lustre/ldlm/ldlm_inodebits.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lustre/ldlm/ldlm_inodebits.c
  32  *
  33  * Author: Peter Braam <braam@clusterfs.com>
  34  * Author: Phil Schwan <phil@clusterfs.com>
  35  */
  36
  37 /**
  38  * This file contains implementation of IBITS lock type
  39  *
  40  * IBITS lock type contains a bit mask determining various properties of an
  41  * object. The meanings of specific bits are specific to the caller and are
  42  * opaque to LDLM code.
  43  *
  44  * Locks with intersecting bitmasks and conflicting lock modes (e.g.  LCK_PW)
  45  * are considered conflicting.  See the lock mode compatibility matrix
  46  * in lustre_dlm.h.
  47  */
  48
  49 #define DEBUG_SUBSYSTEM S_LDLM
  50
  51 #include <lustre_dlm.h>
  52 #include <obd_support.h>
  53 #include <lustre_lib.h>
  54 #include <obd_class.h>
  55
  56 #include "ldlm_internal.h"
  57
  58 #ifdef HAVE_SERVER_SUPPORT
  59
  60 /**
  61  * It should iterate through all waiting locks on a given resource queue and
  62  * attempt to grant them. An optimization is to check only heads waitintg
  63  * locks for each inodebit type.
  64  *
  65  * Must be called with resource lock held.
  66  */
  67 int ldlm_reprocess_inodebits_queue(struct ldlm_resource *res,
  68                                    struct list_head *queue,
  69                                    struct list_head *work_list,
  70                                    enum ldlm_process_intention intention,
  71                                    struct ldlm_lock *hint)
  72 {
  73         __u64 flags;
  74         int rc = LDLM_ITER_CONTINUE;
  75         enum ldlm_error err;
  76         LIST_HEAD(bl_ast_list);
  77         struct ldlm_ibits_queues *queues = res->lr_ibits_queues;
  78         int i;
  79
  80         ENTRY;
  81
  82         check_res_locked(res);
  83
  84         LASSERT(res->lr_type == LDLM_IBITS);
  85         LASSERT(intention == LDLM_PROCESS_RESCAN ||
  86                 intention == LDLM_PROCESS_RECOVERY);
  87
  88         if (intention == LDLM_PROCESS_RECOVERY)
  89                 return ldlm_reprocess_queue(res, queue, work_list, intention,
  90                                             NULL);
  91
  92 restart:
  93         CDEBUG(D_DLMTRACE, "--- Reprocess resource "DLDLMRES" (%p)\n",
  94                PLDLMRES(res), res);
  95
  96         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
  97                 LIST_HEAD(rpc_list);
  98                 struct list_head *head = &queues->liq_waiting[i];
  99                 struct ldlm_lock *pending;
 100                 struct ldlm_ibits_node *node;
 101
 102                 if (list_empty(head))
 103                         continue;
 104                 if (hint && !(hint->l_policy_data.l_inodebits.bits & BIT(i)))
 105                         continue;
 106
 107                 node = list_entry(head->next, struct ldlm_ibits_node,
 108                                   lin_link[i]);
 109
 110                 pending = node->lock;
 111                 LDLM_DEBUG(pending, "Reprocessing lock from queue %d", i);
 112
 113                 flags = 0;
 114                 rc = ldlm_process_inodebits_lock(pending, &flags, intention,
 115                                                  &err, &rpc_list);
 116                 if (ldlm_is_granted(pending)) {
 117                         list_splice(&rpc_list, work_list);
 118                         /* Try to grant more locks from current queue */
 119                         i--;
 120                 } else {
 121                         list_splice(&rpc_list, &bl_ast_list);
 122                 }
 123         }
 124
 125         if (!list_empty(&bl_ast_list)) {
 126                 unlock_res(res);
 127
 128                 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
 129                                        LDLM_WORK_BL_AST);
 130
 131                 lock_res(res);
 132                 if (rc == -ERESTART)
 133                         GOTO(restart, rc);
 134         }
 135
 136         if (!list_empty(&bl_ast_list))
 137                 ldlm_discard_bl_list(&bl_ast_list);
 138
 139         RETURN(rc);
 140 }
 141
 142 /**
 143  * Determine if the lock is compatible with all locks on the queue.
 144  *
 145  * If \a work_list is provided, conflicting locks are linked there.
 146  * If \a work_list is not provided, we exit this function on first conflict.
 147  *
 148  * \retval 0 if there are conflicting locks in the \a queue
 149  * \retval 1 if the lock is compatible to all locks in \a queue
 150  *
 151  * IBITS locks in granted queue are organized in bunches of
 152  * same-mode/same-bits locks called "skip lists". The First lock in the
 153  * bunch contains a pointer to the end of the bunch.  This allows us to
 154  * skip an entire bunch when iterating the list in search for conflicting
 155  * locks if first lock of the bunch is not conflicting with us.
 156  */
 157 static int
 158 ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req,
 159                             __u64 *ldlm_flags, struct list_head *work_list)
 160 {
 161         enum ldlm_mode req_mode = req->l_req_mode;
 162         struct list_head *tmp;
 163         struct ldlm_lock *lock;
 164         __u64 req_bits = req->l_policy_data.l_inodebits.bits;
 165         __u64 *try_bits = &req->l_policy_data.l_inodebits.try_bits;
 166         int compat = 1;
 167
 168         ENTRY;
 169
 170         lockmode_verify(req_mode);
 171
 172         /* There is no sense in lock with no bits set. Also such a lock
 173          * would be compatible with any other bit lock.
 174          * Meanwhile that can be true if there were just try_bits and all
 175          * are failed, so just exit gracefully and let the caller to care.
 176          */
 177         if ((req_bits | *try_bits) == 0)
 178                 RETURN(0);
 179
 180         /* Group lock could be only DOM */
 181         if (unlikely(req_mode == LCK_GROUP &&
 182                      (req_bits | *try_bits) != MDS_INODELOCK_DOM))
 183                 RETURN(-EPROTO);
 184
 185         list_for_each(tmp, queue) {
 186                 struct list_head *mode_tail;
 187
 188                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 189
 190                 /* We stop walking the queue if we hit ourselves so we don't
 191                  * take conflicting locks enqueued after us into account,
 192                  * or we'd wait forever. */
 193                 if (req == lock)
 194                         RETURN(compat);
 195
 196                 /* last lock in mode group */
 197                 LASSERT(lock->l_sl_mode.prev != NULL);
 198                 mode_tail = &list_entry(lock->l_sl_mode.prev, struct ldlm_lock,
 199                                         l_sl_mode)->l_res_link;
 200
 201                 /* if request lock is not COS_INCOMPAT and COS is disabled,
 202                  * they are compatible, IOW this request is from a local
 203                  * transaction on a DNE system. */
 204                 if (lock->l_req_mode == LCK_COS && !ldlm_is_cos_incompat(req) &&
 205                     !ldlm_is_cos_enabled(req)) {
 206                         /* jump to last lock in mode group */
 207                         tmp = mode_tail;
 208                         continue;
 209                 }
 210
 211                 if (lockmode_compat(lock->l_req_mode, req_mode)) {
 212                         /* non group locks are compatible, bits don't matter */
 213                         if (likely(req_mode != LCK_GROUP)) {
 214                                 /* jump to last lock in mode group */
 215                                 tmp = mode_tail;
 216                                 continue;
 217                         }
 218
 219                         if (req->l_policy_data.l_inodebits.li_gid ==
 220                             lock->l_policy_data.l_inodebits.li_gid) {
 221                                 if (ldlm_is_granted(lock))
 222                                         RETURN(2);
 223
 224                                 if (*ldlm_flags & LDLM_FL_BLOCK_NOWAIT)
 225                                         RETURN(-EWOULDBLOCK);
 226
 227                                 /* Place the same group together */
 228                                 ldlm_resource_insert_lock_after(lock, req);
 229                                 RETURN(0);
 230                         }
 231                 }
 232
 233                 /* GROUP locks are placed to a head of the waiting list, but
 234                  * grouped by gid. */
 235                 if (unlikely(req_mode == LCK_GROUP && !ldlm_is_granted(lock))) {
 236                         compat = 0;
 237                         if (lock->l_req_mode != LCK_GROUP) {
 238                                 /* Already not a GROUP lock, insert before. */
 239                                 ldlm_resource_insert_lock_before(lock, req);
 240                                 break;
 241                         }
 242                         /* Still GROUP but a different gid(the same gid would
 243                          * be handled above). Keep searching for the same gid */
 244                         LASSERT(req->l_policy_data.l_inodebits.li_gid !=
 245                                 lock->l_policy_data.l_inodebits.li_gid);
 246                         continue;
 247                 }
 248
 249                 for (;;) {
 250                         struct list_head *head;
 251
 252                         /* Advance loop cursor to last lock in policy group. */
 253                         tmp = &list_entry(lock->l_sl_policy.prev,
 254                                           struct ldlm_lock,
 255                                           l_sl_policy)->l_res_link;
 256
 257                         /* New lock's try_bits are filtered out by ibits
 258                          * of all locks in both granted and waiting queues.
 259                          */
 260                         *try_bits &= ~(lock->l_policy_data.l_inodebits.bits |
 261                                 lock->l_policy_data.l_inodebits.try_bits);
 262
 263                         if ((req_bits | *try_bits) == 0)
 264                                 RETURN(0);
 265
 266                         /* The new lock ibits is more preferable than try_bits
 267                          * of waiting locks so drop conflicting try_bits in
 268                          * the waiting queue.
 269                          * Notice that try_bits of granted locks must be zero.
 270                          */
 271                         lock->l_policy_data.l_inodebits.try_bits &= ~req_bits;
 272
 273                         /* Locks with overlapping bits conflict. */
 274                         if (lock->l_policy_data.l_inodebits.bits & req_bits) {
 275                                 /* COS lock mode has a special compatibility
 276                                  * requirement: it is only compatible with
 277                                  * locks from the same client. */
 278                                 if (lock->l_req_mode == LCK_COS &&
 279                                     !ldlm_is_cos_incompat(req) &&
 280                                     ldlm_is_cos_enabled(req) &&
 281                                     lock->l_client_cookie == req->l_client_cookie)
 282                                         goto skip_work_list;
 283
 284                                 compat = 0;
 285
 286                                 if (unlikely(lock->l_req_mode == LCK_GROUP)) {
 287                                         LASSERT(ldlm_has_dom(lock));
 288
 289                                         if (*ldlm_flags & LDLM_FL_BLOCK_NOWAIT)
 290                                                 RETURN(-EWOULDBLOCK);
 291
 292                                         /* Local combined DOM lock came across
 293                                          * GROUP DOM lock, it makes the thread
 294                                          * to be blocked for a long time, not
 295                                          * allowed, the trybits to be used
 296                                          * instead.
 297                                          */
 298                                         if (!req->l_export &&
 299                                             (req_bits & MDS_INODELOCK_DOM) &&
 300                                             (req_bits & ~MDS_INODELOCK_DOM))
 301                                                 LBUG();
 302
 303                                         goto skip_work_list;
 304                                 }
 305
 306                                 /* Found a conflicting policy group. */
 307                                 if (!work_list)
 308                                         RETURN(0);
 309
 310                                 /* Add locks of the policy group to @work_list
 311                                  * as blocking locks for @req */
 312                                 if (lock->l_blocking_ast)
 313                                         ldlm_add_ast_work_item(lock, req,
 314                                                                work_list);
 315                                 head = &lock->l_sl_policy;
 316                                 list_for_each_entry(lock, head, l_sl_policy)
 317                                         if (lock->l_blocking_ast)
 318                                                 ldlm_add_ast_work_item(lock,
 319                                                                 req, work_list);
 320                         }
 321 skip_work_list:
 322                         if (tmp == mode_tail)
 323                                 break;
 324
 325                         tmp = tmp->next;
 326                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 327                 } /* Loop over policy groups within one mode group. */
 328         } /* Loop over mode groups within @queue. */
 329
 330         RETURN(compat);
 331 }
 332
 333 /**
 334  * Process a granting attempt for IBITS lock.
 335  * Must be called with ns lock held
 336  *
 337  * This function looks for any conflicts for \a lock in the granted or
 338  * waiting queues. The lock is granted if no conflicts are found in
 339  * either queue.
 340  */
 341 int ldlm_process_inodebits_lock(struct ldlm_lock *lock, __u64 *ldlm_flags,
 342                                 enum ldlm_process_intention intention,
 343                                 enum ldlm_error *err,
 344                                 struct list_head *work_list)
 345 {
 346         struct ldlm_resource *res = lock->l_resource;
 347         struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ?
 348                                                         NULL : work_list;
 349         int rc, rc2 = 0;
 350         ENTRY;
 351
 352         *err = ELDLM_LOCK_ABORTED;
 353         LASSERT(!ldlm_is_granted(lock));
 354         check_res_locked(res);
 355
 356         if (intention == LDLM_PROCESS_RESCAN) {
 357                 struct list_head *bl_list =
 358                         *ldlm_flags & LDLM_FL_BLOCK_NOWAIT ? NULL : work_list;
 359
 360                 LASSERT(lock->l_policy_data.l_inodebits.bits != 0);
 361
 362                 /* It is possible that some of granted locks was not canceled
 363                  * but converted and is kept in granted queue. So there is
 364                  * a window where lock with 'ast_sent' might become granted
 365                  * again. Meanwhile a new lock may appear in that window and
 366                  * conflicts with the converted lock so the following scenario
 367                  * is possible:
 368                  *
 369                  * 1) lock1 conflicts with lock2
 370                  * 2) bl_ast was sent for lock2
 371                  * 3) lock3 comes and conflicts with lock2 too
 372                  * 4) no bl_ast sent because lock2->l_bl_ast_sent is 1
 373                  * 5) lock2 was converted for lock1 but not for lock3
 374                  * 6) lock1 granted, lock3 still is waiting for lock2, but
 375                  *    there will never be another bl_ast for that
 376                  *
 377                  * To avoid this scenario the work_list is used below to collect
 378                  * any blocked locks from granted queue during every reprocess
 379                  * and bl_ast will be sent if needed.
 380                  */
 381                 *ldlm_flags = 0;
 382                 rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
 383                                                  ldlm_flags, bl_list);
 384                 if (!rc)
 385                         RETURN(LDLM_ITER_STOP);
 386                 rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock,
 387                                                  ldlm_flags, NULL);
 388                 if (!rc)
 389                         RETURN(LDLM_ITER_STOP);
 390
 391                 /* grant also try_bits if any */
 392                 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
 393                         lock->l_policy_data.l_inodebits.bits |=
 394                                 lock->l_policy_data.l_inodebits.try_bits;
 395                         lock->l_policy_data.l_inodebits.try_bits = 0;
 396                         *ldlm_flags |= LDLM_FL_LOCK_CHANGED;
 397                 }
 398                 ldlm_resource_unlink_lock(lock);
 399                 ldlm_grant_lock(lock, grant_work);
 400
 401                 *err = ELDLM_OK;
 402                 RETURN(LDLM_ITER_CONTINUE);
 403         }
 404
 405         rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
 406                                          ldlm_flags, work_list);
 407         if (rc < 0)
 408                 GOTO(out, *err = rc);
 409
 410         if (rc != 2) {
 411                 rc2 = ldlm_inodebits_compat_queue(&res->lr_waiting, lock,
 412                                                   ldlm_flags, work_list);
 413                 if (rc2 < 0)
 414                         GOTO(out, *err = rc = rc2);
 415         }
 416
 417         if (rc + rc2 != 2) {
 418                 /* if there were only bits to try and all are conflicting */
 419                 if ((lock->l_policy_data.l_inodebits.bits |
 420                      lock->l_policy_data.l_inodebits.try_bits)) {
 421                         /* There is no sense to set LDLM_FL_NO_TIMEOUT to
 422                          * @ldlm_flags for DOM lock while they are enqueued
 423                          * through intents, i.e. @lock here is local which does
 424                          * not timeout. */
 425                         *err = ELDLM_OK;
 426                 }
 427         } else {
 428                 /* grant also all remaining try_bits */
 429                 if (lock->l_policy_data.l_inodebits.try_bits != 0) {
 430                         lock->l_policy_data.l_inodebits.bits |=
 431                                 lock->l_policy_data.l_inodebits.try_bits;
 432                         lock->l_policy_data.l_inodebits.try_bits = 0;
 433                         *ldlm_flags |= LDLM_FL_LOCK_CHANGED;
 434                 }
 435                 LASSERT(lock->l_policy_data.l_inodebits.bits);
 436                 ldlm_resource_unlink_lock(lock);
 437                 ldlm_grant_lock(lock, grant_work);
 438                 *err = ELDLM_OK;
 439         }
 440
 441         RETURN(LDLM_ITER_CONTINUE);
 442 out:
 443         return rc;
 444 }
 445 #endif /* HAVE_SERVER_SUPPORT */
 446
 447 void ldlm_ibits_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
 448                                      union ldlm_policy_data *lpolicy)
 449 {
 450         lpolicy->l_inodebits.bits = wpolicy->l_inodebits.bits;
 451         /**
 452          * try_bits and li_gid are to be handled outside of generic
 453          * write_to_local due to different behavior on a server and client.
 454          */
 455 }
 456
 457 void ldlm_ibits_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
 458                                      union ldlm_wire_policy_data *wpolicy)
 459 {
 460         memset(wpolicy, 0, sizeof(*wpolicy));
 461         wpolicy->l_inodebits.bits = lpolicy->l_inodebits.bits;
 462         wpolicy->l_inodebits.try_bits = lpolicy->l_inodebits.try_bits;
 463         wpolicy->l_inodebits.li_gid = lpolicy->l_inodebits.li_gid;
 464 }
 465
 466 /**
 467  * Attempt to convert already granted IBITS lock with several bits set to
 468  * a lock with less bits (downgrade).
 469  *
 470  * Such lock conversion is used to keep lock with non-blocking bits instead of
 471  * cancelling it, introduced for better support of DoM files.
 472  */
 473 int ldlm_inodebits_drop(struct ldlm_lock *lock, __u64 to_drop)
 474 {
 475         ENTRY;
 476
 477         check_res_locked(lock->l_resource);
 478
 479         /* Just return if there are no conflicting bits */
 480         if ((lock->l_policy_data.l_inodebits.bits & to_drop) == 0) {
 481                 LDLM_WARN(lock, "try to drop unset bits %#llx/%#llx",
 482                           lock->l_policy_data.l_inodebits.bits, to_drop);
 483                 /* nothing to do */
 484                 RETURN(0);
 485         }
 486
 487         /* remove lock from a skiplist and put in the new place
 488          * according with new inodebits */
 489         ldlm_resource_unlink_lock(lock);
 490         lock->l_policy_data.l_inodebits.bits &= ~to_drop;
 491         ldlm_grant_lock_with_skiplist(lock);
 492         RETURN(0);
 493 }
 494 EXPORT_SYMBOL(ldlm_inodebits_drop);
 495
 496 /* convert single lock */
 497 int ldlm_cli_inodebits_convert(struct ldlm_lock *lock,
 498                                enum ldlm_cancel_flags cancel_flags)
 499 {
 500         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 501         struct ldlm_lock_desc ld = { { 0 } };
 502         __u64 drop_bits, new_bits;
 503         __u32 flags = 0;
 504         int rc;
 505
 506         ENTRY;
 507
 508         check_res_locked(lock->l_resource);
 509
 510         /* Lock is being converted already */
 511         if (ldlm_is_converting(lock)) {
 512                 if (!(cancel_flags & LCF_ASYNC)) {
 513                         unlock_res_and_lock(lock);
 514                         wait_event_idle(lock->l_waitq,
 515                                         is_lock_converted(lock));
 516                         lock_res_and_lock(lock);
 517                 }
 518                 RETURN(0);
 519         }
 520
 521         /* lru_cancel may happen in parallel and call ldlm_cli_cancel_list()
 522          * independently.
 523          */
 524         if (ldlm_is_canceling(lock))
 525                 RETURN(-EINVAL);
 526
 527         /* no need in only local convert */
 528         if (lock->l_flags & (LDLM_FL_LOCAL_ONLY | LDLM_FL_CANCEL_ON_BLOCK))
 529                 RETURN(-EINVAL);
 530
 531         drop_bits = lock->l_policy_data.l_inodebits.cancel_bits;
 532         /* no cancel bits - means that caller needs full cancel */
 533         if (drop_bits == 0)
 534                 RETURN(-EINVAL);
 535
 536         new_bits = lock->l_policy_data.l_inodebits.bits & ~drop_bits;
 537         /* check if all lock bits are dropped, proceed with cancel */
 538         if (!new_bits)
 539                 RETURN(-EINVAL);
 540
 541         /* check if no dropped bits, consider this as successful convert */
 542         if (lock->l_policy_data.l_inodebits.bits == new_bits)
 543                 RETURN(0);
 544
 545         ldlm_set_converting(lock);
 546         /* Finally call cancel callback for remaining bits only.
 547          * It is important to have converting flag during that
 548          * so blocking_ast callback can distinguish convert from
 549          * cancels.
 550          */
 551         ld.l_policy_data.l_inodebits.cancel_bits = drop_bits;
 552         unlock_res_and_lock(lock);
 553         lock->l_blocking_ast(lock, &ld, lock->l_ast_data, LDLM_CB_CANCELING);
 554         /* now notify server about convert */
 555         rc = ldlm_cli_convert_req(lock, &flags, new_bits);
 556         lock_res_and_lock(lock);
 557         if (rc)
 558                 GOTO(full_cancel, rc);
 559
 560         /* Finally clear these bits in lock ibits */
 561         ldlm_inodebits_drop(lock, drop_bits);
 562
 563         /* Being locked again check if lock was canceled, it is important
 564          * to do and don't drop cbpending below
 565          */
 566         if (ldlm_is_canceling(lock))
 567                 GOTO(full_cancel, rc = -EINVAL);
 568
 569         /* also check again if more bits to be cancelled appeared */
 570         if (drop_bits != lock->l_policy_data.l_inodebits.cancel_bits)
 571                 GOTO(clear_converting, rc = -EAGAIN);
 572
 573         /* clear cbpending flag early, it is safe to match lock right after
 574          * client convert because it is downgrade always.
 575          */
 576         ldlm_clear_cbpending(lock);
 577         ldlm_clear_bl_ast(lock);
 578         spin_lock(&ns->ns_lock);
 579         if (list_empty(&lock->l_lru))
 580                 ldlm_lock_add_to_lru_nolock(lock);
 581         spin_unlock(&ns->ns_lock);
 582
 583         /* the job is done, zero the cancel_bits. If more conflicts appear,
 584          * it will result in another cycle of ldlm_cli_inodebits_convert().
 585          */
 586 full_cancel:
 587         lock->l_policy_data.l_inodebits.cancel_bits = 0;
 588 clear_converting:
 589         ldlm_clear_converting(lock);
 590         RETURN(rc);
 591 }
 592
 593 int ldlm_inodebits_alloc_lock(struct ldlm_lock *lock)
 594 {
 595         if (ldlm_is_ns_srv(lock)) {
 596                 int i;
 597
 598                 OBD_SLAB_ALLOC_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
 599                 if (lock->l_ibits_node == NULL)
 600                         return -ENOMEM;
 601                 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
 602                         INIT_LIST_HEAD(&lock->l_ibits_node->lin_link[i]);
 603                 lock->l_ibits_node->lock = lock;
 604         } else {
 605                 lock->l_ibits_node = NULL;
 606         }
 607         return 0;
 608 }
 609
 610 void ldlm_inodebits_add_lock(struct ldlm_resource *res, struct list_head *head,
 611                              struct ldlm_lock *lock, bool tail)
 612 {
 613         int i;
 614
 615         if (!ldlm_is_ns_srv(lock))
 616                 return;
 617
 618         if (head == &res->lr_waiting) {
 619                 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
 620                         if (!(lock->l_policy_data.l_inodebits.bits & BIT(i)))
 621                                 continue;
 622                         if (tail)
 623                                 list_add_tail(&lock->l_ibits_node->lin_link[i],
 624                                          &res->lr_ibits_queues->liq_waiting[i]);
 625                         else
 626                                 list_add(&lock->l_ibits_node->lin_link[i],
 627                                          &res->lr_ibits_queues->liq_waiting[i]);
 628                 }
 629         } else if (head == &res->lr_granted && lock->l_ibits_node != NULL) {
 630                 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
 631                         LASSERT(list_empty(&lock->l_ibits_node->lin_link[i]));
 632                 OBD_SLAB_FREE_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
 633                 lock->l_ibits_node = NULL;
 634         } else if (head != &res->lr_granted) {
 635                 /* we are inserting in a middle of a list, after @head */
 636                 struct ldlm_lock *orig = list_entry(head, struct ldlm_lock,
 637                                                     l_res_link);
 638                 LASSERT(orig->l_policy_data.l_inodebits.bits ==
 639                         lock->l_policy_data.l_inodebits.bits);
 640                 /* The is no a use case to insert before with exactly matched
 641                  * set of bits */
 642                 LASSERT(tail == false);
 643
 644                 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
 645                         if (!(lock->l_policy_data.l_inodebits.bits & (1 << i)))
 646                                 continue;
 647                         list_add(&lock->l_ibits_node->lin_link[i],
 648                                  &orig->l_ibits_node->lin_link[i]);
 649                 }
 650         }
 651 }
 652
 653 void ldlm_inodebits_unlink_lock(struct ldlm_lock *lock)
 654 {
 655         int i;
 656
 657         ldlm_unlink_lock_skiplist(lock);
 658         if (!ldlm_is_ns_srv(lock))
 659                 return;
 660
 661         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
 662                 list_del_init(&lock->l_ibits_node->lin_link[i]);
 663 }