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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003 Hewlett-Packard Development Company LP.
  28  * Developed under the sponsorship of the US Government under
  29  * Subcontract No. B514193
  30  *
  31  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  32  * Use is subject to license terms.
  33  *
  34  * Copyright (c) 2010, 2012, Intel Corporation.
  35  */
  36 /*
  37  * This file is part of Lustre, http://www.lustre.org/
  38  * Lustre is a trademark of Sun Microsystems, Inc.
  39  */
  40
  41 /**
  42  * This file implements POSIX lock type for Lustre.
  43  * Its policy properties are start and end of extent and PID.
  44  *
  45  * These locks are only done through MDS due to POSIX semantics requiring
  46  * e.g. that locks could be only partially released and as such split into
  47  * two parts, and also that two adjacent locks from the same process may be
  48  * merged into a single wider lock.
  49  *
  50  * Lock modes are mapped like this:
  51  * PR and PW for READ and WRITE locks
  52  * NL to request a releasing of a portion of the lock
  53  *
  54  * These flock locks never timeout.
  55  */
  56
  57 #define DEBUG_SUBSYSTEM S_LDLM
  58
  59 #ifdef __KERNEL__
  60 #include <lustre_dlm.h>
  61 #include <obd_support.h>
  62 #include <obd_class.h>
  63 #include <lustre_lib.h>
  64 #include <libcfs/list.h>
  65 #else
  66 #include <liblustre.h>
  67 #include <obd_class.h>
  68 #endif
  69
  70 #include "ldlm_internal.h"
  71
  72 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
  73                             void *data, int flag);
  74
  75 /**
  76  * list_for_remaining_safe - iterate over the remaining entries in a list
  77  *              and safeguard against removal of a list entry.
  78  * \param pos   the &struct list_head to use as a loop counter. pos MUST
  79  *              have been initialized prior to using it in this macro.
  80  * \param n     another &struct list_head to use as temporary storage
  81  * \param head  the head for your list.
  82  */
  83 #define list_for_remaining_safe(pos, n, head) \
  84         for (n = pos->next; pos != (head); pos = n, n = pos->next)
  85
  86 static inline int
  87 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
  88 {
  89         return((new->l_policy_data.l_flock.owner ==
  90                 lock->l_policy_data.l_flock.owner) &&
  91                (new->l_export == lock->l_export));
  92 }
  93
  94 static inline int
  95 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
  96 {
  97         return((new->l_policy_data.l_flock.start <=
  98                 lock->l_policy_data.l_flock.end) &&
  99                (new->l_policy_data.l_flock.end >=
 100                 lock->l_policy_data.l_flock.start));
 101 }
 102
 103 static inline void ldlm_flock_blocking_link(struct ldlm_lock *req,
 104                                             struct ldlm_lock *lock)
 105 {
 106         /* For server only */
 107         if (req->l_export == NULL)
 108                 return;
 109
 110         LASSERT(cfs_hlist_unhashed(&req->l_exp_flock_hash));
 111
 112         req->l_policy_data.l_flock.blocking_owner =
 113                 lock->l_policy_data.l_flock.owner;
 114         req->l_policy_data.l_flock.blocking_export =
 115                 lock->l_export;
 116         req->l_policy_data.l_flock.blocking_refs = 0;
 117
 118         cfs_hash_add(req->l_export->exp_flock_hash,
 119                      &req->l_policy_data.l_flock.owner,
 120                      &req->l_exp_flock_hash);
 121 }
 122
 123 static inline void ldlm_flock_blocking_unlink(struct ldlm_lock *req)
 124 {
 125         /* For server only */
 126         if (req->l_export == NULL)
 127                 return;
 128
 129         check_res_locked(req->l_resource);
 130         if (req->l_export->exp_flock_hash != NULL &&
 131             !cfs_hlist_unhashed(&req->l_exp_flock_hash))
 132                 cfs_hash_del(req->l_export->exp_flock_hash,
 133                              &req->l_policy_data.l_flock.owner,
 134                              &req->l_exp_flock_hash);
 135 }
 136
 137 static inline void
 138 ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, __u64 flags)
 139 {
 140         ENTRY;
 141
 142         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%llx)",
 143                    mode, flags);
 144
 145         /* Safe to not lock here, since it should be empty anyway */
 146         LASSERT(cfs_hlist_unhashed(&lock->l_exp_flock_hash));
 147
 148         cfs_list_del_init(&lock->l_res_link);
 149         if (flags == LDLM_FL_WAIT_NOREPROC &&
 150             !(lock->l_flags & LDLM_FL_FAILED)) {
 151                 /* client side - set a flag to prevent sending a CANCEL */
 152                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
 153
 154                 /* when reaching here, it is under lock_res_and_lock(). Thus,
 155                    need call the nolock version of ldlm_lock_decref_internal*/
 156                 ldlm_lock_decref_internal_nolock(lock, mode);
 157         }
 158
 159         ldlm_lock_destroy_nolock(lock);
 160         EXIT;
 161 }
 162
 163 /**
 164  * POSIX locks deadlock detection code.
 165  *
 166  * Given a new lock \a req and an existing lock \a bl_lock it conflicts
 167  * with, we need to iterate through all blocked POSIX locks for this
 168  * export and see if there is a deadlock condition arising. (i.e. when
 169  * one client holds a lock on something and want a lock on something
 170  * else and at the same time another client has the opposite situation).
 171  */
 172
 173 struct ldlm_flock_lookup_cb_data {
 174         __u64 *bl_owner;
 175         struct ldlm_lock *lock;
 176         struct obd_export *exp;
 177 };
 178
 179 static int ldlm_flock_lookup_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 180                                 cfs_hlist_node_t *hnode, void *data)
 181 {
 182         struct ldlm_flock_lookup_cb_data *cb_data = data;
 183         struct obd_export *exp = cfs_hash_object(hs, hnode);
 184         struct ldlm_lock *lock;
 185
 186         lock = cfs_hash_lookup(exp->exp_flock_hash, cb_data->bl_owner);
 187         if (lock == NULL)
 188                 return 0;
 189
 190         /* Stop on first found lock. Same process can't sleep twice */
 191         cb_data->lock = lock;
 192         cb_data->exp = class_export_get(exp);
 193
 194         return 1;
 195 }
 196
 197 static int
 198 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *bl_lock)
 199 {
 200         struct obd_export *req_exp = req->l_export;
 201         struct obd_export *bl_exp = bl_lock->l_export;
 202         __u64 req_owner = req->l_policy_data.l_flock.owner;
 203         __u64 bl_owner = bl_lock->l_policy_data.l_flock.owner;
 204
 205         /* For server only */
 206         if (req_exp == NULL)
 207                 return 0;
 208
 209         class_export_get(bl_exp);
 210         while (1) {
 211                 struct ldlm_flock_lookup_cb_data cb_data = {
 212                                         .bl_owner = &bl_owner,
 213                                         .lock = NULL,
 214                                         .exp = NULL };
 215                 struct obd_export *bl_exp_new;
 216                 struct ldlm_lock *lock = NULL;
 217                 struct ldlm_flock *flock;
 218
 219                 if (bl_exp->exp_flock_hash != NULL) {
 220                         cfs_hash_for_each_key(bl_exp->exp_obd->obd_nid_hash,
 221                                 &bl_exp->exp_connection->c_peer.nid,
 222                                 ldlm_flock_lookup_cb, &cb_data);
 223                         lock = cb_data.lock;
 224                 }
 225                 if (lock == NULL)
 226                         break;
 227
 228                 class_export_put(bl_exp);
 229                 bl_exp = cb_data.exp;
 230
 231                 LASSERT(req != lock);
 232                 flock = &lock->l_policy_data.l_flock;
 233                 LASSERT(flock->owner == bl_owner);
 234                 bl_owner = flock->blocking_owner;
 235                 bl_exp_new = class_export_get(flock->blocking_export);
 236                 class_export_put(bl_exp);
 237
 238                 cfs_hash_put(bl_exp->exp_flock_hash, &lock->l_exp_flock_hash);
 239                 bl_exp = bl_exp_new;
 240
 241                 if (bl_exp->exp_failed)
 242                         break;
 243
 244                 if (bl_owner == req_owner &&
 245                     (bl_exp->exp_connection->c_peer.nid ==
 246                      req_exp->exp_connection->c_peer.nid)) {
 247                         class_export_put(bl_exp);
 248                         return 1;
 249                 }
 250         }
 251         class_export_put(bl_exp);
 252
 253         return 0;
 254 }
 255
 256 static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
 257                                                 cfs_list_t *work_list)
 258 {
 259         CDEBUG(D_INFO, "reprocess deadlock req=%p\n", lock);
 260
 261         if ((exp_connect_flags(lock->l_export) &
 262                                 OBD_CONNECT_FLOCK_DEAD) == 0) {
 263                 CERROR("deadlock found, but client doesn't "
 264                                 "support flock canceliation\n");
 265         } else {
 266                 LASSERT(lock->l_completion_ast);
 267                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
 268                 lock->l_flags |= LDLM_FL_AST_SENT | LDLM_FL_CANCEL_ON_BLOCK |
 269                         LDLM_FL_FLOCK_DEADLOCK;
 270                 ldlm_flock_blocking_unlink(lock);
 271                 ldlm_resource_unlink_lock(lock);
 272                 ldlm_add_ast_work_item(lock, NULL, work_list);
 273         }
 274 }
 275
 276 /**
 277  * Process a granting attempt for flock lock.
 278  * Must be called under ns lock held.
 279  *
 280  * This function looks for any conflicts for \a lock in the granted or
 281  * waiting queues. The lock is granted if no conflicts are found in
 282  * either queue.
 283  *
 284  * It is also responsible for splitting a lock if a portion of the lock
 285  * is released.
 286  *
 287  * If \a first_enq is 0 (ie, called from ldlm_reprocess_queue):
 288  *   - blocking ASTs have already been sent
 289  *
 290  * If \a first_enq is 1 (ie, called from ldlm_lock_enqueue):
 291  *   - blocking ASTs have not been sent yet, so list of conflicting locks
 292  *     would be collected and ASTs sent.
 293  */
 294 int
 295 ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags, int first_enq,
 296                         ldlm_error_t *err, cfs_list_t *work_list)
 297 {
 298         struct ldlm_resource *res = req->l_resource;
 299         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
 300         cfs_list_t *tmp;
 301         cfs_list_t *ownlocks = NULL;
 302         struct ldlm_lock *lock = NULL;
 303         struct ldlm_lock *new = req;
 304         struct ldlm_lock *new2 = NULL;
 305         ldlm_mode_t mode = req->l_req_mode;
 306         int local = ns_is_client(ns);
 307         int added = (mode == LCK_NL);
 308         int overlaps = 0;
 309         int splitted = 0;
 310         const struct ldlm_callback_suite null_cbs = { NULL };
 311         ENTRY;
 312
 313         CDEBUG(D_DLMTRACE, "flags %#llx owner "LPU64" pid %u mode %u start "
 314                LPU64" end "LPU64"\n", *flags,
 315                new->l_policy_data.l_flock.owner,
 316                new->l_policy_data.l_flock.pid, mode,
 317                req->l_policy_data.l_flock.start,
 318                req->l_policy_data.l_flock.end);
 319
 320         *err = ELDLM_OK;
 321
 322         if (local) {
 323                 /* No blocking ASTs are sent to the clients for
 324                  * Posix file & record locks */
 325                 req->l_blocking_ast = NULL;
 326         } else {
 327                 /* Called on the server for lock cancels. */
 328                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 329         }
 330
 331 reprocess:
 332         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 333                 /* This loop determines where this processes locks start
 334                  * in the resource lr_granted list. */
 335                 cfs_list_for_each(tmp, &res->lr_granted) {
 336                         lock = cfs_list_entry(tmp, struct ldlm_lock,
 337                                               l_res_link);
 338                         if (ldlm_same_flock_owner(lock, req)) {
 339                                 ownlocks = tmp;
 340                                 break;
 341                         }
 342                 }
 343         } else {
 344                 int reprocess_failed = 0;
 345                 lockmode_verify(mode);
 346
 347                 /* This loop determines if there are existing locks
 348                  * that conflict with the new lock request. */
 349                 cfs_list_for_each(tmp, &res->lr_granted) {
 350                         lock = cfs_list_entry(tmp, struct ldlm_lock,
 351                                               l_res_link);
 352
 353                         if (ldlm_same_flock_owner(lock, req)) {
 354                                 if (!ownlocks)
 355                                         ownlocks = tmp;
 356                                 continue;
 357                         }
 358
 359                         /* locks are compatible, overlap doesn't matter */
 360                         if (lockmode_compat(lock->l_granted_mode, mode))
 361                                 continue;
 362
 363                         if (!ldlm_flocks_overlap(lock, req))
 364                                 continue;
 365
 366                         if (!first_enq) {
 367                                 reprocess_failed = 1;
 368                                 if (ldlm_flock_deadlock(req, lock)) {
 369                                         ldlm_flock_cancel_on_deadlock(req,
 370                                                         work_list);
 371                                         RETURN(LDLM_ITER_CONTINUE);
 372                                 }
 373                                 continue;
 374                         }
 375
 376                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 377                                 ldlm_flock_destroy(req, mode, *flags);
 378                                 *err = -EAGAIN;
 379                                 RETURN(LDLM_ITER_STOP);
 380                         }
 381
 382                         if (*flags & LDLM_FL_TEST_LOCK) {
 383                                 ldlm_flock_destroy(req, mode, *flags);
 384                                 req->l_req_mode = lock->l_granted_mode;
 385                                 req->l_policy_data.l_flock.pid =
 386                                         lock->l_policy_data.l_flock.pid;
 387                                 req->l_policy_data.l_flock.start =
 388                                         lock->l_policy_data.l_flock.start;
 389                                 req->l_policy_data.l_flock.end =
 390                                         lock->l_policy_data.l_flock.end;
 391                                 *flags |= LDLM_FL_LOCK_CHANGED;
 392                                 RETURN(LDLM_ITER_STOP);
 393                         }
 394
 395                         /* add lock to blocking list before deadlock
 396                          * check to prevent race */
 397                         ldlm_flock_blocking_link(req, lock);
 398
 399                         if (ldlm_flock_deadlock(req, lock)) {
 400                                 ldlm_flock_blocking_unlink(req);
 401                                 ldlm_flock_destroy(req, mode, *flags);
 402                                 *err = -EDEADLK;
 403                                 RETURN(LDLM_ITER_STOP);
 404                         }
 405
 406                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 407                         *flags |= LDLM_FL_BLOCK_GRANTED;
 408                         RETURN(LDLM_ITER_STOP);
 409                 }
 410                 if (reprocess_failed)
 411                         RETURN(LDLM_ITER_CONTINUE);
 412         }
 413
 414         if (*flags & LDLM_FL_TEST_LOCK) {
 415                 ldlm_flock_destroy(req, mode, *flags);
 416                 req->l_req_mode = LCK_NL;
 417                 *flags |= LDLM_FL_LOCK_CHANGED;
 418                 RETURN(LDLM_ITER_STOP);
 419         }
 420
 421         /* In case we had slept on this lock request take it off of the
 422          * deadlock detection hash list. */
 423         ldlm_flock_blocking_unlink(req);
 424
 425         /* Scan the locks owned by this process that overlap this request.
 426          * We may have to merge or split existing locks. */
 427
 428         if (!ownlocks)
 429                 ownlocks = &res->lr_granted;
 430
 431         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 432                 lock = cfs_list_entry(ownlocks, struct ldlm_lock, l_res_link);
 433
 434                 if (!ldlm_same_flock_owner(lock, new))
 435                         break;
 436
 437                 if (lock->l_granted_mode == mode) {
 438                         /* If the modes are the same then we need to process
 439                          * locks that overlap OR adjoin the new lock. The extra
 440                          * logic condition is necessary to deal with arithmetic
 441                          * overflow and underflow. */
 442                         if ((new->l_policy_data.l_flock.start >
 443                              (lock->l_policy_data.l_flock.end + 1))
 444                             && (lock->l_policy_data.l_flock.end !=
 445                                 OBD_OBJECT_EOF))
 446                                 continue;
 447
 448                         if ((new->l_policy_data.l_flock.end <
 449                              (lock->l_policy_data.l_flock.start - 1))
 450                             && (lock->l_policy_data.l_flock.start != 0))
 451                                 break;
 452
 453                         if (new->l_policy_data.l_flock.start <
 454                             lock->l_policy_data.l_flock.start) {
 455                                 lock->l_policy_data.l_flock.start =
 456                                         new->l_policy_data.l_flock.start;
 457                         } else {
 458                                 new->l_policy_data.l_flock.start =
 459                                         lock->l_policy_data.l_flock.start;
 460                         }
 461
 462                         if (new->l_policy_data.l_flock.end >
 463                             lock->l_policy_data.l_flock.end) {
 464                                 lock->l_policy_data.l_flock.end =
 465                                         new->l_policy_data.l_flock.end;
 466                         } else {
 467                                 new->l_policy_data.l_flock.end =
 468                                         lock->l_policy_data.l_flock.end;
 469                         }
 470
 471                         if (added) {
 472                                 ldlm_flock_destroy(lock, mode, *flags);
 473                         } else {
 474                                 new = lock;
 475                                 added = 1;
 476                         }
 477                         continue;
 478                 }
 479
 480                 if (new->l_policy_data.l_flock.start >
 481                     lock->l_policy_data.l_flock.end)
 482                         continue;
 483
 484                 if (new->l_policy_data.l_flock.end <
 485                     lock->l_policy_data.l_flock.start)
 486                         break;
 487
 488                 ++overlaps;
 489
 490                 if (new->l_policy_data.l_flock.start <=
 491                     lock->l_policy_data.l_flock.start) {
 492                         if (new->l_policy_data.l_flock.end <
 493                             lock->l_policy_data.l_flock.end) {
 494                                 lock->l_policy_data.l_flock.start =
 495                                         new->l_policy_data.l_flock.end + 1;
 496                                 break;
 497                         }
 498                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 499                         continue;
 500                 }
 501                 if (new->l_policy_data.l_flock.end >=
 502                     lock->l_policy_data.l_flock.end) {
 503                         lock->l_policy_data.l_flock.end =
 504                                 new->l_policy_data.l_flock.start - 1;
 505                         continue;
 506                 }
 507
 508                 /* split the existing lock into two locks */
 509
 510                 /* if this is an F_UNLCK operation then we could avoid
 511                  * allocating a new lock and use the req lock passed in
 512                  * with the request but this would complicate the reply
 513                  * processing since updates to req get reflected in the
 514                  * reply. The client side replays the lock request so
 515                  * it must see the original lock data in the reply. */
 516
 517                 /* XXX - if ldlm_lock_new() can sleep we should
 518                  * release the lr_lock, allocate the new lock,
 519                  * and restart processing this lock. */
 520                 if (!new2) {
 521                         unlock_res_and_lock(req);
 522                         new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
 523                                                 lock->l_granted_mode, &null_cbs,
 524                                                 NULL, 0, LVB_T_NONE);
 525                         lock_res_and_lock(req);
 526                         if (!new2) {
 527                                 ldlm_flock_destroy(req, lock->l_granted_mode,
 528                                                    *flags);
 529                                 *err = -ENOLCK;
 530                                 RETURN(LDLM_ITER_STOP);
 531                         }
 532                         goto reprocess;
 533                 }
 534
 535                 splitted = 1;
 536
 537                 new2->l_granted_mode = lock->l_granted_mode;
 538                 new2->l_policy_data.l_flock.pid =
 539                         new->l_policy_data.l_flock.pid;
 540                 new2->l_policy_data.l_flock.owner =
 541                         new->l_policy_data.l_flock.owner;
 542                 new2->l_policy_data.l_flock.start =
 543                         lock->l_policy_data.l_flock.start;
 544                 new2->l_policy_data.l_flock.end =
 545                         new->l_policy_data.l_flock.start - 1;
 546                 lock->l_policy_data.l_flock.start =
 547                         new->l_policy_data.l_flock.end + 1;
 548                 new2->l_conn_export = lock->l_conn_export;
 549                 if (lock->l_export != NULL) {
 550                         new2->l_export = class_export_lock_get(lock->l_export, new2);
 551                         if (new2->l_export->exp_lock_hash &&
 552                             cfs_hlist_unhashed(&new2->l_exp_hash))
 553                                 cfs_hash_add(new2->l_export->exp_lock_hash,
 554                                              &new2->l_remote_handle,
 555                                              &new2->l_exp_hash);
 556                 }
 557                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 558                         ldlm_lock_addref_internal_nolock(new2,
 559                                                          lock->l_granted_mode);
 560
 561                 /* insert new2 at lock */
 562                 ldlm_resource_add_lock(res, ownlocks, new2);
 563                 LDLM_LOCK_RELEASE(new2);
 564                 break;
 565         }
 566
 567         /* if new2 is created but never used, destroy it*/
 568         if (splitted == 0 && new2 != NULL)
 569                 ldlm_lock_destroy_nolock(new2);
 570
 571         /* At this point we're granting the lock request. */
 572         req->l_granted_mode = req->l_req_mode;
 573
 574         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 575         if (!added) {
 576                 cfs_list_del_init(&req->l_res_link);
 577                 /* insert new lock before ownlocks in list. */
 578                 ldlm_resource_add_lock(res, ownlocks, req);
 579         }
 580
 581         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 582 #ifdef HAVE_SERVER_SUPPORT
 583                 if (first_enq) {
 584                         /* If this is an unlock, reprocess the waitq and
 585                          * send completions ASTs for locks that can now be
 586                          * granted. The only problem with doing this
 587                          * reprocessing here is that the completion ASTs for
 588                          * newly granted locks will be sent before the unlock
 589                          * completion is sent. It shouldn't be an issue. Also
 590                          * note that ldlm_process_flock_lock() will recurse,
 591                          * but only once because first_enq will be false from
 592                          * ldlm_reprocess_queue. */
 593                         if ((mode == LCK_NL) && overlaps) {
 594                                 CFS_LIST_HEAD(rpc_list);
 595                                 int rc;
 596 restart:
 597                                 ldlm_reprocess_queue(res, &res->lr_waiting,
 598                                                      &rpc_list);
 599
 600                                 unlock_res_and_lock(req);
 601                                 rc = ldlm_run_ast_work(ns, &rpc_list,
 602                                                        LDLM_WORK_CP_AST);
 603                                 lock_res_and_lock(req);
 604                                 if (rc == -ERESTART)
 605                                         GOTO(restart, -ERESTART);
 606                        }
 607                 } else {
 608                         LASSERT(req->l_completion_ast);
 609                         ldlm_add_ast_work_item(req, NULL, work_list);
 610                 }
 611 #else /* !HAVE_SERVER_SUPPORT */
 612                 /* The only one possible case for client-side calls flock
 613                  * policy function is ldlm_flock_completion_ast inside which
 614                  * carries LDLM_FL_WAIT_NOREPROC flag. */
 615                 CERROR("Illegal parameter for client-side-only module.\n");
 616                 LBUG();
 617 #endif /* HAVE_SERVER_SUPPORT */
 618         }
 619
 620         /* In case we're reprocessing the requested lock we can't destroy
 621          * it until after calling ldlm_add_ast_work_item() above so that laawi()
 622          * can bump the reference count on \a req. Otherwise \a req
 623          * could be freed before the completion AST can be sent.  */
 624         if (added)
 625                 ldlm_flock_destroy(req, mode, *flags);
 626
 627         ldlm_resource_dump(D_INFO, res);
 628         RETURN(LDLM_ITER_CONTINUE);
 629 }
 630
 631 struct ldlm_flock_wait_data {
 632         struct ldlm_lock *fwd_lock;
 633         int               fwd_generation;
 634 };
 635
 636 static void
 637 ldlm_flock_interrupted_wait(void *data)
 638 {
 639         struct ldlm_lock *lock;
 640         ENTRY;
 641
 642         lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
 643
 644         /* take lock off the deadlock detection hash list. */
 645         lock_res_and_lock(lock);
 646         ldlm_flock_blocking_unlink(lock);
 647
 648         /* client side - set flag to prevent lock from being put on LRU list */
 649         lock->l_flags |= LDLM_FL_CBPENDING;
 650         unlock_res_and_lock(lock);
 651
 652         EXIT;
 653 }
 654
 655 /**
 656  * Flock completion callback function.
 657  *
 658  * \param lock [in,out]: A lock to be handled
 659  * \param flags    [in]: flags
 660  * \param *data    [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
 661  *
 662  * \retval 0    : success
 663  * \retval <0   : failure
 664  */
 665 int
 666 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 667 {
 668         struct file_lock                *getlk = lock->l_ast_data;
 669         struct obd_device              *obd;
 670         struct obd_import              *imp = NULL;
 671         struct ldlm_flock_wait_data     fwd;
 672         struct l_wait_info              lwi;
 673         ldlm_error_t                    err;
 674         int                             rc = 0;
 675         ENTRY;
 676
 677         CDEBUG(D_DLMTRACE, "flags: 0x%llx data: %p getlk: %p\n",
 678                flags, data, getlk);
 679
 680         /* Import invalidation. We need to actually release the lock
 681          * references being held, so that it can go away. No point in
 682          * holding the lock even if app still believes it has it, since
 683          * server already dropped it anyway. Only for granted locks too. */
 684         if ((lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) ==
 685             (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) {
 686                 if (lock->l_req_mode == lock->l_granted_mode &&
 687                     lock->l_granted_mode != LCK_NL &&
 688                     NULL == data)
 689                         ldlm_lock_decref_internal(lock, lock->l_req_mode);
 690
 691                 /* Need to wake up the waiter if we were evicted */
 692                 wake_up(&lock->l_waitq);
 693                 RETURN(0);
 694         }
 695
 696         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 697
 698         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 699                        LDLM_FL_BLOCK_CONV))) {
 700                 if (NULL == data)
 701                         /* mds granted the lock in the reply */
 702                         goto granted;
 703                 /* CP AST RPC: lock get granted, wake it up */
 704                 wake_up(&lock->l_waitq);
 705                 RETURN(0);
 706         }
 707
 708         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 709                    "sleeping");
 710         fwd.fwd_lock = lock;
 711         obd = class_exp2obd(lock->l_conn_export);
 712
 713         /* if this is a local lock, there is no import */
 714         if (NULL != obd)
 715                 imp = obd->u.cli.cl_import;
 716
 717         if (NULL != imp) {
 718                 spin_lock(&imp->imp_lock);
 719                 fwd.fwd_generation = imp->imp_generation;
 720                 spin_unlock(&imp->imp_lock);
 721         }
 722
 723         lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
 724
 725         /* Go to sleep until the lock is granted. */
 726         rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
 727
 728         if (rc) {
 729                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 730                            rc);
 731                 RETURN(rc);
 732         }
 733
 734 granted:
 735         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
 736
 737         if (lock->l_flags & LDLM_FL_FAILED) {
 738                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed");
 739                 RETURN(-EIO);
 740         }
 741
 742         LDLM_DEBUG(lock, "client-side enqueue granted");
 743
 744         lock_res_and_lock(lock);
 745
 746
 747         /* Protect against race where lock could have been just destroyed
 748          * due to overlap in ldlm_process_flock_lock().
 749          */
 750         if (lock->l_flags & LDLM_FL_DESTROYED) {
 751                 unlock_res_and_lock(lock);
 752                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 753                 RETURN(0);
 754         }
 755
 756         /* take lock off the deadlock detection hash list. */
 757         ldlm_flock_blocking_unlink(lock);
 758
 759         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 760         cfs_list_del_init(&lock->l_res_link);
 761
 762         if (lock->l_flags & LDLM_FL_FLOCK_DEADLOCK) {
 763                 LDLM_DEBUG(lock, "client-side enqueue deadlock received");
 764                 rc = -EDEADLK;
 765         } else if (flags & LDLM_FL_TEST_LOCK) {
 766                 /* fcntl(F_GETLK) request */
 767                 /* The old mode was saved in getlk->fl_type so that if the mode
 768                  * in the lock changes we can decref the appropriate refcount.*/
 769                 ldlm_flock_destroy(lock, flock_type(getlk),
 770                                    LDLM_FL_WAIT_NOREPROC);
 771                 switch (lock->l_granted_mode) {
 772                 case LCK_PR:
 773                         flock_set_type(getlk, F_RDLCK);
 774                         break;
 775                 case LCK_PW:
 776                         flock_set_type(getlk, F_WRLCK);
 777                         break;
 778                 default:
 779                         flock_set_type(getlk, F_UNLCK);
 780                 }
 781                 flock_set_pid(getlk, (pid_t)lock->l_policy_data.l_flock.pid);
 782                 flock_set_start(getlk,
 783                                 (loff_t)lock->l_policy_data.l_flock.start);
 784                 flock_set_end(getlk,
 785                               (loff_t)lock->l_policy_data.l_flock.end);
 786         } else {
 787                 __u64 noreproc = LDLM_FL_WAIT_NOREPROC;
 788
 789                 /* We need to reprocess the lock to do merges or splits
 790                  * with existing locks owned by this process. */
 791                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 792         }
 793         unlock_res_and_lock(lock);
 794         RETURN(rc);
 795 }
 796 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 797
 798 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 799                             void *data, int flag)
 800 {
 801         ENTRY;
 802
 803         LASSERT(lock);
 804         LASSERT(flag == LDLM_CB_CANCELING);
 805
 806         /* take lock off the deadlock detection hash list. */
 807         lock_res_and_lock(lock);
 808         ldlm_flock_blocking_unlink(lock);
 809         unlock_res_and_lock(lock);
 810         RETURN(0);
 811 }
 812
 813 void ldlm_flock_policy_wire18_to_local(const ldlm_wire_policy_data_t *wpolicy,
 814                                        ldlm_policy_data_t *lpolicy)
 815 {
 816         memset(lpolicy, 0, sizeof(*lpolicy));
 817         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
 818         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
 819         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
 820         /* Compat code, old clients had no idea about owner field and
 821          * relied solely on pid for ownership. Introduced in LU-104, 2.1,
 822          * April 2011 */
 823         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_pid;
 824 }
 825
 826
 827 void ldlm_flock_policy_wire21_to_local(const ldlm_wire_policy_data_t *wpolicy,
 828                                        ldlm_policy_data_t *lpolicy)
 829 {
 830         memset(lpolicy, 0, sizeof(*lpolicy));
 831         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
 832         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
 833         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
 834         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
 835 }
 836
 837 void ldlm_flock_policy_local_to_wire(const ldlm_policy_data_t *lpolicy,
 838                                      ldlm_wire_policy_data_t *wpolicy)
 839 {
 840         memset(wpolicy, 0, sizeof(*wpolicy));
 841         wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
 842         wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
 843         wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
 844         wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
 845 }
 846
 847 /*
 848  * Export handle<->flock hash operations.
 849  */
 850 static unsigned
 851 ldlm_export_flock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
 852 {
 853         return cfs_hash_u64_hash(*(__u64 *)key, mask);
 854 }
 855
 856 static void *
 857 ldlm_export_flock_key(cfs_hlist_node_t *hnode)
 858 {
 859         struct ldlm_lock *lock;
 860
 861         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 862         return &lock->l_policy_data.l_flock.owner;
 863 }
 864
 865 static int
 866 ldlm_export_flock_keycmp(const void *key, cfs_hlist_node_t *hnode)
 867 {
 868         return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
 869 }
 870
 871 static void *
 872 ldlm_export_flock_object(cfs_hlist_node_t *hnode)
 873 {
 874         return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 875 }
 876
 877 static void
 878 ldlm_export_flock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 879 {
 880         struct ldlm_lock *lock;
 881         struct ldlm_flock *flock;
 882
 883         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 884         LDLM_LOCK_GET(lock);
 885
 886         flock = &lock->l_policy_data.l_flock;
 887         LASSERT(flock->blocking_export != NULL);
 888         class_export_get(flock->blocking_export);
 889         flock->blocking_refs++;
 890 }
 891
 892 static void
 893 ldlm_export_flock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 894 {
 895         struct ldlm_lock *lock;
 896         struct ldlm_flock *flock;
 897
 898         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 899         LDLM_LOCK_RELEASE(lock);
 900
 901         flock = &lock->l_policy_data.l_flock;
 902         LASSERT(flock->blocking_export != NULL);
 903         class_export_put(flock->blocking_export);
 904         if (--flock->blocking_refs == 0) {
 905                 flock->blocking_owner = 0;
 906                 flock->blocking_export = NULL;
 907         }
 908 }
 909
 910 static cfs_hash_ops_t ldlm_export_flock_ops = {
 911         .hs_hash        = ldlm_export_flock_hash,
 912         .hs_key         = ldlm_export_flock_key,
 913         .hs_keycmp      = ldlm_export_flock_keycmp,
 914         .hs_object      = ldlm_export_flock_object,
 915         .hs_get         = ldlm_export_flock_get,
 916         .hs_put         = ldlm_export_flock_put,
 917         .hs_put_locked  = ldlm_export_flock_put,
 918 };
 919
 920 int ldlm_init_flock_export(struct obd_export *exp)
 921 {
 922         if( strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0)
 923                 RETURN(0);
 924
 925         exp->exp_flock_hash =
 926                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
 927                                 HASH_EXP_LOCK_CUR_BITS,
 928                                 HASH_EXP_LOCK_MAX_BITS,
 929                                 HASH_EXP_LOCK_BKT_BITS, 0,
 930                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 931                                 &ldlm_export_flock_ops,
 932                                 CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE);
 933         if (!exp->exp_flock_hash)
 934                 RETURN(-ENOMEM);
 935
 936         RETURN(0);
 937 }
 938 EXPORT_SYMBOL(ldlm_init_flock_export);
 939
 940 void ldlm_destroy_flock_export(struct obd_export *exp)
 941 {
 942         ENTRY;
 943         if (exp->exp_flock_hash) {
 944                 cfs_hash_putref(exp->exp_flock_hash);
 945                 exp->exp_flock_hash = NULL;
 946         }
 947         EXIT;
 948 }
 949 EXPORT_SYMBOL(ldlm_destroy_flock_export);