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  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 /**
  38  * This file implements POSIX lock type for Lustre.
  39  * Its policy properties are start and end of extent and PID.
  40  *
  41  * These locks are only done through MDS due to POSIX semantics requiring
  42  * e.g. that locks could be only partially released and as such split into
  43  * two parts, and also that two adjacent locks from the same process may be
  44  * merged into a single wider lock.
  45  *
  46  * Lock modes are mapped like this:
  47  * PR and PW for READ and WRITE locks
  48  * NL to request a releasing of a portion of the lock
  49  *
  50  * These flock locks never timeout.
  51  */
  52
  53 #define DEBUG_SUBSYSTEM S_LDLM
  54
  55 #include <linux/list.h>
  56 #include <lustre_dlm.h>
  57 #include <obd_support.h>
  58 #include <obd_class.h>
  59 #include <lustre_lib.h>
  60
  61 #include "ldlm_internal.h"
  62
  63 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
  64                             void *data, int flag);
  65
  66 /**
  67  * list_for_remaining_safe - iterate over the remaining entries in a list
  68  *              and safeguard against removal of a list entry.
  69  * \param pos   the &struct list_head to use as a loop counter. pos MUST
  70  *              have been initialized prior to using it in this macro.
  71  * \param n     another &struct list_head to use as temporary storage
  72  * \param head  the head for your list.
  73  */
  74 #define list_for_remaining_safe(pos, n, head) \
  75         for (n = pos->next; pos != (head); pos = n, n = pos->next)
  76
  77 static inline int
  78 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
  79 {
  80         return((new->l_policy_data.l_flock.owner ==
  81                 lock->l_policy_data.l_flock.owner) &&
  82                (new->l_export == lock->l_export));
  83 }
  84
  85 static inline int
  86 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
  87 {
  88         return((new->l_policy_data.l_flock.start <=
  89                 lock->l_policy_data.l_flock.end) &&
  90                (new->l_policy_data.l_flock.end >=
  91                 lock->l_policy_data.l_flock.start));
  92 }
  93
  94 static inline void ldlm_flock_blocking_link(struct ldlm_lock *req,
  95                                             struct ldlm_lock *lock)
  96 {
  97         /* For server only */
  98         if (req->l_export == NULL)
  99                 return;
 100
 101         LASSERT(hlist_unhashed(&req->l_exp_flock_hash));
 102
 103         req->l_policy_data.l_flock.blocking_owner =
 104                 lock->l_policy_data.l_flock.owner;
 105         req->l_policy_data.l_flock.blocking_export =
 106                 lock->l_export;
 107         atomic_set(&req->l_policy_data.l_flock.blocking_refs, 0);
 108
 109         cfs_hash_add(req->l_export->exp_flock_hash,
 110                      &req->l_policy_data.l_flock.owner,
 111                      &req->l_exp_flock_hash);
 112 }
 113
 114 static inline void ldlm_flock_blocking_unlink(struct ldlm_lock *req)
 115 {
 116         /* For server only */
 117         if (req->l_export == NULL)
 118                 return;
 119
 120         check_res_locked(req->l_resource);
 121         if (req->l_export->exp_flock_hash != NULL &&
 122             !hlist_unhashed(&req->l_exp_flock_hash))
 123                 cfs_hash_del(req->l_export->exp_flock_hash,
 124                              &req->l_policy_data.l_flock.owner,
 125                              &req->l_exp_flock_hash);
 126 }
 127
 128 static inline void
 129 ldlm_flock_destroy(struct ldlm_lock *lock, enum ldlm_mode mode, __u64 flags)
 130 {
 131         ENTRY;
 132
 133         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: %#llx)",
 134                    mode, flags);
 135
 136         /* Safe to not lock here, since it should be empty anyway */
 137         LASSERT(hlist_unhashed(&lock->l_exp_flock_hash));
 138
 139         list_del_init(&lock->l_res_link);
 140         if (flags == LDLM_FL_WAIT_NOREPROC) {
 141                 /* client side - set a flag to prevent sending a CANCEL */
 142                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
 143
 144                 /* when reaching here, it is under lock_res_and_lock(). Thus,
 145                    need call the nolock version of ldlm_lock_decref_internal*/
 146                 ldlm_lock_decref_internal_nolock(lock, mode);
 147         }
 148
 149         ldlm_lock_destroy_nolock(lock);
 150         EXIT;
 151 }
 152
 153 /**
 154  * POSIX locks deadlock detection code.
 155  *
 156  * Given a new lock \a req and an existing lock \a bl_lock it conflicts
 157  * with, we need to iterate through all blocked POSIX locks for this
 158  * export and see if there is a deadlock condition arising. (i.e. when
 159  * one client holds a lock on something and want a lock on something
 160  * else and at the same time another client has the opposite situation).
 161  */
 162
 163 struct ldlm_flock_lookup_cb_data {
 164         __u64 *bl_owner;
 165         struct ldlm_lock *lock;
 166         struct obd_export *exp;
 167 };
 168
 169 static int ldlm_flock_lookup_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 170                                 struct hlist_node *hnode, void *data)
 171 {
 172         struct ldlm_flock_lookup_cb_data *cb_data = data;
 173         struct obd_export *exp = cfs_hash_object(hs, hnode);
 174         struct ldlm_lock *lock;
 175
 176         lock = cfs_hash_lookup(exp->exp_flock_hash, cb_data->bl_owner);
 177         if (lock == NULL)
 178                 return 0;
 179
 180         /* Stop on first found lock. Same process can't sleep twice */
 181         cb_data->lock = lock;
 182         cb_data->exp = class_export_get(exp);
 183
 184         return 1;
 185 }
 186
 187 static int
 188 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *bl_lock)
 189 {
 190         struct obd_export *req_exp = req->l_export;
 191         struct obd_export *bl_exp = bl_lock->l_export;
 192         __u64 req_owner = req->l_policy_data.l_flock.owner;
 193         __u64 bl_owner = bl_lock->l_policy_data.l_flock.owner;
 194
 195         /* For server only */
 196         if (req_exp == NULL)
 197                 return 0;
 198
 199         class_export_get(bl_exp);
 200         while (1) {
 201                 struct ldlm_flock_lookup_cb_data cb_data = {
 202                                         .bl_owner = &bl_owner,
 203                                         .lock = NULL,
 204                                         .exp = NULL };
 205                 struct obd_export *bl_exp_new;
 206                 struct ldlm_lock *lock = NULL;
 207                 struct ldlm_flock *flock;
 208
 209                 if (bl_exp->exp_flock_hash != NULL) {
 210                         cfs_hash_for_each_key(bl_exp->exp_obd->obd_nid_hash,
 211                                 &bl_exp->exp_connection->c_peer.nid,
 212                                 ldlm_flock_lookup_cb, &cb_data);
 213                         lock = cb_data.lock;
 214                 }
 215                 if (lock == NULL)
 216                         break;
 217
 218                 class_export_put(bl_exp);
 219                 bl_exp = cb_data.exp;
 220
 221                 LASSERT(req != lock);
 222                 flock = &lock->l_policy_data.l_flock;
 223                 LASSERT(flock->owner == bl_owner);
 224                 bl_owner = flock->blocking_owner;
 225                 bl_exp_new = class_export_get(flock->blocking_export);
 226                 class_export_put(bl_exp);
 227
 228                 cfs_hash_put(bl_exp->exp_flock_hash, &lock->l_exp_flock_hash);
 229                 bl_exp = bl_exp_new;
 230
 231                 if (bl_exp->exp_failed)
 232                         break;
 233
 234                 if (bl_owner == req_owner &&
 235                     (bl_exp->exp_connection->c_peer.nid ==
 236                      req_exp->exp_connection->c_peer.nid)) {
 237                         class_export_put(bl_exp);
 238                         return 1;
 239                 }
 240         }
 241         class_export_put(bl_exp);
 242
 243         return 0;
 244 }
 245
 246 static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
 247                                           struct list_head *work_list)
 248 {
 249         CDEBUG(D_INFO, "reprocess deadlock req=%p\n", lock);
 250
 251         if ((exp_connect_flags(lock->l_export) &
 252                                 OBD_CONNECT_FLOCK_DEAD) == 0) {
 253                 CERROR("deadlock found, but client doesn't "
 254                                 "support flock canceliation\n");
 255         } else {
 256                 LASSERT(lock->l_completion_ast);
 257                 LASSERT(!ldlm_is_ast_sent(lock));
 258                 lock->l_flags |= LDLM_FL_AST_SENT | LDLM_FL_CANCEL_ON_BLOCK |
 259                         LDLM_FL_FLOCK_DEADLOCK;
 260                 ldlm_flock_blocking_unlink(lock);
 261                 ldlm_resource_unlink_lock(lock);
 262                 ldlm_add_ast_work_item(lock, NULL, work_list);
 263         }
 264 }
 265
 266 /**
 267  * Process a granting attempt for flock lock.
 268  * Must be called under ns lock held.
 269  *
 270  * This function looks for any conflicts for \a lock in the granted or
 271  * waiting queues. The lock is granted if no conflicts are found in
 272  * either queue.
 273  */
 274 int
 275 ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags,
 276                         enum ldlm_process_intention intention,
 277                         enum ldlm_error *err, struct list_head *work_list)
 278 {
 279         struct ldlm_resource *res = req->l_resource;
 280         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
 281         struct list_head *tmp;
 282         struct list_head *ownlocks = NULL;
 283         struct ldlm_lock *lock = NULL;
 284         struct ldlm_lock *new = req;
 285         struct ldlm_lock *new2 = NULL;
 286         enum ldlm_mode mode = req->l_req_mode;
 287         int local = ns_is_client(ns);
 288         int added = (mode == LCK_NL);
 289         int overlaps = 0;
 290         int splitted = 0;
 291         const struct ldlm_callback_suite null_cbs = { NULL };
 292         ENTRY;
 293
 294         CDEBUG(D_DLMTRACE, "flags %#llx owner %llu pid %u mode %u start "
 295                "%llu end %llu\n", *flags,
 296                new->l_policy_data.l_flock.owner,
 297                new->l_policy_data.l_flock.pid, mode,
 298                req->l_policy_data.l_flock.start,
 299                req->l_policy_data.l_flock.end);
 300
 301         *err = ELDLM_OK;
 302
 303         if (local) {
 304                 /* No blocking ASTs are sent to the clients for
 305                  * Posix file & record locks */
 306                 req->l_blocking_ast = NULL;
 307         } else {
 308                 /* Called on the server for lock cancels. */
 309                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 310         }
 311
 312 reprocess:
 313         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 314                 /* This loop determines where this processes locks start
 315                  * in the resource lr_granted list. */
 316                 list_for_each(tmp, &res->lr_granted) {
 317                         lock = list_entry(tmp, struct ldlm_lock,
 318                                               l_res_link);
 319                         if (ldlm_same_flock_owner(lock, req)) {
 320                                 ownlocks = tmp;
 321                                 break;
 322                         }
 323                 }
 324         } else {
 325                 int reprocess_failed = 0;
 326                 lockmode_verify(mode);
 327
 328                 /* This loop determines if there are existing locks
 329                  * that conflict with the new lock request. */
 330                 list_for_each(tmp, &res->lr_granted) {
 331                         lock = list_entry(tmp, struct ldlm_lock,
 332                                               l_res_link);
 333
 334                         if (ldlm_same_flock_owner(lock, req)) {
 335                                 if (!ownlocks)
 336                                         ownlocks = tmp;
 337                                 continue;
 338                         }
 339
 340                         /* locks are compatible, overlap doesn't matter */
 341                         if (lockmode_compat(lock->l_granted_mode, mode))
 342                                 continue;
 343
 344                         if (!ldlm_flocks_overlap(lock, req))
 345                                 continue;
 346
 347                         if (intention != LDLM_PROCESS_ENQUEUE) {
 348                                 reprocess_failed = 1;
 349                                 if (ldlm_flock_deadlock(req, lock)) {
 350                                         ldlm_flock_cancel_on_deadlock(req,
 351                                                         work_list);
 352                                         RETURN(LDLM_ITER_CONTINUE);
 353                                 }
 354                                 continue;
 355                         }
 356
 357                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 358                                 ldlm_flock_destroy(req, mode, *flags);
 359                                 *err = -EAGAIN;
 360                                 RETURN(LDLM_ITER_STOP);
 361                         }
 362
 363                         if (*flags & LDLM_FL_TEST_LOCK) {
 364                                 ldlm_flock_destroy(req, mode, *flags);
 365                                 req->l_req_mode = lock->l_granted_mode;
 366                                 req->l_policy_data.l_flock.pid =
 367                                         lock->l_policy_data.l_flock.pid;
 368                                 req->l_policy_data.l_flock.start =
 369                                         lock->l_policy_data.l_flock.start;
 370                                 req->l_policy_data.l_flock.end =
 371                                         lock->l_policy_data.l_flock.end;
 372                                 *flags |= LDLM_FL_LOCK_CHANGED;
 373                                 RETURN(LDLM_ITER_STOP);
 374                         }
 375
 376                         /* add lock to blocking list before deadlock
 377                          * check to prevent race */
 378                         ldlm_flock_blocking_link(req, lock);
 379
 380                         if (ldlm_flock_deadlock(req, lock)) {
 381                                 ldlm_flock_blocking_unlink(req);
 382                                 ldlm_flock_destroy(req, mode, *flags);
 383                                 *err = -EDEADLK;
 384                                 RETURN(LDLM_ITER_STOP);
 385                         }
 386
 387                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 388                         *flags |= LDLM_FL_BLOCK_GRANTED;
 389                         RETURN(LDLM_ITER_STOP);
 390                 }
 391                 if (reprocess_failed)
 392                         RETURN(LDLM_ITER_CONTINUE);
 393         }
 394
 395         if (*flags & LDLM_FL_TEST_LOCK) {
 396                 ldlm_flock_destroy(req, mode, *flags);
 397                 req->l_req_mode = LCK_NL;
 398                 *flags |= LDLM_FL_LOCK_CHANGED;
 399                 RETURN(LDLM_ITER_STOP);
 400         }
 401
 402         /* In case we had slept on this lock request take it off of the
 403          * deadlock detection hash list. */
 404         ldlm_flock_blocking_unlink(req);
 405
 406         /* Scan the locks owned by this process that overlap this request.
 407          * We may have to merge or split existing locks. */
 408
 409         if (!ownlocks)
 410                 ownlocks = &res->lr_granted;
 411
 412         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 413                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 414
 415                 if (!ldlm_same_flock_owner(lock, new))
 416                         break;
 417
 418                 if (lock->l_granted_mode == mode) {
 419                         /* If the modes are the same then we need to process
 420                          * locks that overlap OR adjoin the new lock. The extra
 421                          * logic condition is necessary to deal with arithmetic
 422                          * overflow and underflow. */
 423                         if ((new->l_policy_data.l_flock.start >
 424                              (lock->l_policy_data.l_flock.end + 1))
 425                             && (lock->l_policy_data.l_flock.end !=
 426                                 OBD_OBJECT_EOF))
 427                                 continue;
 428
 429                         if ((new->l_policy_data.l_flock.end <
 430                              (lock->l_policy_data.l_flock.start - 1))
 431                             && (lock->l_policy_data.l_flock.start != 0))
 432                                 break;
 433
 434                         if (new->l_policy_data.l_flock.start <
 435                             lock->l_policy_data.l_flock.start) {
 436                                 lock->l_policy_data.l_flock.start =
 437                                         new->l_policy_data.l_flock.start;
 438                         } else {
 439                                 new->l_policy_data.l_flock.start =
 440                                         lock->l_policy_data.l_flock.start;
 441                         }
 442
 443                         if (new->l_policy_data.l_flock.end >
 444                             lock->l_policy_data.l_flock.end) {
 445                                 lock->l_policy_data.l_flock.end =
 446                                         new->l_policy_data.l_flock.end;
 447                         } else {
 448                                 new->l_policy_data.l_flock.end =
 449                                         lock->l_policy_data.l_flock.end;
 450                         }
 451
 452                         if (added) {
 453                                 ldlm_flock_destroy(lock, mode, *flags);
 454                         } else {
 455                                 new = lock;
 456                                 added = 1;
 457                         }
 458                         continue;
 459                 }
 460
 461                 if (new->l_policy_data.l_flock.start >
 462                     lock->l_policy_data.l_flock.end)
 463                         continue;
 464
 465                 if (new->l_policy_data.l_flock.end <
 466                     lock->l_policy_data.l_flock.start)
 467                         break;
 468
 469                 ++overlaps;
 470
 471                 if (new->l_policy_data.l_flock.start <=
 472                     lock->l_policy_data.l_flock.start) {
 473                         if (new->l_policy_data.l_flock.end <
 474                             lock->l_policy_data.l_flock.end) {
 475                                 lock->l_policy_data.l_flock.start =
 476                                         new->l_policy_data.l_flock.end + 1;
 477                                 break;
 478                         }
 479                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 480                         continue;
 481                 }
 482                 if (new->l_policy_data.l_flock.end >=
 483                     lock->l_policy_data.l_flock.end) {
 484                         lock->l_policy_data.l_flock.end =
 485                                 new->l_policy_data.l_flock.start - 1;
 486                         continue;
 487                 }
 488
 489                 /* split the existing lock into two locks */
 490
 491                 /* if this is an F_UNLCK operation then we could avoid
 492                  * allocating a new lock and use the req lock passed in
 493                  * with the request but this would complicate the reply
 494                  * processing since updates to req get reflected in the
 495                  * reply. The client side replays the lock request so
 496                  * it must see the original lock data in the reply. */
 497
 498                 /* XXX - if ldlm_lock_new() can sleep we should
 499                  * release the lr_lock, allocate the new lock,
 500                  * and restart processing this lock. */
 501                 if (new2 == NULL) {
 502                         unlock_res_and_lock(req);
 503                         new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
 504                                                 lock->l_granted_mode, &null_cbs,
 505                                                 NULL, 0, LVB_T_NONE);
 506                         lock_res_and_lock(req);
 507                         if (IS_ERR(new2)) {
 508                                 ldlm_flock_destroy(req, lock->l_granted_mode,
 509                                                    *flags);
 510                                 *err = PTR_ERR(new2);
 511                                 RETURN(LDLM_ITER_STOP);
 512                         }
 513                         goto reprocess;
 514                 }
 515
 516                 splitted = 1;
 517
 518                 new2->l_granted_mode = lock->l_granted_mode;
 519                 new2->l_policy_data.l_flock.pid =
 520                         new->l_policy_data.l_flock.pid;
 521                 new2->l_policy_data.l_flock.owner =
 522                         new->l_policy_data.l_flock.owner;
 523                 new2->l_policy_data.l_flock.start =
 524                         lock->l_policy_data.l_flock.start;
 525                 new2->l_policy_data.l_flock.end =
 526                         new->l_policy_data.l_flock.start - 1;
 527                 lock->l_policy_data.l_flock.start =
 528                         new->l_policy_data.l_flock.end + 1;
 529                 new2->l_conn_export = lock->l_conn_export;
 530                 if (lock->l_export != NULL) {
 531                         new2->l_export = class_export_lock_get(lock->l_export, new2);
 532                         if (new2->l_export->exp_lock_hash &&
 533                             hlist_unhashed(&new2->l_exp_hash))
 534                                 cfs_hash_add(new2->l_export->exp_lock_hash,
 535                                              &new2->l_remote_handle,
 536                                              &new2->l_exp_hash);
 537                 }
 538                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 539                         ldlm_lock_addref_internal_nolock(new2,
 540                                                          lock->l_granted_mode);
 541
 542                 /* insert new2 at lock */
 543                 ldlm_resource_add_lock(res, ownlocks, new2);
 544                 LDLM_LOCK_RELEASE(new2);
 545                 break;
 546         }
 547
 548         /* if new2 is created but never used, destroy it*/
 549         if (splitted == 0 && new2 != NULL)
 550                 ldlm_lock_destroy_nolock(new2);
 551
 552         /* At this point we're granting the lock request. */
 553         req->l_granted_mode = req->l_req_mode;
 554
 555         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 556         if (!added) {
 557                 list_del_init(&req->l_res_link);
 558                 /* insert new lock before ownlocks in list. */
 559                 ldlm_resource_add_lock(res, ownlocks, req);
 560         }
 561
 562         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 563 #ifdef HAVE_SERVER_SUPPORT
 564                 if (intention == LDLM_PROCESS_ENQUEUE) {
 565                         /* If this is an unlock, reprocess the waitq and
 566                          * send completions ASTs for locks that can now be
 567                          * granted. The only problem with doing this
 568                          * reprocessing here is that the completion ASTs for
 569                          * newly granted locks will be sent before the unlock
 570                          * completion is sent. It shouldn't be an issue. Also
 571                          * note that ldlm_process_flock_lock() will recurse,
 572                          * but only once because 'intention' won't be
 573                          * LDLM_PROCESS_ENQUEUE from ldlm_reprocess_queue. */
 574                         if ((mode == LCK_NL) && overlaps) {
 575                                 struct list_head rpc_list;
 576                                 int rc;
 577
 578                                 INIT_LIST_HEAD(&rpc_list);
 579 restart:
 580                                 ldlm_reprocess_queue(res, &res->lr_waiting,
 581                                                      &rpc_list,
 582                                                      LDLM_PROCESS_RESCAN);
 583
 584                                 unlock_res_and_lock(req);
 585                                 rc = ldlm_run_ast_work(ns, &rpc_list,
 586                                                        LDLM_WORK_CP_AST);
 587                                 lock_res_and_lock(req);
 588                                 if (rc == -ERESTART)
 589                                         GOTO(restart, rc);
 590                        }
 591                 } else {
 592                         LASSERT(req->l_completion_ast);
 593                         ldlm_add_ast_work_item(req, NULL, work_list);
 594                 }
 595 #else /* !HAVE_SERVER_SUPPORT */
 596                 /* The only one possible case for client-side calls flock
 597                  * policy function is ldlm_flock_completion_ast inside which
 598                  * carries LDLM_FL_WAIT_NOREPROC flag. */
 599                 CERROR("Illegal parameter for client-side-only module.\n");
 600                 LBUG();
 601 #endif /* HAVE_SERVER_SUPPORT */
 602         }
 603
 604         /* In case we're reprocessing the requested lock we can't destroy
 605          * it until after calling ldlm_add_ast_work_item() above so that laawi()
 606          * can bump the reference count on \a req. Otherwise \a req
 607          * could be freed before the completion AST can be sent.  */
 608         if (added)
 609                 ldlm_flock_destroy(req, mode, *flags);
 610
 611         ldlm_resource_dump(D_INFO, res);
 612         RETURN(LDLM_ITER_CONTINUE);
 613 }
 614
 615 struct ldlm_flock_wait_data {
 616         struct ldlm_lock *fwd_lock;
 617         int               fwd_generation;
 618 };
 619
 620 static void
 621 ldlm_flock_interrupted_wait(void *data)
 622 {
 623         struct ldlm_lock *lock;
 624         ENTRY;
 625
 626         lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
 627
 628         /* take lock off the deadlock detection hash list. */
 629         lock_res_and_lock(lock);
 630         ldlm_flock_blocking_unlink(lock);
 631
 632         /* client side - set flag to prevent lock from being put on LRU list */
 633         ldlm_set_cbpending(lock);
 634         unlock_res_and_lock(lock);
 635
 636         EXIT;
 637 }
 638
 639 /**
 640  * Flock completion callback function.
 641  *
 642  * \param lock [in,out]: A lock to be handled
 643  * \param flags    [in]: flags
 644  * \param *data    [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
 645  *
 646  * \retval 0    : success
 647  * \retval <0   : failure
 648  */
 649 int
 650 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 651 {
 652         struct file_lock *getlk = lock->l_ast_data;
 653         struct obd_device *obd;
 654         struct obd_import *imp = NULL;
 655         struct ldlm_flock_wait_data fwd;
 656         struct l_wait_info lwi;
 657         enum ldlm_error err;
 658         int rc = 0;
 659         ENTRY;
 660
 661         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4);
 662         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) {
 663                 lock_res_and_lock(lock);
 664                 lock->l_flags |= LDLM_FL_FAIL_LOC;
 665                 unlock_res_and_lock(lock);
 666                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT3, 4);
 667         }
 668         CDEBUG(D_DLMTRACE, "flags: %#llx data: %p getlk: %p\n",
 669                flags, data, getlk);
 670
 671         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 672
 673         if (flags & LDLM_FL_FAILED)
 674                 goto granted;
 675
 676         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
 677                 if (NULL == data)
 678                         /* mds granted the lock in the reply */
 679                         goto granted;
 680                 /* CP AST RPC: lock get granted, wake it up */
 681                 wake_up(&lock->l_waitq);
 682                 RETURN(0);
 683         }
 684
 685         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 686                    "sleeping");
 687         fwd.fwd_lock = lock;
 688         obd = class_exp2obd(lock->l_conn_export);
 689
 690         /* if this is a local lock, there is no import */
 691         if (NULL != obd)
 692                 imp = obd->u.cli.cl_import;
 693
 694         if (NULL != imp) {
 695                 spin_lock(&imp->imp_lock);
 696                 fwd.fwd_generation = imp->imp_generation;
 697                 spin_unlock(&imp->imp_lock);
 698         }
 699
 700         lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
 701
 702         /* Go to sleep until the lock is granted. */
 703         rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
 704
 705         if (rc) {
 706                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 707                            rc);
 708                 RETURN(rc);
 709         }
 710
 711 granted:
 712         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
 713
 714         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT4)) {
 715                 lock_res_and_lock(lock);
 716                 /* DEADLOCK is always set with CBPENDING */
 717                 lock->l_flags |= LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
 718                 unlock_res_and_lock(lock);
 719                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT4, 4);
 720         }
 721         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT5)) {
 722                 lock_res_and_lock(lock);
 723                 /* DEADLOCK is always set with CBPENDING */
 724                 lock->l_flags |= LDLM_FL_FAIL_LOC |
 725                                  LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
 726                 unlock_res_and_lock(lock);
 727                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT5, 4);
 728         }
 729
 730         lock_res_and_lock(lock);
 731
 732
 733         /* Protect against race where lock could have been just destroyed
 734          * due to overlap in ldlm_process_flock_lock().
 735          */
 736         if (ldlm_is_destroyed(lock)) {
 737                 unlock_res_and_lock(lock);
 738                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 739
 740                 /* An error is still to be returned, to propagate it up to
 741                  * ldlm_cli_enqueue_fini() caller. */
 742                 RETURN(-EIO);
 743         }
 744
 745         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 746         ldlm_resource_unlink_lock(lock);
 747
 748         /* Import invalidation. We need to actually release the lock
 749          * references being held, so that it can go away. No point in
 750          * holding the lock even if app still believes it has it, since
 751          * server already dropped it anyway. Only for granted locks too. */
 752         /* Do the same for DEADLOCK'ed locks. */
 753         if (ldlm_is_failed(lock) || ldlm_is_flock_deadlock(lock)) {
 754                 int mode;
 755
 756                 if (flags & LDLM_FL_TEST_LOCK)
 757                         LASSERT(ldlm_is_test_lock(lock));
 758
 759                 if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock))
 760                         mode = getlk->fl_type;
 761                 else
 762                         mode = lock->l_granted_mode;
 763
 764                 if (ldlm_is_flock_deadlock(lock)) {
 765                         LDLM_DEBUG(lock, "client-side enqueue deadlock "
 766                                    "received");
 767                         rc = -EDEADLK;
 768                 }
 769                 ldlm_flock_destroy(lock, mode, LDLM_FL_WAIT_NOREPROC);
 770                 unlock_res_and_lock(lock);
 771
 772                 /* Need to wake up the waiter if we were evicted */
 773                 wake_up(&lock->l_waitq);
 774
 775                 /* An error is still to be returned, to propagate it up to
 776                  * ldlm_cli_enqueue_fini() caller. */
 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                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 809         }
 810         unlock_res_and_lock(lock);
 811         RETURN(rc);
 812 }
 813 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 814
 815 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 816                             void *data, int flag)
 817 {
 818         ENTRY;
 819
 820         LASSERT(lock);
 821         LASSERT(flag == LDLM_CB_CANCELING);
 822
 823         /* take lock off the deadlock detection hash list. */
 824         lock_res_and_lock(lock);
 825         ldlm_flock_blocking_unlink(lock);
 826         unlock_res_and_lock(lock);
 827         RETURN(0);
 828 }
 829
 830 void ldlm_flock_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
 831                                      union ldlm_policy_data *lpolicy)
 832 {
 833         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
 834         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
 835         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
 836         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
 837 }
 838
 839 void ldlm_flock_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
 840                                      union ldlm_wire_policy_data *wpolicy)
 841 {
 842         memset(wpolicy, 0, sizeof(*wpolicy));
 843         wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
 844         wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
 845         wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
 846         wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
 847 }
 848
 849 /*
 850  * Export handle<->flock hash operations.
 851  */
 852 static unsigned
 853 ldlm_export_flock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
 854 {
 855         return cfs_hash_u64_hash(*(__u64 *)key, mask);
 856 }
 857
 858 static void *
 859 ldlm_export_flock_key(struct hlist_node *hnode)
 860 {
 861         struct ldlm_lock *lock;
 862
 863         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 864         return &lock->l_policy_data.l_flock.owner;
 865 }
 866
 867 static int
 868 ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode)
 869 {
 870         return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
 871 }
 872
 873 static void *
 874 ldlm_export_flock_object(struct hlist_node *hnode)
 875 {
 876         return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 877 }
 878
 879 static void
 880 ldlm_export_flock_get(struct cfs_hash *hs, struct hlist_node *hnode)
 881 {
 882         struct ldlm_lock *lock;
 883         struct ldlm_flock *flock;
 884
 885         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 886         LDLM_LOCK_GET(lock);
 887
 888         flock = &lock->l_policy_data.l_flock;
 889         LASSERT(flock->blocking_export != NULL);
 890         class_export_get(flock->blocking_export);
 891         atomic_inc(&flock->blocking_refs);
 892 }
 893
 894 static void
 895 ldlm_export_flock_put(struct cfs_hash *hs, struct hlist_node *hnode)
 896 {
 897         struct ldlm_lock *lock;
 898         struct ldlm_flock *flock;
 899
 900         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 901
 902         flock = &lock->l_policy_data.l_flock;
 903         LASSERT(flock->blocking_export != NULL);
 904         class_export_put(flock->blocking_export);
 905         if (atomic_dec_and_test(&flock->blocking_refs)) {
 906                 flock->blocking_owner = 0;
 907                 flock->blocking_export = NULL;
 908         }
 909         LDLM_LOCK_RELEASE(lock);
 910 }
 911
 912 static struct cfs_hash_ops ldlm_export_flock_ops = {
 913         .hs_hash        = ldlm_export_flock_hash,
 914         .hs_key         = ldlm_export_flock_key,
 915         .hs_keycmp      = ldlm_export_flock_keycmp,
 916         .hs_object      = ldlm_export_flock_object,
 917         .hs_get         = ldlm_export_flock_get,
 918         .hs_put         = ldlm_export_flock_put,
 919         .hs_put_locked  = ldlm_export_flock_put,
 920 };
 921
 922 int ldlm_init_flock_export(struct obd_export *exp)
 923 {
 924         if( strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0)
 925                 RETURN(0);
 926
 927         exp->exp_flock_hash =
 928                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
 929                                 HASH_EXP_LOCK_CUR_BITS,
 930                                 HASH_EXP_LOCK_MAX_BITS,
 931                                 HASH_EXP_LOCK_BKT_BITS, 0,
 932                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 933                                 &ldlm_export_flock_ops,
 934                                 CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE);
 935         if (!exp->exp_flock_hash)
 936                 RETURN(-ENOMEM);
 937
 938         RETURN(0);
 939 }
 940
 941 void ldlm_destroy_flock_export(struct obd_export *exp)
 942 {
 943         ENTRY;
 944         if (exp->exp_flock_hash) {
 945                 cfs_hash_putref(exp->exp_flock_hash);
 946                 exp->exp_flock_hash = NULL;
 947         }
 948         EXIT;
 949 }