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                  */
 147                 ldlm_lock_decref_internal_nolock(lock, mode);
 148         }
 149
 150         ldlm_lock_destroy_nolock(lock);
 151         EXIT;
 152 }
 153
 154 /**
 155  * POSIX locks deadlock detection code.
 156  *
 157  * Given a new lock \a req and an existing lock \a bl_lock it conflicts
 158  * with, we need to iterate through all blocked POSIX locks for this
 159  * export and see if there is a deadlock condition arising. (i.e. when
 160  * one client holds a lock on something and want a lock on something
 161  * else and at the same time another client has the opposite situation).
 162  */
 163
 164 struct ldlm_flock_lookup_cb_data {
 165         __u64 *bl_owner;
 166         struct ldlm_lock *lock;
 167         struct obd_export *exp;
 168 };
 169
 170 static int ldlm_flock_lookup_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 171                                 struct hlist_node *hnode, void *data)
 172 {
 173         struct ldlm_flock_lookup_cb_data *cb_data = data;
 174         struct obd_export *exp = cfs_hash_object(hs, hnode);
 175         struct ldlm_lock *lock;
 176
 177         lock = cfs_hash_lookup(exp->exp_flock_hash, cb_data->bl_owner);
 178         if (lock == NULL)
 179                 return 0;
 180
 181         /* Stop on first found lock. Same process can't sleep twice */
 182         cb_data->lock = lock;
 183         cb_data->exp = class_export_get(exp);
 184
 185         return 1;
 186 }
 187
 188 static int
 189 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *bl_lock)
 190 {
 191         struct obd_export *req_exp = req->l_export;
 192         struct obd_export *bl_exp = bl_lock->l_export;
 193         __u64 req_owner = req->l_policy_data.l_flock.owner;
 194         __u64 bl_owner = bl_lock->l_policy_data.l_flock.owner;
 195
 196         /* For server only */
 197         if (req_exp == NULL)
 198                 return 0;
 199
 200         class_export_get(bl_exp);
 201         while (1) {
 202                 struct ldlm_flock_lookup_cb_data cb_data = {
 203                         .bl_owner = &bl_owner,
 204                         .lock = NULL,
 205                         .exp = NULL,
 206                 };
 207                 struct obd_export *bl_exp_new;
 208                 struct ldlm_lock *lock = NULL;
 209                 struct ldlm_flock *flock;
 210
 211                 if (bl_exp->exp_flock_hash != NULL) {
 212                         cfs_hash_for_each_key(
 213                                 bl_exp->exp_obd->obd_nid_hash,
 214                                 &bl_exp->exp_connection->c_peer.nid,
 215                                 ldlm_flock_lookup_cb, &cb_data);
 216                         lock = cb_data.lock;
 217                 }
 218                 if (lock == NULL)
 219                         break;
 220
 221                 class_export_put(bl_exp);
 222                 bl_exp = cb_data.exp;
 223
 224                 LASSERT(req != lock);
 225                 flock = &lock->l_policy_data.l_flock;
 226                 LASSERT(flock->owner == bl_owner);
 227                 bl_owner = flock->blocking_owner;
 228                 bl_exp_new = class_export_get(flock->blocking_export);
 229                 class_export_put(bl_exp);
 230
 231                 cfs_hash_put(bl_exp->exp_flock_hash, &lock->l_exp_flock_hash);
 232                 bl_exp = bl_exp_new;
 233
 234                 if (bl_exp->exp_failed)
 235                         break;
 236
 237                 if (bl_owner == req_owner &&
 238                     (bl_exp->exp_connection->c_peer.nid ==
 239                      req_exp->exp_connection->c_peer.nid)) {
 240                         class_export_put(bl_exp);
 241                         return 1;
 242                 }
 243         }
 244         class_export_put(bl_exp);
 245
 246         return 0;
 247 }
 248
 249 static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
 250                                           struct list_head *work_list)
 251 {
 252         CDEBUG(D_INFO, "reprocess deadlock req=%p\n", lock);
 253
 254         if ((exp_connect_flags(lock->l_export) &
 255              OBD_CONNECT_FLOCK_DEAD) == 0) {
 256                 CERROR("deadlock found, but client doesn't support flock canceliation\n");
 257         } else {
 258                 LASSERT(lock->l_completion_ast);
 259                 LASSERT(!ldlm_is_ast_sent(lock));
 260                 lock->l_flags |= (LDLM_FL_AST_SENT | LDLM_FL_CANCEL_ON_BLOCK |
 261                                   LDLM_FL_FLOCK_DEADLOCK);
 262                 ldlm_flock_blocking_unlink(lock);
 263                 ldlm_resource_unlink_lock(lock);
 264                 ldlm_add_ast_work_item(lock, NULL, work_list);
 265         }
 266 }
 267
 268 /**
 269  * Process a granting attempt for flock lock.
 270  * Must be called under ns lock held.
 271  *
 272  * This function looks for any conflicts for \a lock in the granted or
 273  * waiting queues. The lock is granted if no conflicts are found in
 274  * either queue.
 275  */
 276 int
 277 ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags,
 278                         enum ldlm_process_intention intention,
 279                         enum ldlm_error *err, struct list_head *work_list)
 280 {
 281         struct ldlm_resource *res = req->l_resource;
 282         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
 283         struct list_head *tmp;
 284         struct list_head *ownlocks = NULL;
 285         struct ldlm_lock *lock = NULL;
 286         struct ldlm_lock *new = req;
 287         struct ldlm_lock *new2 = NULL;
 288         enum ldlm_mode mode = req->l_req_mode;
 289         int local = ns_is_client(ns);
 290         int added = (mode == LCK_NL);
 291         int overlaps = 0;
 292         int splitted = 0;
 293         const struct ldlm_callback_suite null_cbs = { NULL };
 294         struct list_head *grant_work = (intention == LDLM_PROCESS_ENQUEUE ?
 295                                         NULL : work_list);
 296         ENTRY;
 297
 298         CDEBUG(D_DLMTRACE, "flags %#llx owner %llu pid %u mode %u start "
 299                "%llu end %llu\n", *flags,
 300                new->l_policy_data.l_flock.owner,
 301                new->l_policy_data.l_flock.pid, mode,
 302                req->l_policy_data.l_flock.start,
 303                req->l_policy_data.l_flock.end);
 304
 305         *err = ELDLM_OK;
 306
 307         if (local) {
 308                 /* No blocking ASTs are sent to the clients for
 309                  * Posix file & record locks
 310                  */
 311                 req->l_blocking_ast = NULL;
 312         } else {
 313                 /* Called on the server for lock cancels. */
 314                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 315         }
 316
 317 reprocess:
 318         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 319                 /* This loop determines where this processes locks start
 320                  * in the resource lr_granted list.
 321                  */
 322                 list_for_each(tmp, &res->lr_granted) {
 323                         lock = list_entry(tmp, struct ldlm_lock,
 324                                           l_res_link);
 325                         if (ldlm_same_flock_owner(lock, req)) {
 326                                 ownlocks = tmp;
 327                                 break;
 328                         }
 329                 }
 330         } else {
 331                 int reprocess_failed = 0;
 332                 lockmode_verify(mode);
 333
 334                 /* This loop determines if there are existing locks
 335                  * that conflict with the new lock request.
 336                  */
 337                 list_for_each(tmp, &res->lr_granted) {
 338                         lock = list_entry(tmp, struct ldlm_lock,
 339                                           l_res_link);
 340
 341                         if (ldlm_same_flock_owner(lock, req)) {
 342                                 if (!ownlocks)
 343                                         ownlocks = tmp;
 344                                 continue;
 345                         }
 346
 347                         /* locks are compatible, overlap doesn't matter */
 348                         if (lockmode_compat(lock->l_granted_mode, mode))
 349                                 continue;
 350
 351                         if (!ldlm_flocks_overlap(lock, req))
 352                                 continue;
 353
 354                         if (intention != LDLM_PROCESS_ENQUEUE) {
 355                                 reprocess_failed = 1;
 356                                 if (ldlm_flock_deadlock(req, lock)) {
 357                                         ldlm_flock_cancel_on_deadlock(
 358                                                 req, grant_work);
 359                                         RETURN(LDLM_ITER_CONTINUE);
 360                                 }
 361                                 continue;
 362                         }
 363
 364                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 365                                 ldlm_flock_destroy(req, mode, *flags);
 366                                 *err = -EAGAIN;
 367                                 RETURN(LDLM_ITER_STOP);
 368                         }
 369
 370                         if (*flags & LDLM_FL_TEST_LOCK) {
 371                                 ldlm_flock_destroy(req, mode, *flags);
 372                                 req->l_req_mode = lock->l_granted_mode;
 373                                 req->l_policy_data.l_flock.pid =
 374                                         lock->l_policy_data.l_flock.pid;
 375                                 req->l_policy_data.l_flock.start =
 376                                         lock->l_policy_data.l_flock.start;
 377                                 req->l_policy_data.l_flock.end =
 378                                         lock->l_policy_data.l_flock.end;
 379                                 *flags |= LDLM_FL_LOCK_CHANGED;
 380                                 RETURN(LDLM_ITER_STOP);
 381                         }
 382
 383                         /* add lock to blocking list before deadlock
 384                          * check to prevent race
 385                          */
 386                         ldlm_flock_blocking_link(req, lock);
 387
 388                         if (ldlm_flock_deadlock(req, lock)) {
 389                                 ldlm_flock_blocking_unlink(req);
 390                                 ldlm_flock_destroy(req, mode, *flags);
 391                                 *err = -EDEADLK;
 392                                 RETURN(LDLM_ITER_STOP);
 393                         }
 394
 395                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 396                         *flags |= LDLM_FL_BLOCK_GRANTED;
 397                         RETURN(LDLM_ITER_STOP);
 398                 }
 399                 if (reprocess_failed)
 400                         RETURN(LDLM_ITER_CONTINUE);
 401         }
 402
 403         if (*flags & LDLM_FL_TEST_LOCK) {
 404                 ldlm_flock_destroy(req, mode, *flags);
 405                 req->l_req_mode = LCK_NL;
 406                 *flags |= LDLM_FL_LOCK_CHANGED;
 407                 RETURN(LDLM_ITER_STOP);
 408         }
 409
 410         /* In case we had slept on this lock request take it off of the
 411          * deadlock detection hash list.
 412          */
 413         ldlm_flock_blocking_unlink(req);
 414
 415         /* Scan the locks owned by this process that overlap this request.
 416          * We may have to merge or split existing locks.
 417          */
 418
 419         if (!ownlocks)
 420                 ownlocks = &res->lr_granted;
 421
 422         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 423                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 424
 425                 if (!ldlm_same_flock_owner(lock, new))
 426                         break;
 427
 428                 if (lock->l_granted_mode == mode) {
 429                         /* If the modes are the same then we need to process
 430                          * locks that overlap OR adjoin the new lock. The extra
 431                          * logic condition is necessary to deal with arithmetic
 432                          * overflow and underflow.
 433                          */
 434                         if ((new->l_policy_data.l_flock.start >
 435                              (lock->l_policy_data.l_flock.end + 1))
 436                             && (lock->l_policy_data.l_flock.end !=
 437                                 OBD_OBJECT_EOF))
 438                                 continue;
 439
 440                         if ((new->l_policy_data.l_flock.end <
 441                              (lock->l_policy_data.l_flock.start - 1))
 442                             && (lock->l_policy_data.l_flock.start != 0))
 443                                 break;
 444
 445                         if (new->l_policy_data.l_flock.start <
 446                             lock->l_policy_data.l_flock.start) {
 447                                 lock->l_policy_data.l_flock.start =
 448                                         new->l_policy_data.l_flock.start;
 449                         } else {
 450                                 new->l_policy_data.l_flock.start =
 451                                         lock->l_policy_data.l_flock.start;
 452                         }
 453
 454                         if (new->l_policy_data.l_flock.end >
 455                             lock->l_policy_data.l_flock.end) {
 456                                 lock->l_policy_data.l_flock.end =
 457                                         new->l_policy_data.l_flock.end;
 458                         } else {
 459                                 new->l_policy_data.l_flock.end =
 460                                         lock->l_policy_data.l_flock.end;
 461                         }
 462
 463                         if (added) {
 464                                 ldlm_flock_destroy(lock, mode, *flags);
 465                         } else {
 466                                 new = lock;
 467                                 added = 1;
 468                         }
 469                         continue;
 470                 }
 471
 472                 if (new->l_policy_data.l_flock.start >
 473                     lock->l_policy_data.l_flock.end)
 474                         continue;
 475
 476                 if (new->l_policy_data.l_flock.end <
 477                     lock->l_policy_data.l_flock.start)
 478                         break;
 479
 480                 ++overlaps;
 481
 482                 if (new->l_policy_data.l_flock.start <=
 483                     lock->l_policy_data.l_flock.start) {
 484                         if (new->l_policy_data.l_flock.end <
 485                             lock->l_policy_data.l_flock.end) {
 486                                 lock->l_policy_data.l_flock.start =
 487                                         new->l_policy_data.l_flock.end + 1;
 488                                 break;
 489                         }
 490                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 491                         continue;
 492                 }
 493                 if (new->l_policy_data.l_flock.end >=
 494                     lock->l_policy_data.l_flock.end) {
 495                         lock->l_policy_data.l_flock.end =
 496                                 new->l_policy_data.l_flock.start - 1;
 497                         continue;
 498                 }
 499
 500                 /* split the existing lock into two locks */
 501
 502                 /* if this is an F_UNLCK operation then we could avoid
 503                  * allocating a new lock and use the req lock passed in
 504                  * with the request but this would complicate the reply
 505                  * processing since updates to req get reflected in the
 506                  * reply. The client side replays the lock request so
 507                  * it must see the original lock data in the reply.
 508                  */
 509
 510                 /* XXX - if ldlm_lock_new() can sleep we should
 511                  * release the lr_lock, allocate the new lock,
 512                  * and restart processing this lock.
 513                  */
 514                 if (new2 == NULL) {
 515                         unlock_res_and_lock(req);
 516                         new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
 517                                                 lock->l_granted_mode, &null_cbs,
 518                                                 NULL, 0, LVB_T_NONE);
 519                         lock_res_and_lock(req);
 520                         if (IS_ERR(new2)) {
 521                                 ldlm_flock_destroy(req, lock->l_granted_mode,
 522                                                    *flags);
 523                                 *err = PTR_ERR(new2);
 524                                 RETURN(LDLM_ITER_STOP);
 525                         }
 526                         goto reprocess;
 527                 }
 528
 529                 splitted = 1;
 530
 531                 new2->l_granted_mode = lock->l_granted_mode;
 532                 new2->l_policy_data.l_flock.pid =
 533                         new->l_policy_data.l_flock.pid;
 534                 new2->l_policy_data.l_flock.owner =
 535                         new->l_policy_data.l_flock.owner;
 536                 new2->l_policy_data.l_flock.start =
 537                         lock->l_policy_data.l_flock.start;
 538                 new2->l_policy_data.l_flock.end =
 539                         new->l_policy_data.l_flock.start - 1;
 540                 lock->l_policy_data.l_flock.start =
 541                         new->l_policy_data.l_flock.end + 1;
 542                 new2->l_conn_export = lock->l_conn_export;
 543                 if (lock->l_export != NULL) {
 544                         new2->l_export = class_export_lock_get(lock->l_export,
 545                                                                new2);
 546                         if (new2->l_export->exp_lock_hash &&
 547                             hlist_unhashed(&new2->l_exp_hash))
 548                                 cfs_hash_add(new2->l_export->exp_lock_hash,
 549                                              &new2->l_remote_handle,
 550                                              &new2->l_exp_hash);
 551                 }
 552                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 553                         ldlm_lock_addref_internal_nolock(new2,
 554                                                          lock->l_granted_mode);
 555
 556                 /* insert new2 at lock */
 557                 ldlm_resource_add_lock(res, ownlocks, new2);
 558                 LDLM_LOCK_RELEASE(new2);
 559                 break;
 560         }
 561
 562         /* if new2 is created but never used, destroy it*/
 563         if (splitted == 0 && new2 != NULL)
 564                 ldlm_lock_destroy_nolock(new2);
 565
 566         /* At this point we're granting the lock request. */
 567         req->l_granted_mode = req->l_req_mode;
 568
 569         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 570         if (!added) {
 571                 list_del_init(&req->l_res_link);
 572                 /* insert new lock before ownlocks in list. */
 573                 ldlm_resource_add_lock(res, ownlocks, req);
 574         }
 575
 576         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 577 #ifdef HAVE_SERVER_SUPPORT
 578                 if (intention == LDLM_PROCESS_ENQUEUE) {
 579                         /* If this is an unlock, reprocess the waitq and
 580                          * send completions ASTs for locks that can now be
 581                          * granted. The only problem with doing this
 582                          * reprocessing here is that the completion ASTs for
 583                          * newly granted locks will be sent before the unlock
 584                          * completion is sent. It shouldn't be an issue. Also
 585                          * note that ldlm_process_flock_lock() will recurse,
 586                          * but only once because 'intention' won't be
 587                          * LDLM_PROCESS_ENQUEUE from ldlm_reprocess_queue.
 588                          */
 589                         if ((mode == LCK_NL) && overlaps) {
 590                                 LIST_HEAD(rpc_list);
 591                                 int rc;
 592
 593 restart:
 594                                 ldlm_reprocess_queue(res, &res->lr_waiting,
 595                                                      &rpc_list,
 596                                                      LDLM_PROCESS_RESCAN, NULL);
 597
 598                                 unlock_res_and_lock(req);
 599                                 rc = ldlm_run_ast_work(ns, &rpc_list,
 600                                                        LDLM_WORK_CP_AST);
 601                                 lock_res_and_lock(req);
 602                                 if (rc == -ERESTART)
 603                                         GOTO(restart, rc);
 604                         }
 605                 } else {
 606                         LASSERT(req->l_completion_ast);
 607                         ldlm_add_ast_work_item(req, NULL, grant_work);
 608                 }
 609 #else /* !HAVE_SERVER_SUPPORT */
 610                 /* The only one possible case for client-side calls flock
 611                  * policy function is ldlm_flock_completion_ast inside which
 612                  * carries LDLM_FL_WAIT_NOREPROC flag.
 613                  */
 614                 CERROR("Illegal parameter for client-side-only module.\n");
 615                 LBUG();
 616 #endif /* HAVE_SERVER_SUPPORT */
 617         }
 618
 619         /* In case we're reprocessing the requested lock we can't destroy
 620          * it until after calling ldlm_add_ast_work_item() above so that laawi()
 621          * can bump the reference count on \a req. Otherwise \a req
 622          * could be freed before the completion AST can be sent.
 623          */
 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 /**
 632  * Flock completion callback function.
 633  *
 634  * \param lock [in,out]: A lock to be handled
 635  * \param flags    [in]: flags
 636  * \param *data    [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
 637  *
 638  * \retval 0    : success
 639  * \retval <0   : failure
 640  */
 641 int
 642 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 643 {
 644         struct file_lock *getlk = lock->l_ast_data;
 645         struct obd_device *obd;
 646         struct obd_import *imp = NULL;
 647         enum ldlm_error err;
 648         int rc = 0;
 649         ENTRY;
 650
 651         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4);
 652         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) {
 653                 lock_res_and_lock(lock);
 654                 lock->l_flags |= LDLM_FL_FAIL_LOC;
 655                 unlock_res_and_lock(lock);
 656                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT3, 4);
 657         }
 658         CDEBUG(D_DLMTRACE, "flags: %#llx data: %p getlk: %p\n",
 659                flags, data, getlk);
 660
 661         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 662
 663         if (flags & LDLM_FL_FAILED)
 664                 goto granted;
 665
 666         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
 667                 if (NULL == data)
 668                         /* mds granted the lock in the reply */
 669                         goto granted;
 670                 /* CP AST RPC: lock get granted, wake it up */
 671                 wake_up(&lock->l_waitq);
 672                 RETURN(0);
 673         }
 674
 675         LDLM_DEBUG(lock,
 676                    "client-side enqueue returned a blocked lock, sleeping");
 677         obd = class_exp2obd(lock->l_conn_export);
 678
 679         /* if this is a local lock, there is no import */
 680         if (obd)
 681                 imp = obd->u.cli.cl_import;
 682
 683         /* Go to sleep until the lock is granted. */
 684         rc = l_wait_event_abortable(lock->l_waitq,
 685                                     is_granted_or_cancelled(lock));
 686         if (rc < 0) {
 687                 /* take lock off the deadlock detection hash list. */
 688                 lock_res_and_lock(lock);
 689                 ldlm_flock_blocking_unlink(lock);
 690
 691                 /* client side - set flag to prevent lock from being
 692                  * put on LRU list
 693                  */
 694                 ldlm_set_cbpending(lock);
 695                 unlock_res_and_lock(lock);
 696
 697                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 698                            rc);
 699                 RETURN(rc);
 700         }
 701
 702 granted:
 703         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
 704
 705         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT4)) {
 706                 lock_res_and_lock(lock);
 707                 /* DEADLOCK is always set with CBPENDING */
 708                 lock->l_flags |= LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
 709                 unlock_res_and_lock(lock);
 710                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT4, 4);
 711         }
 712         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT5)) {
 713                 lock_res_and_lock(lock);
 714                 /* DEADLOCK is always set with CBPENDING */
 715                 lock->l_flags |= (LDLM_FL_FAIL_LOC |
 716                                   LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING);
 717                 unlock_res_and_lock(lock);
 718                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT5, 4);
 719         }
 720
 721         lock_res_and_lock(lock);
 722
 723
 724         /* Protect against race where lock could have been just destroyed
 725          * due to overlap in ldlm_process_flock_lock().
 726          */
 727         if (ldlm_is_destroyed(lock)) {
 728                 unlock_res_and_lock(lock);
 729                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 730
 731                 /* An error is still to be returned, to propagate it up to
 732                  * ldlm_cli_enqueue_fini() caller. */
 733                 RETURN(-EIO);
 734         }
 735
 736         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 737         ldlm_resource_unlink_lock(lock);
 738
 739         /* Import invalidation. We need to actually release the lock
 740          * references being held, so that it can go away. No point in
 741          * holding the lock even if app still believes it has it, since
 742          * server already dropped it anyway. Only for granted locks too.
 743          */
 744         /* Do the same for DEADLOCK'ed locks. */
 745         if (ldlm_is_failed(lock) || ldlm_is_flock_deadlock(lock)) {
 746                 int mode;
 747
 748                 if (flags & LDLM_FL_TEST_LOCK)
 749                         LASSERT(ldlm_is_test_lock(lock));
 750
 751                 if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock))
 752                         mode = getlk->fl_type;
 753                 else
 754                         mode = lock->l_granted_mode;
 755
 756                 if (ldlm_is_flock_deadlock(lock)) {
 757                         LDLM_DEBUG(lock, "client-side enqueue deadlock "
 758                                    "received");
 759                         rc = -EDEADLK;
 760                 }
 761                 ldlm_flock_destroy(lock, mode, LDLM_FL_WAIT_NOREPROC);
 762                 unlock_res_and_lock(lock);
 763
 764                 /* Need to wake up the waiter if we were evicted */
 765                 wake_up(&lock->l_waitq);
 766
 767                 /* An error is still to be returned, to propagate it up to
 768                  * ldlm_cli_enqueue_fini() caller.
 769                  */
 770                 RETURN(rc ? : -EIO);
 771         }
 772
 773         LDLM_DEBUG(lock, "client-side enqueue granted");
 774
 775         if (flags & LDLM_FL_TEST_LOCK) {
 776                 /*
 777                  * fcntl(F_GETLK) request
 778                  * The old mode was saved in getlk->fl_type so that if the mode
 779                  * in the lock changes we can decref the appropriate refcount.
 780                  */
 781                 LASSERT(ldlm_is_test_lock(lock));
 782                 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
 783                 switch (lock->l_granted_mode) {
 784                 case LCK_PR:
 785                         getlk->fl_type = F_RDLCK;
 786                         break;
 787                 case LCK_PW:
 788                         getlk->fl_type = F_WRLCK;
 789                         break;
 790                 default:
 791                         getlk->fl_type = F_UNLCK;
 792                 }
 793                 getlk->fl_pid = (pid_t)lock->l_policy_data.l_flock.pid;
 794                 getlk->fl_start = (loff_t)lock->l_policy_data.l_flock.start;
 795                 getlk->fl_end = (loff_t)lock->l_policy_data.l_flock.end;
 796         } else {
 797                 __u64 noreproc = LDLM_FL_WAIT_NOREPROC;
 798
 799                 /* We need to reprocess the lock to do merges or splits
 800                  * with existing locks owned by this process.
 801                  */
 802                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 803         }
 804         unlock_res_and_lock(lock);
 805         RETURN(rc);
 806 }
 807 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 808
 809 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 810                             void *data, int flag)
 811 {
 812         ENTRY;
 813
 814         LASSERT(lock);
 815         LASSERT(flag == LDLM_CB_CANCELING);
 816
 817         /* take lock off the deadlock detection hash list. */
 818         lock_res_and_lock(lock);
 819         ldlm_flock_blocking_unlink(lock);
 820         unlock_res_and_lock(lock);
 821         RETURN(0);
 822 }
 823
 824 void ldlm_flock_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
 825                                      union ldlm_policy_data *lpolicy)
 826 {
 827         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
 828         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
 829         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
 830         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
 831 }
 832
 833 void ldlm_flock_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
 834                                      union ldlm_wire_policy_data *wpolicy)
 835 {
 836         memset(wpolicy, 0, sizeof(*wpolicy));
 837         wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
 838         wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
 839         wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
 840         wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
 841 }
 842
 843 /*
 844  * Export handle<->flock hash operations.
 845  */
 846 static unsigned
 847 ldlm_export_flock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
 848 {
 849         return cfs_hash_u64_hash(*(__u64 *)key, mask);
 850 }
 851
 852 static void *
 853 ldlm_export_flock_key(struct hlist_node *hnode)
 854 {
 855         struct ldlm_lock *lock;
 856
 857         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 858         return &lock->l_policy_data.l_flock.owner;
 859 }
 860
 861 static int
 862 ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode)
 863 {
 864         return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
 865 }
 866
 867 static void *
 868 ldlm_export_flock_object(struct hlist_node *hnode)
 869 {
 870         return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 871 }
 872
 873 static void
 874 ldlm_export_flock_get(struct cfs_hash *hs, struct hlist_node *hnode)
 875 {
 876         struct ldlm_lock *lock;
 877         struct ldlm_flock *flock;
 878
 879         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 880         LDLM_LOCK_GET(lock);
 881
 882         flock = &lock->l_policy_data.l_flock;
 883         LASSERT(flock->blocking_export != NULL);
 884         class_export_get(flock->blocking_export);
 885         atomic_inc(&flock->blocking_refs);
 886 }
 887
 888 static void
 889 ldlm_export_flock_put(struct cfs_hash *hs, struct hlist_node *hnode)
 890 {
 891         struct ldlm_lock *lock;
 892         struct ldlm_flock *flock;
 893
 894         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 895
 896         flock = &lock->l_policy_data.l_flock;
 897         LASSERT(flock->blocking_export != NULL);
 898         class_export_put(flock->blocking_export);
 899         if (atomic_dec_and_test(&flock->blocking_refs)) {
 900                 flock->blocking_owner = 0;
 901                 flock->blocking_export = NULL;
 902         }
 903         LDLM_LOCK_RELEASE(lock);
 904 }
 905
 906 static struct cfs_hash_ops ldlm_export_flock_ops = {
 907         .hs_hash        = ldlm_export_flock_hash,
 908         .hs_key         = ldlm_export_flock_key,
 909         .hs_keycmp      = ldlm_export_flock_keycmp,
 910         .hs_object      = ldlm_export_flock_object,
 911         .hs_get         = ldlm_export_flock_get,
 912         .hs_put         = ldlm_export_flock_put,
 913         .hs_put_locked  = ldlm_export_flock_put,
 914 };
 915
 916 int ldlm_init_flock_export(struct obd_export *exp)
 917 {
 918         if( strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0)
 919                 RETURN(0);
 920
 921         exp->exp_flock_hash =
 922                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
 923                                 HASH_EXP_LOCK_CUR_BITS,
 924                                 HASH_EXP_LOCK_MAX_BITS,
 925                                 HASH_EXP_LOCK_BKT_BITS, 0,
 926                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 927                                 &ldlm_export_flock_ops,
 928                                 CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE);
 929         if (!exp->exp_flock_hash)
 930                 RETURN(-ENOMEM);
 931
 932         RETURN(0);
 933 }
 934
 935 void ldlm_destroy_flock_export(struct obd_export *exp)
 936 {
 937         ENTRY;
 938         if (exp->exp_flock_hash) {
 939                 cfs_hash_putref(exp->exp_flock_hash);
 940                 exp->exp_flock_hash = NULL;
 941         }
 942         EXIT;
 943 }