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 #ifdef HAVE_SERVER_SUPPORT
 155 /**
 156  * POSIX locks deadlock detection code.
 157  *
 158  * Given a new lock \a req and an existing lock \a bl_lock it conflicts
 159  * with, we need to iterate through all blocked POSIX locks for this
 160  * export and see if there is a deadlock condition arising. (i.e. when
 161  * one client holds a lock on something and want a lock on something
 162  * else and at the same time another client has the opposite situation).
 163  */
 164
 165 struct ldlm_flock_lookup_cb_data {
 166         __u64 *bl_owner;
 167         struct ldlm_lock *lock;
 168         struct obd_export *exp;
 169 };
 170
 171 static int ldlm_flock_lookup_cb(struct obd_export *exp, void *data)
 172 {
 173         struct ldlm_flock_lookup_cb_data *cb_data = data;
 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                 };
 206                 struct ptlrpc_connection *bl_exp_conn;
 207                 struct obd_export *bl_exp_new;
 208                 struct ldlm_lock *lock = NULL;
 209                 struct ldlm_flock *flock;
 210
 211                 bl_exp_conn = bl_exp->exp_connection;
 212                 if (bl_exp->exp_flock_hash != NULL) {
 213                         int found;
 214
 215                         found = obd_nid_export_for_each(bl_exp->exp_obd,
 216                                                         bl_exp_conn->c_peer.nid,
 217                                                         ldlm_flock_lookup_cb,
 218                                                         &cb_data);
 219                         if (found)
 220                                 lock = cb_data.lock;
 221                 }
 222                 if (lock == NULL)
 223                         break;
 224
 225                 class_export_put(bl_exp);
 226                 bl_exp = cb_data.exp;
 227
 228                 LASSERT(req != lock);
 229                 flock = &lock->l_policy_data.l_flock;
 230                 LASSERT(flock->owner == bl_owner);
 231                 bl_owner = flock->blocking_owner;
 232                 bl_exp_new = class_export_get(flock->blocking_export);
 233                 class_export_put(bl_exp);
 234
 235                 cfs_hash_put(bl_exp->exp_flock_hash, &lock->l_exp_flock_hash);
 236                 bl_exp = bl_exp_new;
 237
 238                 if (bl_exp->exp_failed)
 239                         break;
 240
 241                 if (bl_owner == req_owner &&
 242                     (bl_exp_conn->c_peer.nid ==
 243                      req_exp->exp_connection->c_peer.nid)) {
 244                         class_export_put(bl_exp);
 245                         return 1;
 246                 }
 247         }
 248         class_export_put(bl_exp);
 249
 250         return 0;
 251 }
 252
 253 static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
 254                                           struct list_head *work_list)
 255 {
 256         CDEBUG(D_INFO, "reprocess deadlock req=%p\n", lock);
 257
 258         if ((exp_connect_flags(lock->l_export) &
 259              OBD_CONNECT_FLOCK_DEAD) == 0) {
 260                 CERROR("deadlock found, but client doesn't support flock canceliation\n");
 261         } else {
 262                 LASSERT(lock->l_completion_ast);
 263                 LASSERT(!ldlm_is_ast_sent(lock));
 264                 lock->l_flags |= (LDLM_FL_AST_SENT | LDLM_FL_CANCEL_ON_BLOCK |
 265                                   LDLM_FL_FLOCK_DEADLOCK);
 266                 ldlm_flock_blocking_unlink(lock);
 267                 ldlm_resource_unlink_lock(lock);
 268                 ldlm_add_ast_work_item(lock, NULL, work_list);
 269         }
 270 }
 271 #endif /* HAVE_SERVER_SUPPORT */
 272
 273 /**
 274  * Process a granting attempt for flock lock.
 275  * Must be called under ns lock held.
 276  *
 277  * This function looks for any conflicts for \a lock in the granted or
 278  * waiting queues. The lock is granted if no conflicts are found in
 279  * either queue.
 280  */
 281 int
 282 ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags,
 283                         enum ldlm_process_intention intention,
 284                         enum ldlm_error *err, struct list_head *work_list)
 285 {
 286         struct ldlm_resource *res = req->l_resource;
 287         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
 288         struct list_head *tmp;
 289         struct list_head *ownlocks = NULL;
 290         struct ldlm_lock *lock = NULL;
 291         struct ldlm_lock *new = req;
 292         struct ldlm_lock *new2 = NULL;
 293         enum ldlm_mode mode = req->l_req_mode;
 294         int local = ns_is_client(ns);
 295         int added = (mode == LCK_NL);
 296         int overlaps = 0;
 297         int splitted = 0;
 298         const struct ldlm_callback_suite null_cbs = { NULL };
 299 #ifdef HAVE_SERVER_SUPPORT
 300         struct list_head *grant_work = (intention == LDLM_PROCESS_ENQUEUE ?
 301                                         NULL : work_list);
 302 #endif
 303         ENTRY;
 304
 305         CDEBUG(D_DLMTRACE, "flags %#llx owner %llu pid %u mode %u start "
 306                "%llu end %llu\n", *flags,
 307                new->l_policy_data.l_flock.owner,
 308                new->l_policy_data.l_flock.pid, mode,
 309                req->l_policy_data.l_flock.start,
 310                req->l_policy_data.l_flock.end);
 311
 312         *err = ELDLM_OK;
 313
 314         if (local) {
 315                 /* No blocking ASTs are sent to the clients for
 316                  * Posix file & record locks
 317                  */
 318                 req->l_blocking_ast = NULL;
 319         } else {
 320                 /* Called on the server for lock cancels. */
 321                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 322         }
 323
 324 reprocess:
 325         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 326                 /* This loop determines where this processes locks start
 327                  * in the resource lr_granted list.
 328                  */
 329                 list_for_each(tmp, &res->lr_granted) {
 330                         lock = list_entry(tmp, struct ldlm_lock,
 331                                           l_res_link);
 332                         if (ldlm_same_flock_owner(lock, req)) {
 333                                 ownlocks = tmp;
 334                                 break;
 335                         }
 336                 }
 337         }
 338 #ifdef HAVE_SERVER_SUPPORT
 339         else {
 340                 int reprocess_failed = 0;
 341                 lockmode_verify(mode);
 342
 343                 /* This loop determines if there are existing locks
 344                  * that conflict with the new lock request.
 345                  */
 346                 list_for_each(tmp, &res->lr_granted) {
 347                         lock = list_entry(tmp, struct ldlm_lock,
 348                                           l_res_link);
 349
 350                         if (ldlm_same_flock_owner(lock, req)) {
 351                                 if (!ownlocks)
 352                                         ownlocks = tmp;
 353                                 continue;
 354                         }
 355
 356                         /* locks are compatible, overlap doesn't matter */
 357                         if (lockmode_compat(lock->l_granted_mode, mode))
 358                                 continue;
 359
 360                         if (!ldlm_flocks_overlap(lock, req))
 361                                 continue;
 362
 363                         if (intention != LDLM_PROCESS_ENQUEUE) {
 364                                 reprocess_failed = 1;
 365                                 if (ldlm_flock_deadlock(req, lock)) {
 366                                         ldlm_flock_cancel_on_deadlock(
 367                                                 req, grant_work);
 368                                         RETURN(LDLM_ITER_CONTINUE);
 369                                 }
 370                                 continue;
 371                         }
 372
 373                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 374                                 ldlm_flock_destroy(req, mode, *flags);
 375                                 *err = -EAGAIN;
 376                                 RETURN(LDLM_ITER_STOP);
 377                         }
 378
 379                         if (*flags & LDLM_FL_TEST_LOCK) {
 380                                 ldlm_flock_destroy(req, mode, *flags);
 381                                 req->l_req_mode = lock->l_granted_mode;
 382                                 req->l_policy_data.l_flock.pid =
 383                                         lock->l_policy_data.l_flock.pid;
 384                                 req->l_policy_data.l_flock.start =
 385                                         lock->l_policy_data.l_flock.start;
 386                                 req->l_policy_data.l_flock.end =
 387                                         lock->l_policy_data.l_flock.end;
 388                                 *flags |= LDLM_FL_LOCK_CHANGED;
 389                                 RETURN(LDLM_ITER_STOP);
 390                         }
 391
 392                         /* add lock to blocking list before deadlock
 393                          * check to prevent race
 394                          */
 395                         ldlm_flock_blocking_link(req, lock);
 396
 397                         if (ldlm_flock_deadlock(req, lock)) {
 398                                 ldlm_flock_blocking_unlink(req);
 399                                 ldlm_flock_destroy(req, mode, *flags);
 400                                 *err = -EDEADLK;
 401                                 RETURN(LDLM_ITER_STOP);
 402                         }
 403
 404                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 405                         *flags |= LDLM_FL_BLOCK_GRANTED;
 406                         RETURN(LDLM_ITER_STOP);
 407                 }
 408                 if (reprocess_failed)
 409                         RETURN(LDLM_ITER_CONTINUE);
 410         }
 411
 412         if (*flags & LDLM_FL_TEST_LOCK) {
 413                 ldlm_flock_destroy(req, mode, *flags);
 414                 req->l_req_mode = LCK_NL;
 415                 *flags |= LDLM_FL_LOCK_CHANGED;
 416                 RETURN(LDLM_ITER_STOP);
 417         }
 418
 419         /* In case we had slept on this lock request take it off of the
 420          * deadlock detection hash list.
 421          */
 422         ldlm_flock_blocking_unlink(req);
 423 #endif /* HAVE_SERVER_SUPPORT */
 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 = 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                          */
 443                         if ((new->l_policy_data.l_flock.start >
 444                              (lock->l_policy_data.l_flock.end + 1))
 445                             && (lock->l_policy_data.l_flock.end !=
 446                                 OBD_OBJECT_EOF))
 447                                 continue;
 448
 449                         if ((new->l_policy_data.l_flock.end <
 450                              (lock->l_policy_data.l_flock.start - 1))
 451                             && (lock->l_policy_data.l_flock.start != 0))
 452                                 break;
 453
 454                         if (new->l_policy_data.l_flock.start <
 455                             lock->l_policy_data.l_flock.start) {
 456                                 lock->l_policy_data.l_flock.start =
 457                                         new->l_policy_data.l_flock.start;
 458                         } else {
 459                                 new->l_policy_data.l_flock.start =
 460                                         lock->l_policy_data.l_flock.start;
 461                         }
 462
 463                         if (new->l_policy_data.l_flock.end >
 464                             lock->l_policy_data.l_flock.end) {
 465                                 lock->l_policy_data.l_flock.end =
 466                                         new->l_policy_data.l_flock.end;
 467                         } else {
 468                                 new->l_policy_data.l_flock.end =
 469                                         lock->l_policy_data.l_flock.end;
 470                         }
 471
 472                         if (added) {
 473                                 ldlm_flock_destroy(lock, mode, *flags);
 474                         } else {
 475                                 new = lock;
 476                                 added = 1;
 477                         }
 478                         continue;
 479                 }
 480
 481                 if (new->l_policy_data.l_flock.start >
 482                     lock->l_policy_data.l_flock.end)
 483                         continue;
 484
 485                 if (new->l_policy_data.l_flock.end <
 486                     lock->l_policy_data.l_flock.start)
 487                         break;
 488
 489                 ++overlaps;
 490
 491                 if (new->l_policy_data.l_flock.start <=
 492                     lock->l_policy_data.l_flock.start) {
 493                         if (new->l_policy_data.l_flock.end <
 494                             lock->l_policy_data.l_flock.end) {
 495                                 lock->l_policy_data.l_flock.start =
 496                                         new->l_policy_data.l_flock.end + 1;
 497                                 break;
 498                         }
 499                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 500                         continue;
 501                 }
 502                 if (new->l_policy_data.l_flock.end >=
 503                     lock->l_policy_data.l_flock.end) {
 504                         lock->l_policy_data.l_flock.end =
 505                                 new->l_policy_data.l_flock.start - 1;
 506                         continue;
 507                 }
 508
 509                 /* split the existing lock into two locks */
 510
 511                 /* if this is an F_UNLCK operation then we could avoid
 512                  * allocating a new lock and use the req lock passed in
 513                  * with the request but this would complicate the reply
 514                  * processing since updates to req get reflected in the
 515                  * reply. The client side replays the lock request so
 516                  * it must see the original lock data in the reply.
 517                  */
 518
 519                 /* XXX - if ldlm_lock_new() can sleep we should
 520                  * release the lr_lock, allocate the new lock,
 521                  * and restart processing this lock.
 522                  */
 523                 if (new2 == NULL) {
 524                         unlock_res_and_lock(req);
 525                         new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
 526                                                 lock->l_granted_mode, &null_cbs,
 527                                                 NULL, 0, LVB_T_NONE);
 528                         lock_res_and_lock(req);
 529                         if (IS_ERR(new2)) {
 530                                 ldlm_flock_destroy(req, lock->l_granted_mode,
 531                                                    *flags);
 532                                 *err = PTR_ERR(new2);
 533                                 RETURN(LDLM_ITER_STOP);
 534                         }
 535                         goto reprocess;
 536                 }
 537
 538                 splitted = 1;
 539
 540                 new2->l_granted_mode = lock->l_granted_mode;
 541                 new2->l_policy_data.l_flock.pid =
 542                         new->l_policy_data.l_flock.pid;
 543                 new2->l_policy_data.l_flock.owner =
 544                         new->l_policy_data.l_flock.owner;
 545                 new2->l_policy_data.l_flock.start =
 546                         lock->l_policy_data.l_flock.start;
 547                 new2->l_policy_data.l_flock.end =
 548                         new->l_policy_data.l_flock.start - 1;
 549                 lock->l_policy_data.l_flock.start =
 550                         new->l_policy_data.l_flock.end + 1;
 551                 new2->l_conn_export = lock->l_conn_export;
 552                 if (lock->l_export != NULL) {
 553                         new2->l_export = class_export_lock_get(lock->l_export,
 554                                                                new2);
 555                         if (new2->l_export->exp_lock_hash &&
 556                             hlist_unhashed(&new2->l_exp_hash))
 557                                 cfs_hash_add(new2->l_export->exp_lock_hash,
 558                                              &new2->l_remote_handle,
 559                                              &new2->l_exp_hash);
 560                 }
 561                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 562                         ldlm_lock_addref_internal_nolock(new2,
 563                                                          lock->l_granted_mode);
 564
 565                 /* insert new2 at lock */
 566                 ldlm_resource_add_lock(res, ownlocks, new2);
 567                 LDLM_LOCK_RELEASE(new2);
 568                 break;
 569         }
 570
 571         /* if new2 is created but never used, destroy it*/
 572         if (splitted == 0 && new2 != NULL)
 573                 ldlm_lock_destroy_nolock(new2);
 574
 575         /* At this point we're granting the lock request. */
 576         req->l_granted_mode = req->l_req_mode;
 577
 578         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 579         if (!added) {
 580                 list_del_init(&req->l_res_link);
 581                 /* insert new lock before ownlocks in list. */
 582                 ldlm_resource_add_lock(res, ownlocks, req);
 583         }
 584
 585         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 586 #ifdef HAVE_SERVER_SUPPORT
 587                 if (intention == LDLM_PROCESS_ENQUEUE) {
 588                         /* If this is an unlock, reprocess the waitq and
 589                          * send completions ASTs for locks that can now be
 590                          * granted. The only problem with doing this
 591                          * reprocessing here is that the completion ASTs for
 592                          * newly granted locks will be sent before the unlock
 593                          * completion is sent. It shouldn't be an issue. Also
 594                          * note that ldlm_process_flock_lock() will recurse,
 595                          * but only once because 'intention' won't be
 596                          * LDLM_PROCESS_ENQUEUE from ldlm_reprocess_queue.
 597                          */
 598                         if ((mode == LCK_NL) && overlaps) {
 599                                 LIST_HEAD(rpc_list);
 600                                 int rc;
 601
 602 restart:
 603                                 ldlm_reprocess_queue(res, &res->lr_waiting,
 604                                                      &rpc_list,
 605                                                      LDLM_PROCESS_RESCAN, NULL);
 606
 607                                 unlock_res_and_lock(req);
 608                                 rc = ldlm_run_ast_work(ns, &rpc_list,
 609                                                        LDLM_WORK_CP_AST);
 610                                 lock_res_and_lock(req);
 611                                 if (rc == -ERESTART)
 612                                         GOTO(restart, rc);
 613                         }
 614                 } else {
 615                         LASSERT(req->l_completion_ast);
 616                         ldlm_add_ast_work_item(req, NULL, grant_work);
 617                 }
 618 #else /* !HAVE_SERVER_SUPPORT */
 619                 /* The only one possible case for client-side calls flock
 620                  * policy function is ldlm_flock_completion_ast inside which
 621                  * carries LDLM_FL_WAIT_NOREPROC flag.
 622                  */
 623                 CERROR("Illegal parameter for client-side-only module.\n");
 624                 LBUG();
 625 #endif /* HAVE_SERVER_SUPPORT */
 626         }
 627
 628         /* In case we're reprocessing the requested lock we can't destroy
 629          * it until after calling ldlm_add_ast_work_item() above so that laawi()
 630          * can bump the reference count on \a req. Otherwise \a req
 631          * could be freed before the completion AST can be sent.
 632          */
 633         if (added)
 634                 ldlm_flock_destroy(req, mode, *flags);
 635
 636         ldlm_resource_dump(D_INFO, res);
 637         RETURN(LDLM_ITER_CONTINUE);
 638 }
 639
 640 /**
 641  * Flock completion callback function.
 642  *
 643  * \param lock [in,out]: A lock to be handled
 644  * \param flags    [in]: flags
 645  * \param *data    [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
 646  *
 647  * \retval 0    : success
 648  * \retval <0   : failure
 649  */
 650 int
 651 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 652 {
 653         struct file_lock *getlk = lock->l_ast_data;
 654         struct obd_device *obd;
 655         struct obd_import *imp = NULL;
 656         enum ldlm_error err;
 657         int rc = 0;
 658         ENTRY;
 659
 660         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4);
 661         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) {
 662                 lock_res_and_lock(lock);
 663                 lock->l_flags |= LDLM_FL_FAIL_LOC;
 664                 unlock_res_and_lock(lock);
 665                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT3, 4);
 666         }
 667         CDEBUG(D_DLMTRACE, "flags: %#llx data: %p getlk: %p\n",
 668                flags, data, getlk);
 669
 670         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 671
 672         if (flags & LDLM_FL_FAILED)
 673                 goto granted;
 674
 675         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
 676                 if (NULL == data)
 677                         /* mds granted the lock in the reply */
 678                         goto granted;
 679                 /* CP AST RPC: lock get granted, wake it up */
 680                 wake_up(&lock->l_waitq);
 681                 RETURN(0);
 682         }
 683
 684         LDLM_DEBUG(lock,
 685                    "client-side enqueue returned a blocked lock, sleeping");
 686         obd = class_exp2obd(lock->l_conn_export);
 687
 688         /* if this is a local lock, there is no import */
 689         if (obd)
 690                 imp = obd->u.cli.cl_import;
 691
 692         /* Go to sleep until the lock is granted. */
 693         rc = l_wait_event_abortable(lock->l_waitq,
 694                                     is_granted_or_cancelled(lock));
 695         if (rc < 0) {
 696                 /* take lock off the deadlock detection hash list. */
 697                 lock_res_and_lock(lock);
 698                 ldlm_flock_blocking_unlink(lock);
 699
 700                 /* client side - set flag to prevent lock from being
 701                  * put on LRU list
 702                  */
 703                 ldlm_set_cbpending(lock);
 704                 unlock_res_and_lock(lock);
 705
 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          */
 753         /* Do the same for DEADLOCK'ed locks. */
 754         if (ldlm_is_failed(lock) || ldlm_is_flock_deadlock(lock)) {
 755                 int mode;
 756
 757                 if (flags & LDLM_FL_TEST_LOCK)
 758                         LASSERT(ldlm_is_test_lock(lock));
 759
 760                 if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock))
 761                         mode = getlk->fl_type;
 762                 else
 763                         mode = lock->l_granted_mode;
 764
 765                 if (ldlm_is_flock_deadlock(lock)) {
 766                         LDLM_DEBUG(lock, "client-side enqueue deadlock "
 767                                    "received");
 768                         rc = -EDEADLK;
 769                 }
 770                 ldlm_flock_destroy(lock, mode, LDLM_FL_WAIT_NOREPROC);
 771                 unlock_res_and_lock(lock);
 772
 773                 /* Need to wake up the waiter if we were evicted */
 774                 wake_up(&lock->l_waitq);
 775
 776                 /* An error is still to be returned, to propagate it up to
 777                  * ldlm_cli_enqueue_fini() caller.
 778                  */
 779                 RETURN(rc ? : -EIO);
 780         }
 781
 782         LDLM_DEBUG(lock, "client-side enqueue granted");
 783
 784         if (flags & LDLM_FL_TEST_LOCK) {
 785                 /*
 786                  * fcntl(F_GETLK) request
 787                  * The old mode was saved in getlk->fl_type so that if the mode
 788                  * in the lock changes we can decref the appropriate refcount.
 789                  */
 790                 LASSERT(ldlm_is_test_lock(lock));
 791                 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
 792                 switch (lock->l_granted_mode) {
 793                 case LCK_PR:
 794                         getlk->fl_type = F_RDLCK;
 795                         break;
 796                 case LCK_PW:
 797                         getlk->fl_type = F_WRLCK;
 798                         break;
 799                 default:
 800                         getlk->fl_type = F_UNLCK;
 801                 }
 802                 getlk->fl_pid = (pid_t)lock->l_policy_data.l_flock.pid;
 803                 getlk->fl_start = (loff_t)lock->l_policy_data.l_flock.start;
 804                 getlk->fl_end = (loff_t)lock->l_policy_data.l_flock.end;
 805         } else {
 806                 __u64 noreproc = LDLM_FL_WAIT_NOREPROC;
 807
 808                 /* We need to reprocess the lock to do merges or splits
 809                  * with existing locks owned by this process.
 810                  */
 811                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 812         }
 813         unlock_res_and_lock(lock);
 814         RETURN(rc);
 815 }
 816 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 817
 818 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 819                             void *data, int flag)
 820 {
 821         ENTRY;
 822
 823         LASSERT(lock);
 824         LASSERT(flag == LDLM_CB_CANCELING);
 825
 826         /* take lock off the deadlock detection hash list. */
 827         lock_res_and_lock(lock);
 828         ldlm_flock_blocking_unlink(lock);
 829         unlock_res_and_lock(lock);
 830         RETURN(0);
 831 }
 832
 833 void ldlm_flock_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
 834                                      union ldlm_policy_data *lpolicy)
 835 {
 836         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
 837         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
 838         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
 839         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
 840 }
 841
 842 void ldlm_flock_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
 843                                      union ldlm_wire_policy_data *wpolicy)
 844 {
 845         memset(wpolicy, 0, sizeof(*wpolicy));
 846         wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
 847         wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
 848         wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
 849         wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
 850 }
 851
 852 /*
 853  * Export handle<->flock hash operations.
 854  */
 855 static unsigned
 856 ldlm_export_flock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
 857 {
 858         return cfs_hash_u64_hash(*(__u64 *)key, mask);
 859 }
 860
 861 static void *
 862 ldlm_export_flock_key(struct hlist_node *hnode)
 863 {
 864         struct ldlm_lock *lock;
 865
 866         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 867         return &lock->l_policy_data.l_flock.owner;
 868 }
 869
 870 static int
 871 ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode)
 872 {
 873         return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
 874 }
 875
 876 static void *
 877 ldlm_export_flock_object(struct hlist_node *hnode)
 878 {
 879         return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 880 }
 881
 882 static void
 883 ldlm_export_flock_get(struct cfs_hash *hs, struct hlist_node *hnode)
 884 {
 885         struct ldlm_lock *lock;
 886         struct ldlm_flock *flock;
 887
 888         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 889         LDLM_LOCK_GET(lock);
 890
 891         flock = &lock->l_policy_data.l_flock;
 892         LASSERT(flock->blocking_export != NULL);
 893         class_export_get(flock->blocking_export);
 894         atomic_inc(&flock->blocking_refs);
 895 }
 896
 897 static void
 898 ldlm_export_flock_put(struct cfs_hash *hs, struct hlist_node *hnode)
 899 {
 900         struct ldlm_lock *lock;
 901         struct ldlm_flock *flock;
 902
 903         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 904
 905         flock = &lock->l_policy_data.l_flock;
 906         LASSERT(flock->blocking_export != NULL);
 907         class_export_put(flock->blocking_export);
 908         if (atomic_dec_and_test(&flock->blocking_refs)) {
 909                 flock->blocking_owner = 0;
 910                 flock->blocking_export = NULL;
 911         }
 912         LDLM_LOCK_RELEASE(lock);
 913 }
 914
 915 static struct cfs_hash_ops ldlm_export_flock_ops = {
 916         .hs_hash        = ldlm_export_flock_hash,
 917         .hs_key         = ldlm_export_flock_key,
 918         .hs_keycmp      = ldlm_export_flock_keycmp,
 919         .hs_object      = ldlm_export_flock_object,
 920         .hs_get         = ldlm_export_flock_get,
 921         .hs_put         = ldlm_export_flock_put,
 922         .hs_put_locked  = ldlm_export_flock_put,
 923 };
 924
 925 int ldlm_init_flock_export(struct obd_export *exp)
 926 {
 927         if( strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0)
 928                 RETURN(0);
 929
 930         exp->exp_flock_hash =
 931                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
 932                                 HASH_EXP_LOCK_CUR_BITS,
 933                                 HASH_EXP_LOCK_MAX_BITS,
 934                                 HASH_EXP_LOCK_BKT_BITS, 0,
 935                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 936                                 &ldlm_export_flock_ops,
 937                                 CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE);
 938         if (!exp->exp_flock_hash)
 939                 RETURN(-ENOMEM);
 940
 941         RETURN(0);
 942 }
 943
 944 void ldlm_destroy_flock_export(struct obd_export *exp)
 945 {
 946         ENTRY;
 947         if (exp->exp_flock_hash) {
 948                 cfs_hash_putref(exp->exp_flock_hash);
 949                 exp->exp_flock_hash = NULL;
 950         }
 951         EXIT;
 952 }