lustre/ldlm/ldlm_flock.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003 Hewlett-Packard Development Company LP.
  24  * Developed under the sponsorship of the US Government under
  25  * Subcontract No. B514193
  26  *
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2010, 2017, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  */
  35
  36 /**
  37  * This file implements POSIX lock type for Lustre.
  38  * Its policy properties are start and end of extent and PID.
  39  *
  40  * These locks are only done through MDS due to POSIX semantics requiring
  41  * e.g. that locks could be only partially released and as such split into
  42  * two parts, and also that two adjacent locks from the same process may be
  43  * merged into a single wider lock.
  44  *
  45  * Lock modes are mapped like this:
  46  * PR and PW for READ and WRITE locks
  47  * NL to request a releasing of a portion of the lock
  48  *
  49  * These flock locks never timeout.
  50  */
  51
  52 #define DEBUG_SUBSYSTEM S_LDLM
  53
  54 #include <linux/list.h>
  55 #include <lustre_dlm.h>
  56 #include <obd_support.h>
  57 #include <obd_class.h>
  58 #include <lustre_lib.h>
  59
  60 #include "ldlm_internal.h"
  61
  62 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
  63                             void *data, int flag);
  64
  65 /**
  66  * list_for_remaining_safe - iterate over the remaining entries in a list
  67  *              and safeguard against removal of a list entry.
  68  * \param pos   the &struct list_head to use as a loop counter. pos MUST
  69  *              have been initialized prior to using it in this macro.
  70  * \param n     another &struct list_head to use as temporary storage
  71  * \param head  the head for your list.
  72  */
  73 #define list_for_remaining_safe(pos, n, head) \
  74         for (n = pos->next; pos != (head); pos = n, n = pos->next)
  75
  76 static inline int
  77 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
  78 {
  79         return ((new->l_policy_data.l_flock.owner ==
  80                  lock->l_policy_data.l_flock.owner) &&
  81                 (new->l_export == lock->l_export));
  82 }
  83
  84 static inline int
  85 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
  86 {
  87         return ((new->l_policy_data.l_flock.start <=
  88                  lock->l_policy_data.l_flock.end) &&
  89                 (new->l_policy_data.l_flock.end >=
  90                  lock->l_policy_data.l_flock.start));
  91 }
  92
  93 static inline void ldlm_flock_blocking_link(struct ldlm_lock *req,
  94                                             struct ldlm_lock *lock)
  95 {
  96         /* For server only */
  97         if (req->l_export == NULL)
  98                 return;
  99
 100         LASSERT(hlist_unhashed(&req->l_exp_flock_hash));
 101
 102         req->l_policy_data.l_flock.blocking_owner =
 103                 lock->l_policy_data.l_flock.owner;
 104         req->l_policy_data.l_flock.blocking_export =
 105                 lock->l_export;
 106         atomic_set(&req->l_policy_data.l_flock.blocking_refs, 0);
 107
 108         cfs_hash_add(req->l_export->exp_flock_hash,
 109                      &req->l_policy_data.l_flock.owner,
 110                      &req->l_exp_flock_hash);
 111 }
 112
 113 static inline void ldlm_flock_blocking_unlink(struct ldlm_lock *req)
 114 {
 115         /* For server only */
 116         if (req->l_export == NULL)
 117                 return;
 118
 119         check_res_locked(req->l_resource);
 120         if (req->l_export->exp_flock_hash != NULL &&
 121             !hlist_unhashed(&req->l_exp_flock_hash))
 122                 cfs_hash_del(req->l_export->exp_flock_hash,
 123                              &req->l_policy_data.l_flock.owner,
 124                              &req->l_exp_flock_hash);
 125 }
 126
 127 static inline void
 128 ldlm_flock_destroy(struct ldlm_lock *lock, enum ldlm_mode mode, __u64 flags)
 129 {
 130         ENTRY;
 131
 132         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: %#llx)",
 133                    mode, flags);
 134
 135         /* Safe to not lock here, since it should be empty anyway */
 136         LASSERT(hlist_unhashed(&lock->l_exp_flock_hash));
 137
 138         list_del_init(&lock->l_res_link);
 139         if (flags == LDLM_FL_WAIT_NOREPROC) {
 140                 /* client side - set a flag to prevent sending a CANCEL */
 141                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
 142
 143                 /* when reaching here, it is under lock_res_and_lock(). Thus,
 144                  * need call the nolock version of ldlm_lock_decref_internal
 145                  */
 146                 ldlm_lock_decref_internal_nolock(lock, mode);
 147         }
 148
 149         ldlm_lock_destroy_nolock(lock);
 150         EXIT;
 151 }
 152
 153 #ifdef HAVE_SERVER_SUPPORT
 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 obd_export *exp, void *data)
 171 {
 172         struct ldlm_flock_lookup_cb_data *cb_data = data;
 173         struct ldlm_lock *lock;
 174
 175         if (exp->exp_failed)
 176                 return 0;
 177
 178         lock = cfs_hash_lookup(exp->exp_flock_hash, cb_data->bl_owner);
 179         if (lock == NULL)
 180                 return 0;
 181
 182         /* Stop on first found lock. Same process can't sleep twice */
 183         cb_data->lock = lock;
 184         cb_data->exp = class_export_get(exp);
 185
 186         return 1;
 187 }
 188
 189 static int
 190 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *bl_lock)
 191 {
 192         struct obd_export *req_exp = req->l_export;
 193         struct obd_export *bl_exp = bl_lock->l_export;
 194         __u64 req_owner = req->l_policy_data.l_flock.owner;
 195         __u64 bl_owner = bl_lock->l_policy_data.l_flock.owner;
 196
 197         /* For server only */
 198         if (req_exp == NULL)
 199                 return 0;
 200
 201         class_export_get(bl_exp);
 202         while (1) {
 203                 struct ldlm_flock_lookup_cb_data cb_data = {
 204                         .bl_owner = &bl_owner,
 205                         .lock = NULL,
 206                         .exp = NULL,
 207                 };
 208                 struct ptlrpc_connection *bl_exp_conn;
 209                 struct obd_export *bl_exp_new;
 210                 struct ldlm_lock *lock = NULL;
 211                 struct ldlm_flock *flock;
 212
 213                 bl_exp_conn = bl_exp->exp_connection;
 214                 if (bl_exp->exp_flock_hash != NULL) {
 215                         int found;
 216
 217                         found = obd_nid_export_for_each(bl_exp->exp_obd,
 218                                                         &bl_exp_conn->c_peer.nid,
 219                                                         ldlm_flock_lookup_cb,
 220                                                         &cb_data);
 221                         if (found)
 222                                 lock = cb_data.lock;
 223                 }
 224                 if (lock == NULL)
 225                         break;
 226
 227                 class_export_put(bl_exp);
 228                 bl_exp = cb_data.exp;
 229
 230                 LASSERT(req != lock);
 231                 flock = &lock->l_policy_data.l_flock;
 232                 LASSERT(flock->owner == bl_owner);
 233                 bl_owner = flock->blocking_owner;
 234                 bl_exp_new = class_export_get(flock->blocking_export);
 235                 class_export_put(bl_exp);
 236
 237                 cfs_hash_put(bl_exp->exp_flock_hash, &lock->l_exp_flock_hash);
 238                 bl_exp = bl_exp_new;
 239
 240                 if (bl_exp->exp_failed)
 241                         break;
 242
 243                 if (bl_owner == req_owner &&
 244                     nid_same(&bl_exp_conn->c_peer.nid,
 245                               &req_exp->exp_connection->c_peer.nid)) {
 246                         class_export_put(bl_exp);
 247                         return 1;
 248                 }
 249         }
 250         class_export_put(bl_exp);
 251
 252         return 0;
 253 }
 254
 255 static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
 256                                           struct list_head *work_list)
 257 {
 258         CDEBUG(D_INFO, "reprocess deadlock req=%p\n", lock);
 259
 260         if ((exp_connect_flags(lock->l_export) &
 261              OBD_CONNECT_FLOCK_DEAD) == 0) {
 262                 CERROR("deadlock found, but client doesn't support flock canceliation\n");
 263         } else {
 264                 LASSERT(lock->l_completion_ast);
 265                 LASSERT(!ldlm_is_ast_sent(lock));
 266                 lock->l_flags |= (LDLM_FL_AST_SENT | LDLM_FL_CANCEL_ON_BLOCK |
 267                                   LDLM_FL_FLOCK_DEADLOCK);
 268                 ldlm_flock_blocking_unlink(lock);
 269                 ldlm_resource_unlink_lock(lock);
 270                 ldlm_add_ast_work_item(lock, NULL, work_list);
 271         }
 272 }
 273 #endif /* HAVE_SERVER_SUPPORT */
 274
 275 /**
 276  * Process a granting attempt for flock lock.
 277  * Must be called under ns lock held.
 278  *
 279  * This function looks for any conflicts for \a lock in the granted or
 280  * waiting queues. The lock is granted if no conflicts are found in
 281  * either queue.
 282  */
 283 int
 284 ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags,
 285                         enum ldlm_process_intention intention,
 286                         enum ldlm_error *err, struct list_head *work_list)
 287 {
 288         struct ldlm_resource *res = req->l_resource;
 289         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
 290         struct list_head *tmp;
 291         struct list_head *ownlocks = NULL;
 292         struct ldlm_lock *lock = NULL;
 293         struct ldlm_lock *new = req;
 294         struct ldlm_lock *new2 = NULL;
 295         enum ldlm_mode mode = req->l_req_mode;
 296         int local = ns_is_client(ns);
 297         int added = (mode == LCK_NL);
 298         int overlaps = 0;
 299         int splitted = 0;
 300         const struct ldlm_callback_suite null_cbs = { NULL };
 301 #ifdef HAVE_SERVER_SUPPORT
 302         struct list_head *grant_work = (intention == LDLM_PROCESS_ENQUEUE ?
 303                                         NULL : work_list);
 304 #endif
 305         ENTRY;
 306
 307         CDEBUG(D_DLMTRACE, "flags %#llx owner %llu pid %u mode %u start "
 308                "%llu end %llu\n", *flags,
 309                new->l_policy_data.l_flock.owner,
 310                new->l_policy_data.l_flock.pid, mode,
 311                req->l_policy_data.l_flock.start,
 312                req->l_policy_data.l_flock.end);
 313
 314         *err = ELDLM_OK;
 315
 316         if (local) {
 317                 /* No blocking ASTs are sent to the clients for
 318                  * Posix file & record locks
 319                  */
 320                 req->l_blocking_ast = NULL;
 321         } else {
 322                 /* Called on the server for lock cancels. */
 323                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 324         }
 325
 326 reprocess:
 327         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 328                 /* This loop determines where this processes locks start
 329                  * in the resource lr_granted list.
 330                  */
 331                 list_for_each(tmp, &res->lr_granted) {
 332                         lock = list_entry(tmp, struct ldlm_lock,
 333                                           l_res_link);
 334                         if (ldlm_same_flock_owner(lock, req)) {
 335                                 ownlocks = tmp;
 336                                 break;
 337                         }
 338                 }
 339         }
 340 #ifdef HAVE_SERVER_SUPPORT
 341         else {
 342                 int reprocess_failed = 0;
 343                 lockmode_verify(mode);
 344
 345                 /* This loop determines if there are existing locks
 346                  * that conflict with the new lock request.
 347                  */
 348                 list_for_each(tmp, &res->lr_granted) {
 349                         lock = list_entry(tmp, struct ldlm_lock,
 350                                           l_res_link);
 351
 352                         if (ldlm_same_flock_owner(lock, req)) {
 353                                 if (!ownlocks)
 354                                         ownlocks = tmp;
 355                                 continue;
 356                         }
 357
 358                         if (req->l_req_mode == LCK_PR &&
 359                             lock->l_granted_mode == LCK_PR &&
 360                             lock->l_policy_data.l_flock.start <=
 361                                 req->l_policy_data.l_flock.start &&
 362                             lock->l_policy_data.l_flock.end >=
 363                                 req->l_policy_data.l_flock.end) {
 364                                 /* there can't be granted WR lock */
 365                                 break;
 366                         }
 367                         /* locks are compatible, overlap doesn't matter */
 368                         if (lockmode_compat(lock->l_granted_mode, mode))
 369                                 continue;
 370
 371                         if (!ldlm_flocks_overlap(lock, req))
 372                                 continue;
 373
 374                         if (intention != LDLM_PROCESS_ENQUEUE) {
 375                                 if (ldlm_flock_deadlock(req, lock)) {
 376                                         ldlm_flock_cancel_on_deadlock(
 377                                                 req, grant_work);
 378                                         RETURN(LDLM_ITER_CONTINUE);
 379                                 }
 380                                 reprocess_failed = 1;
 381                                 break;
 382                         }
 383
 384                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 385                                 ldlm_flock_destroy(req, mode, *flags);
 386                                 *err = -EAGAIN;
 387                                 RETURN(LDLM_ITER_STOP);
 388                         }
 389
 390                         if (*flags & LDLM_FL_TEST_LOCK) {
 391                                 ldlm_flock_destroy(req, mode, *flags);
 392                                 req->l_req_mode = lock->l_granted_mode;
 393                                 req->l_policy_data.l_flock.pid =
 394                                         lock->l_policy_data.l_flock.pid;
 395                                 req->l_policy_data.l_flock.start =
 396                                         lock->l_policy_data.l_flock.start;
 397                                 req->l_policy_data.l_flock.end =
 398                                         lock->l_policy_data.l_flock.end;
 399                                 *flags |= LDLM_FL_LOCK_CHANGED;
 400                                 RETURN(LDLM_ITER_STOP);
 401                         }
 402
 403                         /* add lock to blocking list before deadlock
 404                          * check to prevent race
 405                          */
 406                         ldlm_flock_blocking_link(req, lock);
 407
 408                         if (ldlm_flock_deadlock(req, lock)) {
 409                                 ldlm_flock_blocking_unlink(req);
 410                                 ldlm_flock_destroy(req, mode, *flags);
 411                                 *err = -EDEADLK;
 412                                 RETURN(LDLM_ITER_STOP);
 413                         }
 414
 415                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 416                         *flags |= LDLM_FL_BLOCK_GRANTED;
 417                         RETURN(LDLM_ITER_STOP);
 418                 }
 419                 if (reprocess_failed)
 420                         RETURN(LDLM_ITER_CONTINUE);
 421         }
 422
 423         if (*flags & LDLM_FL_TEST_LOCK) {
 424                 ldlm_flock_destroy(req, mode, *flags);
 425                 req->l_req_mode = LCK_NL;
 426                 *flags |= LDLM_FL_LOCK_CHANGED;
 427                 RETURN(LDLM_ITER_STOP);
 428         }
 429
 430         /* In case we had slept on this lock request take it off of the
 431          * deadlock detection hash list.
 432          */
 433         ldlm_flock_blocking_unlink(req);
 434 #endif /* HAVE_SERVER_SUPPORT */
 435
 436         /* Scan the locks owned by this process that overlap this request.
 437          * We may have to merge or split existing locks.
 438          */
 439         if (!ownlocks)
 440                 ownlocks = &res->lr_granted;
 441
 442         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 443                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 444
 445                 if (!ldlm_same_flock_owner(lock, new))
 446                         break;
 447
 448                 if (lock->l_granted_mode == mode) {
 449                         /* If the modes are the same then we need to process
 450                          * locks that overlap OR adjoin the new lock. The extra
 451                          * logic condition is necessary to deal with arithmetic
 452                          * overflow and underflow.
 453                          */
 454                         if ((new->l_policy_data.l_flock.start >
 455                              (lock->l_policy_data.l_flock.end + 1))
 456                             && (lock->l_policy_data.l_flock.end !=
 457                                 OBD_OBJECT_EOF))
 458                                 continue;
 459
 460                         if ((new->l_policy_data.l_flock.end <
 461                              (lock->l_policy_data.l_flock.start - 1))
 462                             && (lock->l_policy_data.l_flock.start != 0))
 463                                 break;
 464
 465                         if (new->l_policy_data.l_flock.start <
 466                             lock->l_policy_data.l_flock.start) {
 467                                 lock->l_policy_data.l_flock.start =
 468                                         new->l_policy_data.l_flock.start;
 469                         } else {
 470                                 new->l_policy_data.l_flock.start =
 471                                         lock->l_policy_data.l_flock.start;
 472                         }
 473
 474                         if (new->l_policy_data.l_flock.end >
 475                             lock->l_policy_data.l_flock.end) {
 476                                 lock->l_policy_data.l_flock.end =
 477                                         new->l_policy_data.l_flock.end;
 478                         } else {
 479                                 new->l_policy_data.l_flock.end =
 480                                         lock->l_policy_data.l_flock.end;
 481                         }
 482
 483                         if (added) {
 484                                 ldlm_flock_destroy(lock, mode, *flags);
 485                         } else {
 486                                 new = lock;
 487                                 added = 1;
 488                         }
 489                         continue;
 490                 }
 491
 492                 if (new->l_policy_data.l_flock.start >
 493                     lock->l_policy_data.l_flock.end)
 494                         continue;
 495
 496                 if (new->l_policy_data.l_flock.end <
 497                     lock->l_policy_data.l_flock.start)
 498                         break;
 499
 500                 ++overlaps;
 501
 502                 if (new->l_policy_data.l_flock.start <=
 503                     lock->l_policy_data.l_flock.start) {
 504                         if (new->l_policy_data.l_flock.end <
 505                             lock->l_policy_data.l_flock.end) {
 506                                 lock->l_policy_data.l_flock.start =
 507                                         new->l_policy_data.l_flock.end + 1;
 508                                 break;
 509                         }
 510                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 511                         continue;
 512                 }
 513                 if (new->l_policy_data.l_flock.end >=
 514                     lock->l_policy_data.l_flock.end) {
 515                         lock->l_policy_data.l_flock.end =
 516                                 new->l_policy_data.l_flock.start - 1;
 517                         continue;
 518                 }
 519
 520                 /* split the existing lock into two locks */
 521
 522                 /* if this is an F_UNLCK operation then we could avoid
 523                  * allocating a new lock and use the req lock passed in
 524                  * with the request but this would complicate the reply
 525                  * processing since updates to req get reflected in the
 526                  * reply. The client side replays the lock request so
 527                  * it must see the original lock data in the reply.
 528                  */
 529
 530                 /* XXX - if ldlm_lock_new() can sleep we should
 531                  * release the lr_lock, allocate the new lock,
 532                  * and restart processing this lock.
 533                  */
 534                 if (new2 == NULL) {
 535                         unlock_res_and_lock(req);
 536                         new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
 537                                                 lock->l_granted_mode, &null_cbs,
 538                                                 NULL, 0, LVB_T_NONE);
 539                         lock_res_and_lock(req);
 540                         if (IS_ERR(new2)) {
 541                                 ldlm_flock_destroy(req, lock->l_granted_mode,
 542                                                    *flags);
 543                                 *err = PTR_ERR(new2);
 544                                 RETURN(LDLM_ITER_STOP);
 545                         }
 546                         goto reprocess;
 547                 }
 548
 549                 splitted = 1;
 550
 551                 new2->l_granted_mode = lock->l_granted_mode;
 552                 new2->l_policy_data.l_flock.pid =
 553                         new->l_policy_data.l_flock.pid;
 554                 new2->l_policy_data.l_flock.owner =
 555                         new->l_policy_data.l_flock.owner;
 556                 new2->l_policy_data.l_flock.start =
 557                         lock->l_policy_data.l_flock.start;
 558                 new2->l_policy_data.l_flock.end =
 559                         new->l_policy_data.l_flock.start - 1;
 560                 lock->l_policy_data.l_flock.start =
 561                         new->l_policy_data.l_flock.end + 1;
 562                 new2->l_conn_export = lock->l_conn_export;
 563                 if (lock->l_export != NULL) {
 564                         new2->l_export = class_export_lock_get(lock->l_export,
 565                                                                new2);
 566                         if (new2->l_export->exp_lock_hash &&
 567                             hlist_unhashed(&new2->l_exp_hash))
 568                                 cfs_hash_add(new2->l_export->exp_lock_hash,
 569                                              &new2->l_remote_handle,
 570                                              &new2->l_exp_hash);
 571                 }
 572                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 573                         ldlm_lock_addref_internal_nolock(new2,
 574                                                          lock->l_granted_mode);
 575
 576                 /* insert new2 at lock */
 577                 ldlm_resource_add_lock(res, ownlocks, new2);
 578                 LDLM_LOCK_RELEASE(new2);
 579                 break;
 580         }
 581
 582         /* if new2 is created but never used, destroy it*/
 583         if (splitted == 0 && new2 != NULL)
 584                 ldlm_lock_destroy_nolock(new2);
 585
 586         /* At this point we're granting the lock request. */
 587         req->l_granted_mode = req->l_req_mode;
 588
 589         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 590         if (!added) {
 591                 list_del_init(&req->l_res_link);
 592                 /* insert new lock before ownlocks in list. */
 593                 ldlm_resource_add_lock(res, ownlocks, req);
 594         }
 595
 596         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 597 #ifdef HAVE_SERVER_SUPPORT
 598                 if (intention == LDLM_PROCESS_ENQUEUE) {
 599                         /* If this is an unlock, reprocess the waitq and
 600                          * send completions ASTs for locks that can now be
 601                          * granted. The only problem with doing this
 602                          * reprocessing here is that the completion ASTs for
 603                          * newly granted locks will be sent before the unlock
 604                          * completion is sent. It shouldn't be an issue. Also
 605                          * note that ldlm_process_flock_lock() will recurse,
 606                          * but only once because 'intention' won't be
 607                          * LDLM_PROCESS_ENQUEUE from ldlm_reprocess_queue.
 608                          */
 609                         if ((mode == LCK_NL) && overlaps) {
 610                                 LIST_HEAD(rpc_list);
 611                                 int rc;
 612
 613 restart:
 614                                 ldlm_reprocess_queue(res, &res->lr_waiting,
 615                                                      &rpc_list,
 616                                                      LDLM_PROCESS_RESCAN, 0);
 617
 618                                 unlock_res_and_lock(req);
 619                                 rc = ldlm_run_ast_work(ns, &rpc_list,
 620                                                        LDLM_WORK_CP_AST);
 621                                 lock_res_and_lock(req);
 622                                 if (rc == -ERESTART)
 623                                         GOTO(restart, rc);
 624                         }
 625                 } else {
 626                         LASSERT(req->l_completion_ast);
 627                         ldlm_add_ast_work_item(req, NULL, grant_work);
 628                 }
 629 #else /* !HAVE_SERVER_SUPPORT */
 630                 /* The only one possible case for client-side calls flock
 631                  * policy function is ldlm_flock_completion_ast inside which
 632                  * carries LDLM_FL_WAIT_NOREPROC flag.
 633                  */
 634                 CERROR("Illegal parameter for client-side-only module.\n");
 635                 LBUG();
 636 #endif /* HAVE_SERVER_SUPPORT */
 637         }
 638
 639         /* In case we're reprocessing the requested lock we can't destroy
 640          * it until after calling ldlm_add_ast_work_item() above so that laawi()
 641          * can bump the reference count on \a req. Otherwise \a req
 642          * could be freed before the completion AST can be sent.
 643          */
 644         if (added)
 645                 ldlm_flock_destroy(req, mode, *flags);
 646
 647         ldlm_resource_dump(D_INFO, res);
 648         RETURN(LDLM_ITER_CONTINUE);
 649 }
 650
 651 /**
 652  * Flock completion callback function.
 653  *
 654  * \param lock [in,out]: A lock to be handled
 655  * \param flags    [in]: flags
 656  * \param *data    [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
 657  *
 658  * \retval 0    : success
 659  * \retval <0   : failure
 660  */
 661 int
 662 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 663 {
 664         struct file_lock *getlk = lock->l_ast_data;
 665         struct obd_device *obd;
 666         enum ldlm_error err;
 667         int rc = 0;
 668         ENTRY;
 669
 670         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4);
 671         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) {
 672                 lock_res_and_lock(lock);
 673                 lock->l_flags |= LDLM_FL_FAIL_LOC;
 674                 unlock_res_and_lock(lock);
 675                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT3, 4);
 676         }
 677         CDEBUG(D_DLMTRACE, "flags: %#llx data: %p getlk: %p\n",
 678                flags, data, getlk);
 679
 680         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 681
 682         if (flags & LDLM_FL_FAILED)
 683                 goto granted;
 684
 685         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
 686                 if (NULL == data)
 687                         /* mds granted the lock in the reply */
 688                         goto granted;
 689                 /* CP AST RPC: lock get granted, wake it up */
 690                 wake_up(&lock->l_waitq);
 691                 RETURN(0);
 692         }
 693
 694         LDLM_DEBUG(lock,
 695                    "client-side enqueue returned a blocked lock, sleeping");
 696         obd = class_exp2obd(lock->l_conn_export);
 697
 698         /* Go to sleep until the lock is granted. */
 699         rc = l_wait_event_abortable(lock->l_waitq,
 700                                     is_granted_or_cancelled(lock));
 701         if (rc < 0) {
 702                 /* take lock off the deadlock detection hash list. */
 703                 lock_res_and_lock(lock);
 704                 ldlm_flock_blocking_unlink(lock);
 705
 706                 /* client side - set flag to prevent lock from being
 707                  * put on LRU list
 708                  */
 709                 ldlm_set_cbpending(lock);
 710                 unlock_res_and_lock(lock);
 711
 712                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 713                            rc);
 714                 RETURN(rc);
 715         }
 716
 717 granted:
 718         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
 719
 720         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT4)) {
 721                 lock_res_and_lock(lock);
 722                 /* DEADLOCK is always set with CBPENDING */
 723                 lock->l_flags |= LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
 724                 unlock_res_and_lock(lock);
 725                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT4, 4);
 726         }
 727         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT5)) {
 728                 lock_res_and_lock(lock);
 729                 /* DEADLOCK is always set with CBPENDING */
 730                 lock->l_flags |= (LDLM_FL_FAIL_LOC |
 731                                   LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING);
 732                 unlock_res_and_lock(lock);
 733                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT5, 4);
 734         }
 735
 736         lock_res_and_lock(lock);
 737
 738
 739         /* Protect against race where lock could have been just destroyed
 740          * due to overlap in ldlm_process_flock_lock().
 741          */
 742         if (ldlm_is_destroyed(lock)) {
 743                 unlock_res_and_lock(lock);
 744                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 745
 746                 /* An error is still to be returned, to propagate it up to
 747                  * ldlm_cli_enqueue_fini() caller. */
 748                 RETURN(-EIO);
 749         }
 750
 751         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 752         ldlm_resource_unlink_lock(lock);
 753
 754         /* Import invalidation. We need to actually release the lock
 755          * references being held, so that it can go away. No point in
 756          * holding the lock even if app still believes it has it, since
 757          * server already dropped it anyway. Only for granted locks too.
 758          */
 759         /* Do the same for DEADLOCK'ed locks. */
 760         if (ldlm_is_failed(lock) || ldlm_is_flock_deadlock(lock)) {
 761                 int mode;
 762
 763                 if (flags & LDLM_FL_TEST_LOCK)
 764                         LASSERT(ldlm_is_test_lock(lock));
 765
 766                 if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock))
 767                         mode = getlk->fl_type;
 768                 else
 769                         mode = lock->l_req_mode;
 770
 771                 if (ldlm_is_flock_deadlock(lock)) {
 772                         LDLM_DEBUG(lock, "client-side enqueue deadlock "
 773                                    "received");
 774                         rc = -EDEADLK;
 775                 }
 776                 ldlm_flock_destroy(lock, mode, LDLM_FL_WAIT_NOREPROC);
 777                 unlock_res_and_lock(lock);
 778
 779                 /* Need to wake up the waiter if we were evicted */
 780                 wake_up(&lock->l_waitq);
 781
 782                 /* An error is still to be returned, to propagate it up to
 783                  * ldlm_cli_enqueue_fini() caller.
 784                  */
 785                 RETURN(rc ? : -EIO);
 786         }
 787
 788         LDLM_DEBUG(lock, "client-side enqueue granted");
 789
 790         if (flags & LDLM_FL_TEST_LOCK) {
 791                 /*
 792                  * fcntl(F_GETLK) request
 793                  * The old mode was saved in getlk->fl_type so that if the mode
 794                  * in the lock changes we can decref the appropriate refcount.
 795                  */
 796                 LASSERT(ldlm_is_test_lock(lock));
 797                 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
 798                 switch (lock->l_granted_mode) {
 799                 case LCK_PR:
 800                         getlk->fl_type = F_RDLCK;
 801                         break;
 802                 case LCK_PW:
 803                         getlk->fl_type = F_WRLCK;
 804                         break;
 805                 default:
 806                         getlk->fl_type = F_UNLCK;
 807                 }
 808                 getlk->fl_pid = (pid_t)lock->l_policy_data.l_flock.pid;
 809                 getlk->fl_start = (loff_t)lock->l_policy_data.l_flock.start;
 810                 getlk->fl_end = (loff_t)lock->l_policy_data.l_flock.end;
 811         } else {
 812                 __u64 noreproc = LDLM_FL_WAIT_NOREPROC;
 813
 814                 /* We need to reprocess the lock to do merges or splits
 815                  * with existing locks owned by this process.
 816                  */
 817                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 818         }
 819         unlock_res_and_lock(lock);
 820         RETURN(rc);
 821 }
 822 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 823
 824 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 825                             void *data, int flag)
 826 {
 827         ENTRY;
 828
 829         LASSERT(lock);
 830         LASSERT(flag == LDLM_CB_CANCELING);
 831
 832         /* take lock off the deadlock detection hash list. */
 833         lock_res_and_lock(lock);
 834         ldlm_flock_blocking_unlink(lock);
 835         unlock_res_and_lock(lock);
 836         RETURN(0);
 837 }
 838
 839 void ldlm_flock_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
 840                                      union ldlm_policy_data *lpolicy)
 841 {
 842         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
 843         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
 844         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
 845         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
 846 }
 847
 848 void ldlm_flock_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
 849                                      union ldlm_wire_policy_data *wpolicy)
 850 {
 851         memset(wpolicy, 0, sizeof(*wpolicy));
 852         wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
 853         wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
 854         wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
 855         wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
 856 }
 857
 858 /*
 859  * Export handle<->flock hash operations.
 860  */
 861 static unsigned
 862 ldlm_export_flock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
 863 {
 864         return cfs_hash_u64_hash(*(__u64 *)key, mask);
 865 }
 866
 867 static void *
 868 ldlm_export_flock_key(struct hlist_node *hnode)
 869 {
 870         struct ldlm_lock *lock;
 871
 872         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 873         return &lock->l_policy_data.l_flock.owner;
 874 }
 875
 876 static int
 877 ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode)
 878 {
 879         return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
 880 }
 881
 882 static void *
 883 ldlm_export_flock_object(struct hlist_node *hnode)
 884 {
 885         return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 886 }
 887
 888 static void
 889 ldlm_export_flock_get(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         LDLM_LOCK_GET(lock);
 896
 897         flock = &lock->l_policy_data.l_flock;
 898         LASSERT(flock->blocking_export != NULL);
 899         class_export_get(flock->blocking_export);
 900         atomic_inc(&flock->blocking_refs);
 901 }
 902
 903 static void
 904 ldlm_export_flock_put(struct cfs_hash *hs, struct hlist_node *hnode)
 905 {
 906         struct ldlm_lock *lock;
 907         struct ldlm_flock *flock;
 908
 909         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 910
 911         flock = &lock->l_policy_data.l_flock;
 912         LASSERT(flock->blocking_export != NULL);
 913         class_export_put(flock->blocking_export);
 914         if (atomic_dec_and_test(&flock->blocking_refs)) {
 915                 flock->blocking_owner = 0;
 916                 flock->blocking_export = NULL;
 917         }
 918         LDLM_LOCK_RELEASE(lock);
 919 }
 920
 921 static struct cfs_hash_ops ldlm_export_flock_ops = {
 922         .hs_hash        = ldlm_export_flock_hash,
 923         .hs_key         = ldlm_export_flock_key,
 924         .hs_keycmp      = ldlm_export_flock_keycmp,
 925         .hs_object      = ldlm_export_flock_object,
 926         .hs_get         = ldlm_export_flock_get,
 927         .hs_put         = ldlm_export_flock_put,
 928         .hs_put_locked  = ldlm_export_flock_put,
 929 };
 930
 931 int ldlm_init_flock_export(struct obd_export *exp)
 932 {
 933         if( strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0)
 934                 RETURN(0);
 935
 936         exp->exp_flock_hash =
 937                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
 938                                 HASH_EXP_LOCK_CUR_BITS,
 939                                 HASH_EXP_LOCK_MAX_BITS,
 940                                 HASH_EXP_LOCK_BKT_BITS, 0,
 941                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 942                                 &ldlm_export_flock_ops,
 943                                 CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE);
 944         if (!exp->exp_flock_hash)
 945                 RETURN(-ENOMEM);
 946
 947         RETURN(0);
 948 }
 949
 950 void ldlm_destroy_flock_export(struct obd_export *exp)
 951 {
 952         ENTRY;
 953         if (exp->exp_flock_hash) {
 954                 cfs_hash_putref(exp->exp_flock_hash);
 955                 exp->exp_flock_hash = NULL;
 956         }
 957         EXIT;
 958 }