 spin_lock(&blp->blp_lock);
1405                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1406                 wake_up(&blp->blp_waitq);
1407                 spin_unlock(&blp->blp_lock);
1408
1409                 wait_for_completion(&blp->blp_comp);
1410         }
1411         OBD_FREE(blp, sizeof(*blp));
1412
1413         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1414         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1415         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1416         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1417         ldlm_proc_cleanup();
1418
1419         expired_lock_thread.elt_state = ELT_TERMINATE;
1420         wake_up(&expired_lock_thread.elt_waitq);
1421         wait_event(expired_lock_thread.elt_waitq,
1422                    expired_lock_thread.elt_state == ELT_STOPPED);
1423
1424 #endif
1425
1426         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1427         ldlm_state = NULL;
1428
1429         RETURN(0);
1430 }
1431
1432 int __init ldlm_init(void)
1433 {
1434         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1435                                                sizeof(struct ldlm_resource), 0,
1436                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1437         if (ldlm_resource_slab == NULL)
1438                 return -ENOMEM;
1439
1440         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1441                                            sizeof(struct ldlm_lock), 0,
1442                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1443         if (ldlm_lock_slab == NULL) {
1444                 kmem_cache_destroy(ldlm_resource_slab);
1445                 return -ENOMEM;
1446         }
1447
1448         l_lock_init(&ldlm_handle_lock);
1449
1450         return 0;
1451 }
1452
1453 void __exit ldlm_exit(void)
1454 {
1455         if ( ldlm_refcount )
1456                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1457         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
1458                 CERROR("couldn't free ldlm resource slab\n");
1459         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
1460                 CERROR("couldn't free ldlm lock slab\n");
1461 }
1462
1463 /* ldlm_flock.c */
1464 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1465
1466 /* ldlm_extent.c */
1467 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1468
1469 /* ldlm_lock.c */
1470 EXPORT_SYMBOL(ldlm_get_processing_policy);
1471 EXPORT_SYMBOL(ldlm_lock2desc);
1472 EXPORT_SYMBOL(ldlm_register_intent);
1473 EXPORT_SYMBOL(ldlm_lockname);
1474 EXPORT_SYMBOL(ldlm_typename);
1475 EXPORT_SYMBOL(ldlm_lock2handle);
1476 EXPORT_SYMBOL(__ldlm_handle2lock);
1477 EXPORT_SYMBOL(ldlm_lock_get);
1478 EXPORT_SYMBOL(ldlm_lock_put);
1479 EXPORT_SYMBOL(ldlm_lock_match);
1480 EXPORT_SYMBOL(ldlm_lock_cancel);
1481 EXPORT_SYMBOL(ldlm_lock_addref);
1482 EXPORT_SYMBOL(ldlm_lock_decref);
1483 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1484 EXPORT_SYMBOL(ldlm_lock_change_resource);
1485 EXPORT_SYMBOL(ldlm_lock_set_data);
1486 EXPORT_SYMBOL(ldlm_it2str);
1487 EXPORT_SYMBOL(ldlm_lock_dump);
1488 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1489 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1490 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1491 EXPORT_SYMBOL(ldlm_lock_allow_match);
1492
1493 /* ldlm_request.c */
1494 EXPORT_SYMBOL(ldlm_completion_ast);
1495 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1496 EXPORT_SYMBOL(ldlm_cli_convert);
1497 EXPORT_SYMBOL(ldlm_cli_enqueue);
1498 EXPORT_SYMBOL(ldlm_cli_cancel);
1499 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1500 EXPORT_SYMBOL(ldlm_replay_locks);
1501 EXPORT_SYMBOL(ldlm_resource_foreach);
1502 EXPORT_SYMBOL(ldlm_namespace_foreach);
1503 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1504 EXPORT_SYMBOL(ldlm_change_cbdata);
1505
1506 /* ldlm_lockd.c */
1507 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1508 EXPORT_SYMBOL(ldlm_server_completion_ast);
1509 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1510 EXPORT_SYMBOL(ldlm_handle_enqueue);
1511 EXPORT_SYMBOL(ldlm_handle_cancel);
1512 EXPORT_SYMBOL(ldlm_handle_convert);
1513 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1514 EXPORT_SYMBOL(ldlm_get_ref);
1515 EXPORT_SYMBOL(ldlm_put_ref);
1516
1517 #if 0
1518 /* ldlm_test.c */
1519 EXPORT_SYMBOL(ldlm_test);
1520 EXPORT_SYMBOL(ldlm_regression_start);
1521 EXPORT_SYMBOL(ldlm_regression_stop);
1522 #endif
1523
1524 /* ldlm_resource.c */
1525 EXPORT_SYMBOL(ldlm_namespace_new);
1526 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1527 EXPORT_SYMBOL(ldlm_namespace_free);
1528 EXPORT_SYMBOL(ldlm_namespace_dump);
1529 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1530 EXPORT_SYMBOL(ldlm_resource_get);
1531 EXPORT_SYMBOL(ldlm_resource_putref);
1532
1533 /* l_lock.c */
1534 EXPORT_SYMBOL(l_lock);
1535 EXPORT_SYMBOL(l_unlock);
1536
1537 /* ldlm_lib.c */
1538 EXPORT_SYMBOL(client_obd_setup);
1539 EXPORT_SYMBOL(client_obd_cleanup);
1540 EXPORT_SYMBOL(client_connect_import);
1541 EXPORT_SYMBOL(client_disconnect_export);
1542 EXPORT_SYMBOL(target_abort_recovery);
1543 EXPORT_SYMBOL(target_handle_connect);
1544 EXPORT_SYMBOL(target_destroy_export);
1545 EXPORT_SYMBOL(target_cancel_recovery_timer);
1546 EXPORT_SYMBOL(target_send_reply);
1547 EXPORT_SYMBOL(target_queue_recovery_request);
1548 EXPORT_SYMBOL(target_handle_ping);
1549 EXPORT_SYMBOL(target_handle_disconnect);
1550 EXPORT_SYMBOL(target_queue_final_reply);