lustre/ldlm/ldlm_flock.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) 2003 Hewlett-Packard Development Company LP.
  24  * Developed under the sponsorship of the US Government under
  25  * Subcontract No. B514193
  26  *
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2010, 2017, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  */
  35
  36 /**
  37  * This file implements POSIX lock type for Lustre.
  38  * Its policy properties are start and end of extent and PID.
  39  *
  40  * These locks are only done through MDS due to POSIX semantics requiring
  41  * e.g. that locks could be only partially released and as such split into
  42  * two parts, and also that two adjacent locks from the same process may be
  43  * merged into a single wider lock.
  44  *
  45  * Lock modes are mapped like this:
  46  * PR and PW for READ and WRITE locks
  47  * NL to request a releasing of a portion of the lock
  48  *
  49  * These flock locks never timeout.
  50  */
  51
  52 #define DEBUG_SUBSYSTEM S_LDLM
  53
  54 #include <linux/list.h>
  55 #include <lustre_dlm.h>
  56 #include <obd_support.h>
  57 #include <obd_class.h>
  58 #include <lustre_lib.h>
  59
  60 #include "ldlm_internal.h"
  61
  62 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
  63                             void *data, int flag);
  64
  65 static inline int
  66 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
  67 {
  68         return ((new->l_policy_data.l_flock.owner ==
  69                  lock->l_policy_data.l_flock.owner) &&
  70                 (new->l_export == lock->l_export));
  71 }
  72
  73 static inline int
  74 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
  75 {
  76         return ((new->l_policy_data.l_flock.start <=
  77                  lock->l_policy_data.l_flock.end) &&
  78                 (new->l_policy_data.l_flock.end >=
  79                  lock->l_policy_data.l_flock.start));
  80 }
  81
  82 static inline void ldlm_flock_blocking_link(struct ldlm_lock *req,
  83                                             struct ldlm_lock *lock)
  84 {
  85         /* For server only */
  86         if (req->l_export == NULL)
  87                 return;
  88
  89         LASSERT(hlist_unhashed(&req->l_exp_flock_hash));
  90
  91         req->l_policy_data.l_flock.blocking_owner =
  92                 lock->l_policy_data.l_flock.owner;
  93         req->l_policy_data.l_flock.blocking_export =
  94                 lock->l_export;
  95         atomic_set(&req->l_policy_data.l_flock.blocking_refs, 0);
  96
  97         cfs_hash_add(req->l_export->exp_flock_hash,
  98                      &req->l_policy_data.l_flock.owner,
  99                      &req->l_exp_flock_hash);
 100 }
 101
 102 static inline void ldlm_flock_blocking_unlink(struct ldlm_lock *req)
 103 {
 104         /* For server only */
 105         if (req->l_export == NULL)
 106                 return;
 107
 108         check_res_locked(req->l_resource);
 109         if (req->l_export->exp_flock_hash != NULL &&
 110             !hlist_unhashed(&req->l_exp_flock_hash))
 111                 cfs_hash_del(req->l_export->exp_flock_hash,
 112                              &req->l_policy_data.l_flock.owner,
 113                              &req->l_exp_flock_hash);
 114 }
 115
 116 static inline void
 117 ldlm_flock_destroy(struct ldlm_lock *lock, enum ldlm_mode mode, __u64 flags)
 118 {
 119         ENTRY;
 120
 121         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: %#llx)",
 122                    mode, flags);
 123
 124         /* Safe to not lock here, since it should be empty anyway */
 125         LASSERT(hlist_unhashed(&lock->l_exp_flock_hash));
 126
 127         list_del_init(&lock->l_res_link);
 128         if (flags == LDLM_FL_WAIT_NOREPROC) {
 129                 /* client side - set a flag to prevent sending a CANCEL */
 130                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
 131
 132                 /* when reaching here, it is under lock_res_and_lock(). Thus,
 133                  * need call the nolock version of ldlm_lock_decref_internal
 134                  */
 135                 ldlm_lock_decref_internal_nolock(lock, mode);
 136         }
 137
 138         ldlm_lock_destroy_nolock(lock);
 139         EXIT;
 140 }
 141
 142 #ifdef HAVE_SERVER_SUPPORT
 143 /**
 144  * POSIX locks deadlock detection code.
 145  *
 146  * Given a new lock \a req and an existing lock \a bl_lock it conflicts
 147  * with, we need to iterate through all blocked POSIX locks for this
 148  * export and see if there is a deadlock condition arising. (i.e. when
 149  * one client holds a lock on something and want a lock on something
 150  * else and at the same time another client has the opposite situation).
 151  */
 152
 153 struct ldlm_flock_lookup_cb_data {
 154         __u64 *bl_owner;
 155         struct ldlm_lock *lock;
 156         struct obd_export *exp;
 157 };
 158
 159 static int ldlm_flock_lookup_cb(struct obd_export *exp, void *data)
 160 {
 161         struct ldlm_flock_lookup_cb_data *cb_data = data;
 162         struct ldlm_lock *lock;
 163
 164         if (exp->exp_failed)
 165                 return 0;
 166
 167         lock = cfs_hash_lookup(exp->exp_flock_hash, cb_data->bl_owner);
 168         if (lock == NULL)
 169                 return 0;
 170
 171         /* Stop on first found lock. Same process can't sleep twice */
 172         cb_data->lock = lock;
 173         cb_data->exp = class_export_get(exp);
 174
 175         return 1;
 176 }
 177
 178 static int
 179 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *bl_lock)
 180 {
 181         struct obd_export *req_exp = req->l_export;
 182         struct obd_export *bl_exp = bl_lock->l_export;
 183         __u64 req_owner = req->l_policy_data.l_flock.owner;
 184         __u64 bl_owner = bl_lock->l_policy_data.l_flock.owner;
 185
 186         /* For server only */
 187         if (req_exp == NULL)
 188                 return 0;
 189
 190         class_export_get(bl_exp);
 191         while (1) {
 192                 struct ldlm_flock_lookup_cb_data cb_data = {
 193                         .bl_owner = &bl_owner,
 194                         .lock = NULL,
 195                         .exp = NULL,
 196                 };
 197                 struct ptlrpc_connection *bl_exp_conn;
 198                 struct obd_export *bl_exp_new;
 199                 struct ldlm_lock *lock = NULL;
 200                 struct ldlm_flock *flock;
 201
 202                 bl_exp_conn = bl_exp->exp_connection;
 203                 if (bl_exp->exp_flock_hash != NULL) {
 204                         int found;
 205
 206                         found = obd_nid_export_for_each(bl_exp->exp_obd,
 207                                                         &bl_exp_conn->c_peer.nid,
 208                                                         ldlm_flock_lookup_cb,
 209                                                         &cb_data);
 210                         if (found)
 211                                 lock = cb_data.lock;
 212                 }
 213                 if (lock == NULL)
 214                         break;
 215
 216                 class_export_put(bl_exp);
 217                 bl_exp = cb_data.exp;
 218
 219                 LASSERT(req != lock);
 220                 flock = &lock->l_policy_data.l_flock;
 221                 LASSERT(flock->owner == bl_owner);
 222                 bl_owner = flock->blocking_owner;
 223                 bl_exp_new = class_export_get(flock->blocking_export);
 224                 class_export_put(bl_exp);
 225
 226                 cfs_hash_put(bl_exp->exp_flock_hash, &lock->l_exp_flock_hash);
 227                 bl_exp = bl_exp_new;
 228
 229                 if (bl_exp->exp_failed)
 230                         break;
 231
 232                 if (bl_owner == req_owner &&
 233                     nid_same(&bl_exp_conn->c_peer.nid,
 234                               &req_exp->exp_connection->c_peer.nid)) {
 235                         class_export_put(bl_exp);
 236                         return 1;
 237                 }
 238         }
 239         class_export_put(bl_exp);
 240
 241         return 0;
 242 }
 243
 244 static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
 245                                           struct list_head *work_list)
 246 {
 247         CDEBUG(D_INFO, "reprocess deadlock req=%p\n", lock);
 248
 249         if ((exp_connect_flags(lock->l_export) &
 250              OBD_CONNECT_FLOCK_DEAD) == 0) {
 251                 CERROR("deadlock found, but client doesn't support flock canceliation\n");
 252         } else {
 253                 LASSERT(lock->l_completion_ast);
 254                 LASSERT(!ldlm_is_ast_sent(lock));
 255                 lock->l_flags |= (LDLM_FL_AST_SENT | LDLM_FL_CANCEL_ON_BLOCK |
 256                                   LDLM_FL_FLOCK_DEADLOCK);
 257                 ldlm_flock_blocking_unlink(lock);
 258                 ldlm_resource_unlink_lock(lock);
 259                 ldlm_add_ast_work_item(lock, NULL, work_list);
 260         }
 261 }
 262 #endif /* HAVE_SERVER_SUPPORT */
 263
 264 /**
 265  * Process a granting attempt for flock lock.
 266  * Must be called under ns lock held.
 267  *
 268  * This function looks for any conflicts for \a lock in the granted or
 269  * waiting queues. The lock is granted if no conflicts are found in
 270  * either queue.
 271  */
 272 int
 273 ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags,
 274                         enum ldlm_process_intention intention,
 275                         enum ldlm_error *err, struct list_head *work_list)
 276 {
 277         struct ldlm_resource *res = req->l_resource;
 278         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
 279         struct ldlm_lock *tmp;
 280         struct ldlm_lock *ownlocks = NULL;
 281         struct ldlm_lock *lock = NULL;
 282         struct ldlm_lock *new = req;
 283         struct ldlm_lock *new2 = NULL;
 284         enum ldlm_mode mode = req->l_req_mode;
 285         int local = ns_is_client(ns);
 286         int added = (mode == LCK_NL);
 287         int splitted = 0;
 288         const struct ldlm_callback_suite null_cbs = { NULL };
 289 #ifdef HAVE_SERVER_SUPPORT
 290         struct list_head *grant_work = (intention == LDLM_PROCESS_ENQUEUE ?
 291                                         NULL : work_list);
 292 #endif
 293
 294         ENTRY;
 295         CDEBUG(D_DLMTRACE, "flags %#llx owner %llu pid %u mode %u start "
 296                "%llu end %llu\n", *flags,
 297                new->l_policy_data.l_flock.owner,
 298                new->l_policy_data.l_flock.pid, mode,
 299                req->l_policy_data.l_flock.start,
 300                req->l_policy_data.l_flock.end);
 301
 302         *err = ELDLM_OK;
 303
 304         if (local) {
 305                 /* No blocking ASTs are sent to the clients for
 306                  * Posix file & record locks
 307                  */
 308                 req->l_blocking_ast = NULL;
 309         } else {
 310                 /* Called on the server for lock cancels. */
 311                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 312         }
 313
 314 reprocess:
 315         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 316                 /* This loop determines where this processes locks start
 317                  * in the resource lr_granted list.
 318                  */
 319                 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
 320                         if (ldlm_same_flock_owner(lock, req)) {
 321                                 ownlocks = lock;
 322                                 break;
 323                         }
 324                 }
 325         }
 326 #ifdef HAVE_SERVER_SUPPORT
 327         else {
 328                 int reprocess_failed = 0;
 329                 lockmode_verify(mode);
 330
 331                 /* This loop determines if there are existing locks
 332                  * that conflict with the new lock request.
 333                  */
 334                 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
 335                         if (ldlm_same_flock_owner(lock, req)) {
 336                                 if (!ownlocks)
 337                                         ownlocks = lock;
 338                                 continue;
 339                         }
 340
 341                         if (req->l_req_mode == LCK_PR &&
 342                             lock->l_granted_mode == LCK_PR &&
 343                             lock->l_policy_data.l_flock.start <=
 344                                 req->l_policy_data.l_flock.start &&
 345                             lock->l_policy_data.l_flock.end >=
 346                                 req->l_policy_data.l_flock.end) {
 347                                 /* there can't be granted WR lock */
 348                                 break;
 349                         }
 350                         /* locks are compatible, overlap doesn't matter */
 351                         if (lockmode_compat(lock->l_granted_mode, mode))
 352                                 continue;
 353
 354                         if (!ldlm_flocks_overlap(lock, req))
 355                                 continue;
 356
 357                         if (intention != LDLM_PROCESS_ENQUEUE) {
 358                                 if (ldlm_flock_deadlock(req, lock)) {
 359                                         ldlm_flock_cancel_on_deadlock(
 360                                                 req, grant_work);
 361                                         RETURN(LDLM_ITER_CONTINUE);
 362                                 }
 363                                 reprocess_failed = 1;
 364                                 break;
 365                         }
 366
 367                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 368                                 ldlm_flock_destroy(req, mode, *flags);
 369                                 *err = -EAGAIN;
 370                                 RETURN(LDLM_ITER_STOP);
 371                         }
 372
 373                         if (*flags & LDLM_FL_TEST_LOCK) {
 374                                 ldlm_flock_destroy(req, mode, *flags);
 375                                 req->l_req_mode = lock->l_granted_mode;
 376                                 req->l_policy_data.l_flock.pid =
 377                                         lock->l_policy_data.l_flock.pid;
 378                                 req->l_policy_data.l_flock.start =
 379                                         lock->l_policy_data.l_flock.start;
 380                                 req->l_policy_data.l_flock.end =
 381                                         lock->l_policy_data.l_flock.end;
 382                                 *flags |= LDLM_FL_LOCK_CHANGED;
 383                                 RETURN(LDLM_ITER_STOP);
 384                         }
 385
 386                         /* add lock to blocking list before deadlock
 387                          * check to prevent race
 388                          */
 389                         ldlm_flock_blocking_link(req, lock);
 390
 391                         if (ldlm_flock_deadlock(req, lock)) {
 392                                 ldlm_flock_blocking_unlink(req);
 393                                 ldlm_flock_destroy(req, mode, *flags);
 394                                 *err = -EDEADLK;
 395                                 RETURN(LDLM_ITER_STOP);
 396                         }
 397
 398                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 399                         *flags |= LDLM_FL_BLOCK_GRANTED;
 400                         RETURN(LDLM_ITER_STOP);
 401                 }
 402                 if (reprocess_failed)
 403                         RETURN(LDLM_ITER_CONTINUE);
 404         }
 405
 406         if (*flags & LDLM_FL_TEST_LOCK) {
 407                 ldlm_flock_destroy(req, mode, *flags);
 408                 req->l_req_mode = LCK_NL;
 409                 *flags |= LDLM_FL_LOCK_CHANGED;
 410                 RETURN(LDLM_ITER_STOP);
 411         }
 412
 413         /* In case we had slept on this lock request take it off of the
 414          * deadlock detection hash list.
 415          */
 416         ldlm_flock_blocking_unlink(req);
 417 #endif /* HAVE_SERVER_SUPPORT */
 418
 419         /* Scan the locks owned by this process to find the insertion point
 420          * (as locks are ordered), and to handle overlaps.
 421          * We may have to merge or split existing locks.
 422          */
 423         if (ownlocks)
 424                 lock = ownlocks;
 425         else
 426                 lock = list_entry(&res->lr_granted,
 427                                   struct ldlm_lock, l_res_link);
 428         list_for_each_entry_safe_from(lock, tmp, &res->lr_granted, l_res_link) {
 429                 if (!ldlm_same_flock_owner(lock, new))
 430                         break;
 431
 432                 if (lock->l_granted_mode == mode) {
 433                         /* If the modes are the same then we need to process
 434                          * locks that overlap OR adjoin the new lock. The extra
 435                          * logic condition is necessary to deal with arithmetic
 436                          * overflow and underflow.
 437                          */
 438                         if ((new->l_policy_data.l_flock.start >
 439                              (lock->l_policy_data.l_flock.end + 1))
 440                             && (lock->l_policy_data.l_flock.end !=
 441                                 OBD_OBJECT_EOF))
 442                                 continue;
 443
 444                         if ((new->l_policy_data.l_flock.end <
 445                              (lock->l_policy_data.l_flock.start - 1))
 446                             && (lock->l_policy_data.l_flock.start != 0))
 447                                 break;
 448
 449                         if (new->l_policy_data.l_flock.start <
 450                             lock->l_policy_data.l_flock.start) {
 451                                 lock->l_policy_data.l_flock.start =
 452                                         new->l_policy_data.l_flock.start;
 453                         } else {
 454                                 new->l_policy_data.l_flock.start =
 455                                         lock->l_policy_data.l_flock.start;
 456                         }
 457
 458                         if (new->l_policy_data.l_flock.end >
 459                             lock->l_policy_data.l_flock.end) {
 460                                 lock->l_policy_data.l_flock.end =
 461                                         new->l_policy_data.l_flock.end;
 462                         } else {
 463                                 new->l_policy_data.l_flock.end =
 464                                         lock->l_policy_data.l_flock.end;
 465                         }
 466
 467                         if (added) {
 468                                 ldlm_flock_destroy(lock, mode, *flags);
 469                         } else {
 470                                 new = lock;
 471                                 added = 1;
 472                         }
 473                         continue;
 474                 }
 475
 476                 if (new->l_policy_data.l_flock.start >
 477                     lock->l_policy_data.l_flock.end)
 478                         continue;
 479
 480                 if (new->l_policy_data.l_flock.end <
 481                     lock->l_policy_data.l_flock.start)
 482                         break;
 483
 484                 res->lr_flock_node.lfn_needs_reprocess = true;
 485
 486                 if (new->l_policy_data.l_flock.start <=
 487                     lock->l_policy_data.l_flock.start) {
 488                         if (new->l_policy_data.l_flock.end <
 489                             lock->l_policy_data.l_flock.end) {
 490                                 lock->l_policy_data.l_flock.start =
 491                                         new->l_policy_data.l_flock.end + 1;
 492                                 break;
 493                         }
 494                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 495                         continue;
 496                 }
 497                 if (new->l_policy_data.l_flock.end >=
 498                     lock->l_policy_data.l_flock.end) {
 499                         lock->l_policy_data.l_flock.end =
 500                                 new->l_policy_data.l_flock.start - 1;
 501                         continue;
 502                 }
 503
 504                 /* split the existing lock into two locks */
 505
 506                 /* if this is an F_UNLCK operation then we could avoid
 507                  * allocating a new lock and use the req lock passed in
 508                  * with the request but this would complicate the reply
 509                  * processing since updates to req get reflected in the
 510                  * reply. The client side replays the lock request so
 511                  * it must see the original lock data in the reply.
 512                  */
 513
 514                 /* XXX - if ldlm_lock_new() can sleep we should
 515                  * release the lr_lock, allocate the new lock,
 516                  * and restart processing this lock.
 517                  */
 518                 if (new2 == NULL) {
 519                         unlock_res_and_lock(req);
 520                         new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
 521                                                 lock->l_granted_mode, &null_cbs,
 522                                                 NULL, 0, LVB_T_NONE);
 523                         lock_res_and_lock(req);
 524                         if (IS_ERR(new2)) {
 525                                 ldlm_flock_destroy(req, lock->l_granted_mode,
 526                                                    *flags);
 527                                 *err = PTR_ERR(new2);
 528                                 RETURN(LDLM_ITER_STOP);
 529                         }
 530                         goto reprocess;
 531                 }
 532
 533                 splitted = 1;
 534
 535                 new2->l_granted_mode = lock->l_granted_mode;
 536                 new2->l_policy_data.l_flock.pid =
 537                         new->l_policy_data.l_flock.pid;
 538                 new2->l_policy_data.l_flock.owner =
 539                         new->l_policy_data.l_flock.owner;
 540                 new2->l_policy_data.l_flock.start =
 541                         lock->l_policy_data.l_flock.start;
 542                 new2->l_policy_data.l_flock.end =
 543                         new->l_policy_data.l_flock.start - 1;
 544                 lock->l_policy_data.l_flock.start =
 545                         new->l_policy_data.l_flock.end + 1;
 546                 new2->l_conn_export = lock->l_conn_export;
 547                 if (lock->l_export != NULL) {
 548                         new2->l_export = class_export_lock_get(lock->l_export,
 549                                                                new2);
 550                         if (new2->l_export->exp_lock_hash &&
 551                             hlist_unhashed(&new2->l_exp_hash))
 552                                 cfs_hash_add(new2->l_export->exp_lock_hash,
 553                                              &new2->l_remote_handle,
 554                                              &new2->l_exp_hash);
 555                 }
 556                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 557                         ldlm_lock_addref_internal_nolock(new2,
 558                                                          lock->l_granted_mode);
 559
 560                 /* insert new2 at lock */
 561                 ldlm_resource_add_lock(res, &lock->l_res_link, new2);
 562                 LDLM_LOCK_RELEASE(new2);
 563                 break;
 564         }
 565
 566         /* if new2 is created but never used, destroy it*/
 567         if (splitted == 0 && new2 != NULL)
 568                 ldlm_lock_destroy_nolock(new2);
 569
 570         /* At this point we're granting the lock request. */
 571         req->l_granted_mode = req->l_req_mode;
 572
 573         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 574         if (!added) {
 575                 list_del_init(&req->l_res_link);
 576                 /* insert new lock before "lock", which might be the
 577                  * next lock for this owner, or might be the first
 578                  * lock for the next owner, or might not be a lock at
 579                  * all, but instead points at the head of the list
 580                  */
 581                 ldlm_resource_add_lock(res, &lock->l_res_link, req);
 582         }
 583
 584         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 585 #ifdef HAVE_SERVER_SUPPORT
 586                 if (intention == LDLM_PROCESS_ENQUEUE) {
 587                         /* If this is an unlock, reprocess the waitq and
 588                          * send completions ASTs for locks that can now be
 589                          * granted. The only problem with doing this
 590                          * reprocessing here is that the completion ASTs for
 591                          * newly granted locks will be sent before the unlock
 592                          * completion is sent. It shouldn't be an issue. Also
 593                          * note that ldlm_process_flock_lock() will recurse,
 594                          * but only once because 'intention' won't be
 595                          * LDLM_PROCESS_ENQUEUE from ldlm_reprocess_queue.
 596                          */
 597                         struct ldlm_flock_node *fn = &res->lr_flock_node;
 598 restart:
 599                         if (mode == LCK_NL && fn->lfn_needs_reprocess &&
 600                             atomic_read(&fn->lfn_unlock_pending) == 0) {
 601                                 LIST_HEAD(rpc_list);
 602                                 int rc;
 603
 604                                 ldlm_reprocess_queue(res, &res->lr_waiting,
 605                                                      &rpc_list,
 606                                                      LDLM_PROCESS_RESCAN, 0);
 607                                 fn->lfn_needs_reprocess = false;
 608                                 unlock_res_and_lock(req);
 609                                 rc = ldlm_run_ast_work(ns, &rpc_list,
 610                                                        LDLM_WORK_CP_AST);
 611                                 lock_res_and_lock(req);
 612                                 if (rc == -ERESTART) {
 613                                         fn->lfn_needs_reprocess = true;
 614                                         GOTO(restart, rc);
 615                                 }
 616                         }
 617                 } else {
 618                         LASSERT(req->l_completion_ast);
 619                         ldlm_add_ast_work_item(req, NULL, grant_work);
 620                 }
 621 #else /* !HAVE_SERVER_SUPPORT */
 622                 /* The only one possible case for client-side calls flock
 623                  * policy function is ldlm_flock_completion_ast inside which
 624                  * carries LDLM_FL_WAIT_NOREPROC flag.
 625                  */
 626                 CERROR("Illegal parameter for client-side-only module.\n");
 627                 LBUG();
 628 #endif /* HAVE_SERVER_SUPPORT */
 629         }
 630
 631         /* In case we're reprocessing the requested lock we can't destroy
 632          * it until after calling ldlm_add_ast_work_item() above so that laawi()
 633          * can bump the reference count on \a req. Otherwise \a req
 634          * could be freed before the completion AST can be sent.
 635          */
 636         if (added)
 637                 ldlm_flock_destroy(req, mode, *flags);
 638
 639         ldlm_resource_dump(D_INFO, res);
 640         RETURN(LDLM_ITER_CONTINUE);
 641 }
 642
 643 /**
 644  * Flock completion callback function.
 645  *
 646  * \param lock [in,out]: A lock to be handled
 647  * \param flags    [in]: flags
 648  * \param *data    [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
 649  *
 650  * \retval 0    : success
 651  * \retval <0   : failure
 652  */
 653 int
 654 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 655 {
 656         struct file_lock *getlk = lock->l_ast_data;
 657         struct obd_device *obd;
 658         enum ldlm_error err;
 659         int rc = 0;
 660         ENTRY;
 661
 662         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4);
 663         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) {
 664                 lock_res_and_lock(lock);
 665                 lock->l_flags |= LDLM_FL_FAIL_LOC;
 666                 unlock_res_and_lock(lock);
 667                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT3, 4);
 668         }
 669         CDEBUG(D_DLMTRACE, "flags: %#llx data: %p getlk: %p\n",
 670                flags, data, getlk);
 671
 672         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 673
 674         if (flags & LDLM_FL_FAILED)
 675                 goto granted;
 676
 677         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
 678                 if (NULL == data)
 679                         /* mds granted the lock in the reply */
 680                         goto granted;
 681                 /* CP AST RPC: lock get granted, wake it up */
 682                 wake_up(&lock->l_waitq);
 683                 RETURN(0);
 684         }
 685
 686         LDLM_DEBUG(lock,
 687                    "client-side enqueue returned a blocked lock, sleeping");
 688         obd = class_exp2obd(lock->l_conn_export);
 689
 690         /* Go to sleep until the lock is granted. */
 691         rc = l_wait_event_abortable(lock->l_waitq,
 692                                     is_granted_or_cancelled(lock));
 693         if (rc < 0) {
 694                 /* take lock off the deadlock detection hash list. */
 695                 lock_res_and_lock(lock);
 696                 ldlm_flock_blocking_unlink(lock);
 697
 698                 /* client side - set flag to prevent lock from being
 699                  * put on LRU list
 700                  */
 701                 ldlm_set_cbpending(lock);
 702                 unlock_res_and_lock(lock);
 703
 704                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 705                            rc);
 706                 RETURN(rc);
 707         }
 708
 709 granted:
 710         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
 711
 712         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT4)) {
 713                 lock_res_and_lock(lock);
 714                 /* DEADLOCK is always set with CBPENDING */
 715                 lock->l_flags |= LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
 716                 unlock_res_and_lock(lock);
 717                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT4, 4);
 718         }
 719         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT5)) {
 720                 lock_res_and_lock(lock);
 721                 /* DEADLOCK is always set with CBPENDING */
 722                 lock->l_flags |= (LDLM_FL_FAIL_LOC |
 723                                   LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING);
 724                 unlock_res_and_lock(lock);
 725                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT5, 4);
 726         }
 727
 728         lock_res_and_lock(lock);
 729
 730
 731         /* Protect against race where lock could have been just destroyed
 732          * due to overlap in ldlm_process_flock_lock().
 733          */
 734         if (ldlm_is_destroyed(lock)) {
 735                 unlock_res_and_lock(lock);
 736                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 737
 738                 /* An error is still to be returned, to propagate it up to
 739                  * ldlm_cli_enqueue_fini() caller. */
 740                 RETURN(-EIO);
 741         }
 742
 743         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 744         ldlm_resource_unlink_lock(lock);
 745
 746         /* Import invalidation. We need to actually release the lock
 747          * references being held, so that it can go away. No point in
 748          * holding the lock even if app still believes it has it, since
 749          * server already dropped it anyway. Only for granted locks too.
 750          */
 751         /* Do the same for DEADLOCK'ed locks. */
 752         if (ldlm_is_failed(lock) || ldlm_is_flock_deadlock(lock)) {
 753                 int mode;
 754
 755                 if (flags & LDLM_FL_TEST_LOCK)
 756                         LASSERT(ldlm_is_test_lock(lock));
 757
 758                 if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock))
 759                         mode = getlk->fl_type;
 760                 else
 761                         mode = lock->l_req_mode;
 762
 763                 if (ldlm_is_flock_deadlock(lock)) {
 764                         LDLM_DEBUG(lock, "client-side enqueue deadlock "
 765                                    "received");
 766                         rc = -EDEADLK;
 767                 }
 768                 ldlm_flock_destroy(lock, mode, LDLM_FL_WAIT_NOREPROC);
 769                 unlock_res_and_lock(lock);
 770
 771                 /* Need to wake up the waiter if we were evicted */
 772                 wake_up(&lock->l_waitq);
 773
 774                 /* An error is still to be returned, to propagate it up to
 775                  * ldlm_cli_enqueue_fini() caller.
 776                  */
 777                 RETURN(rc ? : -EIO);
 778         }
 779
 780         LDLM_DEBUG(lock, "client-side enqueue granted");
 781
 782         if (flags & LDLM_FL_TEST_LOCK) {
 783                 /*
 784                  * fcntl(F_GETLK) request
 785                  * The old mode was saved in getlk->fl_type so that if the mode
 786                  * in the lock changes we can decref the appropriate refcount.
 787                  */
 788                 LASSERT(ldlm_is_test_lock(lock));
 789                 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
 790                 switch (lock->l_granted_mode) {
 791                 case LCK_PR:
 792                         getlk->fl_type = F_RDLCK;
 793                         break;
 794                 case LCK_PW:
 795                         getlk->fl_type = F_WRLCK;
 796                         break;
 797                 default:
 798                         getlk->fl_type = F_UNLCK;
 799                 }
 800                 getlk->fl_pid = (pid_t)lock->l_policy_data.l_flock.pid;
 801                 getlk->fl_start = (loff_t)lock->l_policy_data.l_flock.start;
 802                 getlk->fl_end = (loff_t)lock->l_policy_data.l_flock.end;
 803         } else {
 804                 __u64 noreproc = LDLM_FL_WAIT_NOREPROC;
 805
 806                 /* We need to reprocess the lock to do merges or splits
 807                  * with existing locks owned by this process.
 808                  */
 809                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 810         }
 811         unlock_res_and_lock(lock);
 812         RETURN(rc);
 813 }
 814 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 815
 816 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 817                             void *data, int flag)
 818 {
 819         ENTRY;
 820
 821         LASSERT(lock);
 822         LASSERT(flag == LDLM_CB_CANCELING);
 823
 824         /* take lock off the deadlock detection hash list. */
 825         lock_res_and_lock(lock);
 826         ldlm_flock_blocking_unlink(lock);
 827         unlock_res_and_lock(lock);
 828         RETURN(0);
 829 }
 830
 831 void ldlm_flock_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
 832                                      union ldlm_policy_data *lpolicy)
 833 {
 834         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
 835         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
 836         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
 837         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
 838 }
 839
 840 void ldlm_flock_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
 841                                      union ldlm_wire_policy_data *wpolicy)
 842 {
 843         memset(wpolicy, 0, sizeof(*wpolicy));
 844         wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
 845         wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
 846         wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
 847         wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
 848 }
 849
 850 /*
 851  * Export handle<->flock hash operations.
 852  */
 853 static unsigned
 854 ldlm_export_flock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
 855 {
 856         return cfs_hash_64(*(__u64 *)key, 0) & mask;
 857 }
 858
 859 static void *
 860 ldlm_export_flock_key(struct hlist_node *hnode)
 861 {
 862         struct ldlm_lock *lock;
 863
 864         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 865         return &lock->l_policy_data.l_flock.owner;
 866 }
 867
 868 static int
 869 ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode)
 870 {
 871         return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
 872 }
 873
 874 static void *
 875 ldlm_export_flock_object(struct hlist_node *hnode)
 876 {
 877         return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 878 }
 879
 880 static void
 881 ldlm_export_flock_get(struct cfs_hash *hs, struct hlist_node *hnode)
 882 {
 883         struct ldlm_lock *lock;
 884         struct ldlm_flock *flock;
 885
 886         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 887         LDLM_LOCK_GET(lock);
 888
 889         flock = &lock->l_policy_data.l_flock;
 890         LASSERT(flock->blocking_export != NULL);
 891         class_export_get(flock->blocking_export);
 892         atomic_inc(&flock->blocking_refs);
 893 }
 894
 895 static void
 896 ldlm_export_flock_put(struct cfs_hash *hs, struct hlist_node *hnode)
 897 {
 898         struct ldlm_lock *lock;
 899         struct ldlm_flock *flock;
 900
 901         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 902
 903         flock = &lock->l_policy_data.l_flock;
 904         LASSERT(flock->blocking_export != NULL);
 905         class_export_put(flock->blocking_export);
 906         if (atomic_dec_and_test(&flock->blocking_refs)) {
 907                 flock->blocking_owner = 0;
 908                 flock->blocking_export = NULL;
 909         }
 910         LDLM_LOCK_RELEASE(lock);
 911 }
 912
 913 static struct cfs_hash_ops ldlm_export_flock_ops = {
 914         .hs_hash        = ldlm_export_flock_hash,
 915         .hs_key         = ldlm_export_flock_key,
 916         .hs_keycmp      = ldlm_export_flock_keycmp,
 917         .hs_object      = ldlm_export_flock_object,
 918         .hs_get         = ldlm_export_flock_get,
 919         .hs_put         = ldlm_export_flock_put,
 920         .hs_put_locked  = ldlm_export_flock_put,
 921 };
 922
 923 int ldlm_init_flock_export(struct obd_export *exp)
 924 {
 925         if( strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0)
 926                 RETURN(0);
 927
 928         exp->exp_flock_hash =
 929                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
 930                                 HASH_EXP_LOCK_CUR_BITS,
 931                                 HASH_EXP_LOCK_MAX_BITS,
 932                                 HASH_EXP_LOCK_BKT_BITS, 0,
 933                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 934                                 &ldlm_export_flock_ops,
 935                                 CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE);
 936         if (!exp->exp_flock_hash)
 937                 RETURN(-ENOMEM);
 938
 939         RETURN(0);
 940 }
 941
 942 void ldlm_destroy_flock_export(struct obd_export *exp)
 943 {
 944         ENTRY;
 945         if (exp->exp_flock_hash) {
 946                 cfs_hash_putref(exp->exp_flock_hash);
 947                 exp->exp_flock_hash = NULL;
 948         }
 949         EXIT;
 950 }