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         if (exp->exp_failed)
 177                 return 0;
 178
 179         lock = cfs_hash_lookup(exp->exp_flock_hash, cb_data->bl_owner);
 180         if (lock == NULL)
 181                 return 0;
 182
 183         /* Stop on first found lock. Same process can't sleep twice */
 184         cb_data->lock = lock;
 185         cb_data->exp = class_export_get(exp);
 186
 187         return 1;
 188 }
 189
 190 static int
 191 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *bl_lock)
 192 {
 193         struct obd_export *req_exp = req->l_export;
 194         struct obd_export *bl_exp = bl_lock->l_export;
 195         __u64 req_owner = req->l_policy_data.l_flock.owner;
 196         __u64 bl_owner = bl_lock->l_policy_data.l_flock.owner;
 197
 198         /* For server only */
 199         if (req_exp == NULL)
 200                 return 0;
 201
 202         class_export_get(bl_exp);
 203         while (1) {
 204                 struct ldlm_flock_lookup_cb_data cb_data = {
 205                         .bl_owner = &bl_owner,
 206                         .lock = NULL,
 207                         .exp = NULL,
 208                 };
 209                 struct ptlrpc_connection *bl_exp_conn;
 210                 struct obd_export *bl_exp_new;
 211                 struct ldlm_lock *lock = NULL;
 212                 struct ldlm_flock *flock;
 213
 214                 bl_exp_conn = bl_exp->exp_connection;
 215                 if (bl_exp->exp_flock_hash != NULL) {
 216                         int found;
 217
 218                         found = obd_nid_export_for_each(bl_exp->exp_obd,
 219                                                         bl_exp_conn->c_peer.nid,
 220                                                         ldlm_flock_lookup_cb,
 221                                                         &cb_data);
 222                         if (found)
 223                                 lock = cb_data.lock;
 224                 }
 225                 if (lock == NULL)
 226                         break;
 227
 228                 class_export_put(bl_exp);
 229                 bl_exp = cb_data.exp;
 230
 231                 LASSERT(req != lock);
 232                 flock = &lock->l_policy_data.l_flock;
 233                 LASSERT(flock->owner == bl_owner);
 234                 bl_owner = flock->blocking_owner;
 235                 bl_exp_new = class_export_get(flock->blocking_export);
 236                 class_export_put(bl_exp);
 237
 238                 cfs_hash_put(bl_exp->exp_flock_hash, &lock->l_exp_flock_hash);
 239                 bl_exp = bl_exp_new;
 240
 241                 if (bl_exp->exp_failed)
 242                         break;
 243
 244                 if (bl_owner == req_owner &&
 245                     (bl_exp_conn->c_peer.nid ==
 246                      req_exp->exp_connection->c_peer.nid)) {
 247                         class_export_put(bl_exp);
 248                         return 1;
 249                 }
 250         }
 251         class_export_put(bl_exp);
 252
 253         return 0;
 254 }
 255
 256 static void ldlm_flock_cancel_on_deadlock(struct ldlm_lock *lock,
 257                                           struct list_head *work_list)
 258 {
 259         CDEBUG(D_INFO, "reprocess deadlock req=%p\n", lock);
 260
 261         if ((exp_connect_flags(lock->l_export) &
 262              OBD_CONNECT_FLOCK_DEAD) == 0) {
 263                 CERROR("deadlock found, but client doesn't support flock canceliation\n");
 264         } else {
 265                 LASSERT(lock->l_completion_ast);
 266                 LASSERT(!ldlm_is_ast_sent(lock));
 267                 lock->l_flags |= (LDLM_FL_AST_SENT | LDLM_FL_CANCEL_ON_BLOCK |
 268                                   LDLM_FL_FLOCK_DEADLOCK);
 269                 ldlm_flock_blocking_unlink(lock);
 270                 ldlm_resource_unlink_lock(lock);
 271                 ldlm_add_ast_work_item(lock, NULL, work_list);
 272         }
 273 }
 274 #endif /* HAVE_SERVER_SUPPORT */
 275
 276 /**
 277  * Process a granting attempt for flock lock.
 278  * Must be called under ns lock held.
 279  *
 280  * This function looks for any conflicts for \a lock in the granted or
 281  * waiting queues. The lock is granted if no conflicts are found in
 282  * either queue.
 283  */
 284 int
 285 ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags,
 286                         enum ldlm_process_intention intention,
 287                         enum ldlm_error *err, struct list_head *work_list)
 288 {
 289         struct ldlm_resource *res = req->l_resource;
 290         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
 291         struct list_head *tmp;
 292         struct list_head *ownlocks = NULL;
 293         struct ldlm_lock *lock = NULL;
 294         struct ldlm_lock *new = req;
 295         struct ldlm_lock *new2 = NULL;
 296         enum ldlm_mode mode = req->l_req_mode;
 297         int local = ns_is_client(ns);
 298         int added = (mode == LCK_NL);
 299         int overlaps = 0;
 300         int splitted = 0;
 301         const struct ldlm_callback_suite null_cbs = { NULL };
 302 #ifdef HAVE_SERVER_SUPPORT
 303         struct list_head *grant_work = (intention == LDLM_PROCESS_ENQUEUE ?
 304                                         NULL : work_list);
 305 #endif
 306         ENTRY;
 307
 308         CDEBUG(D_DLMTRACE, "flags %#llx owner %llu pid %u mode %u start "
 309                "%llu end %llu\n", *flags,
 310                new->l_policy_data.l_flock.owner,
 311                new->l_policy_data.l_flock.pid, mode,
 312                req->l_policy_data.l_flock.start,
 313                req->l_policy_data.l_flock.end);
 314
 315         *err = ELDLM_OK;
 316
 317         if (local) {
 318                 /* No blocking ASTs are sent to the clients for
 319                  * Posix file & record locks
 320                  */
 321                 req->l_blocking_ast = NULL;
 322         } else {
 323                 /* Called on the server for lock cancels. */
 324                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 325         }
 326
 327 reprocess:
 328         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 329                 /* This loop determines where this processes locks start
 330                  * in the resource lr_granted list.
 331                  */
 332                 list_for_each(tmp, &res->lr_granted) {
 333                         lock = list_entry(tmp, struct ldlm_lock,
 334                                           l_res_link);
 335                         if (ldlm_same_flock_owner(lock, req)) {
 336                                 ownlocks = tmp;
 337                                 break;
 338                         }
 339                 }
 340         }
 341 #ifdef HAVE_SERVER_SUPPORT
 342         else {
 343                 int reprocess_failed = 0;
 344                 lockmode_verify(mode);
 345
 346                 /* This loop determines if there are existing locks
 347                  * that conflict with the new lock request.
 348                  */
 349                 list_for_each(tmp, &res->lr_granted) {
 350                         lock = list_entry(tmp, struct ldlm_lock,
 351                                           l_res_link);
 352
 353                         if (ldlm_same_flock_owner(lock, req)) {
 354                                 if (!ownlocks)
 355                                         ownlocks = tmp;
 356                                 continue;
 357                         }
 358
 359                         /* locks are compatible, overlap doesn't matter */
 360                         if (lockmode_compat(lock->l_granted_mode, mode))
 361                                 continue;
 362
 363                         if (!ldlm_flocks_overlap(lock, req))
 364                                 continue;
 365
 366                         if (intention != LDLM_PROCESS_ENQUEUE) {
 367                                 reprocess_failed = 1;
 368                                 if (ldlm_flock_deadlock(req, lock)) {
 369                                         ldlm_flock_cancel_on_deadlock(
 370                                                 req, grant_work);
 371                                         RETURN(LDLM_ITER_CONTINUE);
 372                                 }
 373                                 continue;
 374                         }
 375
 376                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 377                                 ldlm_flock_destroy(req, mode, *flags);
 378                                 *err = -EAGAIN;
 379                                 RETURN(LDLM_ITER_STOP);
 380                         }
 381
 382                         if (*flags & LDLM_FL_TEST_LOCK) {
 383                                 ldlm_flock_destroy(req, mode, *flags);
 384                                 req->l_req_mode = lock->l_granted_mode;
 385                                 req->l_policy_data.l_flock.pid =
 386                                         lock->l_policy_data.l_flock.pid;
 387                                 req->l_policy_data.l_flock.start =
 388                                         lock->l_policy_data.l_flock.start;
 389                                 req->l_policy_data.l_flock.end =
 390                                         lock->l_policy_data.l_flock.end;
 391                                 *flags |= LDLM_FL_LOCK_CHANGED;
 392                                 RETURN(LDLM_ITER_STOP);
 393                         }
 394
 395                         /* add lock to blocking list before deadlock
 396                          * check to prevent race
 397                          */
 398                         ldlm_flock_blocking_link(req, lock);
 399
 400                         if (ldlm_flock_deadlock(req, lock)) {
 401                                 ldlm_flock_blocking_unlink(req);
 402                                 ldlm_flock_destroy(req, mode, *flags);
 403                                 *err = -EDEADLK;
 404                                 RETURN(LDLM_ITER_STOP);
 405                         }
 406
 407                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 408                         *flags |= LDLM_FL_BLOCK_GRANTED;
 409                         RETURN(LDLM_ITER_STOP);
 410                 }
 411                 if (reprocess_failed)
 412                         RETURN(LDLM_ITER_CONTINUE);
 413         }
 414
 415         if (*flags & LDLM_FL_TEST_LOCK) {
 416                 ldlm_flock_destroy(req, mode, *flags);
 417                 req->l_req_mode = LCK_NL;
 418                 *flags |= LDLM_FL_LOCK_CHANGED;
 419                 RETURN(LDLM_ITER_STOP);
 420         }
 421
 422         /* In case we had slept on this lock request take it off of the
 423          * deadlock detection hash list.
 424          */
 425         ldlm_flock_blocking_unlink(req);
 426 #endif /* HAVE_SERVER_SUPPORT */
 427
 428         /* Scan the locks owned by this process that overlap this request.
 429          * We may have to merge or split existing locks.
 430          */
 431         if (!ownlocks)
 432                 ownlocks = &res->lr_granted;
 433
 434         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 435                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 436
 437                 if (!ldlm_same_flock_owner(lock, new))
 438                         break;
 439
 440                 if (lock->l_granted_mode == mode) {
 441                         /* If the modes are the same then we need to process
 442                          * locks that overlap OR adjoin the new lock. The extra
 443                          * logic condition is necessary to deal with arithmetic
 444                          * overflow and underflow.
 445                          */
 446                         if ((new->l_policy_data.l_flock.start >
 447                              (lock->l_policy_data.l_flock.end + 1))
 448                             && (lock->l_policy_data.l_flock.end !=
 449                                 OBD_OBJECT_EOF))
 450                                 continue;
 451
 452                         if ((new->l_policy_data.l_flock.end <
 453                              (lock->l_policy_data.l_flock.start - 1))
 454                             && (lock->l_policy_data.l_flock.start != 0))
 455                                 break;
 456
 457                         if (new->l_policy_data.l_flock.start <
 458                             lock->l_policy_data.l_flock.start) {
 459                                 lock->l_policy_data.l_flock.start =
 460                                         new->l_policy_data.l_flock.start;
 461                         } else {
 462                                 new->l_policy_data.l_flock.start =
 463                                         lock->l_policy_data.l_flock.start;
 464                         }
 465
 466                         if (new->l_policy_data.l_flock.end >
 467                             lock->l_policy_data.l_flock.end) {
 468                                 lock->l_policy_data.l_flock.end =
 469                                         new->l_policy_data.l_flock.end;
 470                         } else {
 471                                 new->l_policy_data.l_flock.end =
 472                                         lock->l_policy_data.l_flock.end;
 473                         }
 474
 475                         if (added) {
 476                                 ldlm_flock_destroy(lock, mode, *flags);
 477                         } else {
 478                                 new = lock;
 479                                 added = 1;
 480                         }
 481                         continue;
 482                 }
 483
 484                 if (new->l_policy_data.l_flock.start >
 485                     lock->l_policy_data.l_flock.end)
 486                         continue;
 487
 488                 if (new->l_policy_data.l_flock.end <
 489                     lock->l_policy_data.l_flock.start)
 490                         break;
 491
 492                 ++overlaps;
 493
 494                 if (new->l_policy_data.l_flock.start <=
 495                     lock->l_policy_data.l_flock.start) {
 496                         if (new->l_policy_data.l_flock.end <
 497                             lock->l_policy_data.l_flock.end) {
 498                                 lock->l_policy_data.l_flock.start =
 499                                         new->l_policy_data.l_flock.end + 1;
 500                                 break;
 501                         }
 502                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 503                         continue;
 504                 }
 505                 if (new->l_policy_data.l_flock.end >=
 506                     lock->l_policy_data.l_flock.end) {
 507                         lock->l_policy_data.l_flock.end =
 508                                 new->l_policy_data.l_flock.start - 1;
 509                         continue;
 510                 }
 511
 512                 /* split the existing lock into two locks */
 513
 514                 /* if this is an F_UNLCK operation then we could avoid
 515                  * allocating a new lock and use the req lock passed in
 516                  * with the request but this would complicate the reply
 517                  * processing since updates to req get reflected in the
 518                  * reply. The client side replays the lock request so
 519                  * it must see the original lock data in the reply.
 520                  */
 521
 522                 /* XXX - if ldlm_lock_new() can sleep we should
 523                  * release the lr_lock, allocate the new lock,
 524                  * and restart processing this lock.
 525                  */
 526                 if (new2 == NULL) {
 527                         unlock_res_and_lock(req);
 528                         new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
 529                                                 lock->l_granted_mode, &null_cbs,
 530                                                 NULL, 0, LVB_T_NONE);
 531                         lock_res_and_lock(req);
 532                         if (IS_ERR(new2)) {
 533                                 ldlm_flock_destroy(req, lock->l_granted_mode,
 534                                                    *flags);
 535                                 *err = PTR_ERR(new2);
 536                                 RETURN(LDLM_ITER_STOP);
 537                         }
 538                         goto reprocess;
 539                 }
 540
 541                 splitted = 1;
 542
 543                 new2->l_granted_mode = lock->l_granted_mode;
 544                 new2->l_policy_data.l_flock.pid =
 545                         new->l_policy_data.l_flock.pid;
 546                 new2->l_policy_data.l_flock.owner =
 547                         new->l_policy_data.l_flock.owner;
 548                 new2->l_policy_data.l_flock.start =
 549                         lock->l_policy_data.l_flock.start;
 550                 new2->l_policy_data.l_flock.end =
 551                         new->l_policy_data.l_flock.start - 1;
 552                 lock->l_policy_data.l_flock.start =
 553                         new->l_policy_data.l_flock.end + 1;
 554                 new2->l_conn_export = lock->l_conn_export;
 555                 if (lock->l_export != NULL) {
 556                         new2->l_export = class_export_lock_get(lock->l_export,
 557                                                                new2);
 558                         if (new2->l_export->exp_lock_hash &&
 559                             hlist_unhashed(&new2->l_exp_hash))
 560                                 cfs_hash_add(new2->l_export->exp_lock_hash,
 561                                              &new2->l_remote_handle,
 562                                              &new2->l_exp_hash);
 563                 }
 564                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 565                         ldlm_lock_addref_internal_nolock(new2,
 566                                                          lock->l_granted_mode);
 567
 568                 /* insert new2 at lock */
 569                 ldlm_resource_add_lock(res, ownlocks, new2);
 570                 LDLM_LOCK_RELEASE(new2);
 571                 break;
 572         }
 573
 574         /* if new2 is created but never used, destroy it*/
 575         if (splitted == 0 && new2 != NULL)
 576                 ldlm_lock_destroy_nolock(new2);
 577
 578         /* At this point we're granting the lock request. */
 579         req->l_granted_mode = req->l_req_mode;
 580
 581         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 582         if (!added) {
 583                 list_del_init(&req->l_res_link);
 584                 /* insert new lock before ownlocks in list. */
 585                 ldlm_resource_add_lock(res, ownlocks, req);
 586         }
 587
 588         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 589 #ifdef HAVE_SERVER_SUPPORT
 590                 if (intention == LDLM_PROCESS_ENQUEUE) {
 591                         /* If this is an unlock, reprocess the waitq and
 592                          * send completions ASTs for locks that can now be
 593                          * granted. The only problem with doing this
 594                          * reprocessing here is that the completion ASTs for
 595                          * newly granted locks will be sent before the unlock
 596                          * completion is sent. It shouldn't be an issue. Also
 597                          * note that ldlm_process_flock_lock() will recurse,
 598                          * but only once because 'intention' won't be
 599                          * LDLM_PROCESS_ENQUEUE from ldlm_reprocess_queue.
 600                          */
 601                         if ((mode == LCK_NL) && overlaps) {
 602                                 LIST_HEAD(rpc_list);
 603                                 int rc;
 604
 605 restart:
 606                                 ldlm_reprocess_queue(res, &res->lr_waiting,
 607                                                      &rpc_list,
 608                                                      LDLM_PROCESS_RESCAN, NULL);
 609
 610                                 unlock_res_and_lock(req);
 611                                 rc = ldlm_run_ast_work(ns, &rpc_list,
 612                                                        LDLM_WORK_CP_AST);
 613                                 lock_res_and_lock(req);
 614                                 if (rc == -ERESTART)
 615                                         GOTO(restart, rc);
 616                         }
 617                 } else {
 618                         LASSERT(req->l_completion_ast);
 619                         ldlm_add_ast_work_item(req, NULL, grant_work);
 620                 }
 621 #else /* !HAVE_SERVER_SUPPORT */
 622                 /* The only one possible case for client-side calls flock
 623                  * policy function is ldlm_flock_completion_ast inside which
 624                  * carries LDLM_FL_WAIT_NOREPROC flag.
 625                  */
 626                 CERROR("Illegal parameter for client-side-only module.\n");
 627                 LBUG();
 628 #endif /* HAVE_SERVER_SUPPORT */
 629         }
 630
 631         /* In case we're reprocessing the requested lock we can't destroy
 632          * it until after calling ldlm_add_ast_work_item() above so that laawi()
 633          * can bump the reference count on \a req. Otherwise \a req
 634          * could be freed before the completion AST can be sent.
 635          */
 636         if (added)
 637                 ldlm_flock_destroy(req, mode, *flags);
 638
 639         ldlm_resource_dump(D_INFO, res);
 640         RETURN(LDLM_ITER_CONTINUE);
 641 }
 642
 643 /**
 644  * Flock completion callback function.
 645  *
 646  * \param lock [in,out]: A lock to be handled
 647  * \param flags    [in]: flags
 648  * \param *data    [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
 649  *
 650  * \retval 0    : success
 651  * \retval <0   : failure
 652  */
 653 int
 654 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 655 {
 656         struct file_lock *getlk = lock->l_ast_data;
 657         struct obd_device *obd;
 658         struct obd_import *imp = NULL;
 659         enum ldlm_error err;
 660         int rc = 0;
 661         ENTRY;
 662
 663         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4);
 664         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) {
 665                 lock_res_and_lock(lock);
 666                 lock->l_flags |= LDLM_FL_FAIL_LOC;
 667                 unlock_res_and_lock(lock);
 668                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT3, 4);
 669         }
 670         CDEBUG(D_DLMTRACE, "flags: %#llx data: %p getlk: %p\n",
 671                flags, data, getlk);
 672
 673         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 674
 675         if (flags & LDLM_FL_FAILED)
 676                 goto granted;
 677
 678         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
 679                 if (NULL == data)
 680                         /* mds granted the lock in the reply */
 681                         goto granted;
 682                 /* CP AST RPC: lock get granted, wake it up */
 683                 wake_up(&lock->l_waitq);
 684                 RETURN(0);
 685         }
 686
 687         LDLM_DEBUG(lock,
 688                    "client-side enqueue returned a blocked lock, sleeping");
 689         obd = class_exp2obd(lock->l_conn_export);
 690
 691         /* if this is a local lock, there is no import */
 692         if (obd)
 693                 imp = obd->u.cli.cl_import;
 694
 695         /* Go to sleep until the lock is granted. */
 696         rc = l_wait_event_abortable(lock->l_waitq,
 697                                     is_granted_or_cancelled(lock));
 698         if (rc < 0) {
 699                 /* take lock off the deadlock detection hash list. */
 700                 lock_res_and_lock(lock);
 701                 ldlm_flock_blocking_unlink(lock);
 702
 703                 /* client side - set flag to prevent lock from being
 704                  * put on LRU list
 705                  */
 706                 ldlm_set_cbpending(lock);
 707                 unlock_res_and_lock(lock);
 708
 709                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 710                            rc);
 711                 RETURN(rc);
 712         }
 713
 714 granted:
 715         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
 716
 717         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT4)) {
 718                 lock_res_and_lock(lock);
 719                 /* DEADLOCK is always set with CBPENDING */
 720                 lock->l_flags |= LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
 721                 unlock_res_and_lock(lock);
 722                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT4, 4);
 723         }
 724         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT5)) {
 725                 lock_res_and_lock(lock);
 726                 /* DEADLOCK is always set with CBPENDING */
 727                 lock->l_flags |= (LDLM_FL_FAIL_LOC |
 728                                   LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING);
 729                 unlock_res_and_lock(lock);
 730                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT5, 4);
 731         }
 732
 733         lock_res_and_lock(lock);
 734
 735
 736         /* Protect against race where lock could have been just destroyed
 737          * due to overlap in ldlm_process_flock_lock().
 738          */
 739         if (ldlm_is_destroyed(lock)) {
 740                 unlock_res_and_lock(lock);
 741                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 742
 743                 /* An error is still to be returned, to propagate it up to
 744                  * ldlm_cli_enqueue_fini() caller. */
 745                 RETURN(-EIO);
 746         }
 747
 748         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 749         ldlm_resource_unlink_lock(lock);
 750
 751         /* Import invalidation. We need to actually release the lock
 752          * references being held, so that it can go away. No point in
 753          * holding the lock even if app still believes it has it, since
 754          * server already dropped it anyway. Only for granted locks too.
 755          */
 756         /* Do the same for DEADLOCK'ed locks. */
 757         if (ldlm_is_failed(lock) || ldlm_is_flock_deadlock(lock)) {
 758                 int mode;
 759
 760                 if (flags & LDLM_FL_TEST_LOCK)
 761                         LASSERT(ldlm_is_test_lock(lock));
 762
 763                 if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock))
 764                         mode = getlk->fl_type;
 765                 else
 766                         mode = lock->l_req_mode;
 767
 768                 if (ldlm_is_flock_deadlock(lock)) {
 769                         LDLM_DEBUG(lock, "client-side enqueue deadlock "
 770                                    "received");
 771                         rc = -EDEADLK;
 772                 }
 773                 ldlm_flock_destroy(lock, mode, LDLM_FL_WAIT_NOREPROC);
 774                 unlock_res_and_lock(lock);
 775
 776                 /* Need to wake up the waiter if we were evicted */
 777                 wake_up(&lock->l_waitq);
 778
 779                 /* An error is still to be returned, to propagate it up to
 780                  * ldlm_cli_enqueue_fini() caller.
 781                  */
 782                 RETURN(rc ? : -EIO);
 783         }
 784
 785         LDLM_DEBUG(lock, "client-side enqueue granted");
 786
 787         if (flags & LDLM_FL_TEST_LOCK) {
 788                 /*
 789                  * fcntl(F_GETLK) request
 790                  * The old mode was saved in getlk->fl_type so that if the mode
 791                  * in the lock changes we can decref the appropriate refcount.
 792                  */
 793                 LASSERT(ldlm_is_test_lock(lock));
 794                 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
 795                 switch (lock->l_granted_mode) {
 796                 case LCK_PR:
 797                         getlk->fl_type = F_RDLCK;
 798                         break;
 799                 case LCK_PW:
 800                         getlk->fl_type = F_WRLCK;
 801                         break;
 802                 default:
 803                         getlk->fl_type = F_UNLCK;
 804                 }
 805                 getlk->fl_pid = (pid_t)lock->l_policy_data.l_flock.pid;
 806                 getlk->fl_start = (loff_t)lock->l_policy_data.l_flock.start;
 807                 getlk->fl_end = (loff_t)lock->l_policy_data.l_flock.end;
 808         } else {
 809                 __u64 noreproc = LDLM_FL_WAIT_NOREPROC;
 810
 811                 /* We need to reprocess the lock to do merges or splits
 812                  * with existing locks owned by this process.
 813                  */
 814                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 815         }
 816         unlock_res_and_lock(lock);
 817         RETURN(rc);
 818 }
 819 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 820
 821 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 822                             void *data, int flag)
 823 {
 824         ENTRY;
 825
 826         LASSERT(lock);
 827         LASSERT(flag == LDLM_CB_CANCELING);
 828
 829         /* take lock off the deadlock detection hash list. */
 830         lock_res_and_lock(lock);
 831         ldlm_flock_blocking_unlink(lock);
 832         unlock_res_and_lock(lock);
 833         RETURN(0);
 834 }
 835
 836 void ldlm_flock_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
 837                                      union ldlm_policy_data *lpolicy)
 838 {
 839         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
 840         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
 841         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
 842         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
 843 }
 844
 845 void ldlm_flock_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
 846                                      union ldlm_wire_policy_data *wpolicy)
 847 {
 848         memset(wpolicy, 0, sizeof(*wpolicy));
 849         wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
 850         wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
 851         wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
 852         wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
 853 }
 854
 855 /*
 856  * Export handle<->flock hash operations.
 857  */
 858 static unsigned
 859 ldlm_export_flock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
 860 {
 861         return cfs_hash_u64_hash(*(__u64 *)key, mask);
 862 }
 863
 864 static void *
 865 ldlm_export_flock_key(struct hlist_node *hnode)
 866 {
 867         struct ldlm_lock *lock;
 868
 869         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 870         return &lock->l_policy_data.l_flock.owner;
 871 }
 872
 873 static int
 874 ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode)
 875 {
 876         return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
 877 }
 878
 879 static void *
 880 ldlm_export_flock_object(struct hlist_node *hnode)
 881 {
 882         return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 883 }
 884
 885 static void
 886 ldlm_export_flock_get(struct cfs_hash *hs, struct hlist_node *hnode)
 887 {
 888         struct ldlm_lock *lock;
 889         struct ldlm_flock *flock;
 890
 891         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 892         LDLM_LOCK_GET(lock);
 893
 894         flock = &lock->l_policy_data.l_flock;
 895         LASSERT(flock->blocking_export != NULL);
 896         class_export_get(flock->blocking_export);
 897         atomic_inc(&flock->blocking_refs);
 898 }
 899
 900 static void
 901 ldlm_export_flock_put(struct cfs_hash *hs, struct hlist_node *hnode)
 902 {
 903         struct ldlm_lock *lock;
 904         struct ldlm_flock *flock;
 905
 906         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 907
 908         flock = &lock->l_policy_data.l_flock;
 909         LASSERT(flock->blocking_export != NULL);
 910         class_export_put(flock->blocking_export);
 911         if (atomic_dec_and_test(&flock->blocking_refs)) {
 912                 flock->blocking_owner = 0;
 913                 flock->blocking_export = NULL;
 914         }
 915         LDLM_LOCK_RELEASE(lock);
 916 }
 917
 918 static struct cfs_hash_ops ldlm_export_flock_ops = {
 919         .hs_hash        = ldlm_export_flock_hash,
 920         .hs_key         = ldlm_export_flock_key,
 921         .hs_keycmp      = ldlm_export_flock_keycmp,
 922         .hs_object      = ldlm_export_flock_object,
 923         .hs_get         = ldlm_export_flock_get,
 924         .hs_put         = ldlm_export_flock_put,
 925         .hs_put_locked  = ldlm_export_flock_put,
 926 };
 927
 928 int ldlm_init_flock_export(struct obd_export *exp)
 929 {
 930         if( strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0)
 931                 RETURN(0);
 932
 933         exp->exp_flock_hash =
 934                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
 935                                 HASH_EXP_LOCK_CUR_BITS,
 936                                 HASH_EXP_LOCK_MAX_BITS,
 937                                 HASH_EXP_LOCK_BKT_BITS, 0,
 938                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 939                                 &ldlm_export_flock_ops,
 940                                 CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE);
 941         if (!exp->exp_flock_hash)
 942                 RETURN(-ENOMEM);
 943
 944         RETURN(0);
 945 }
 946
 947 void ldlm_destroy_flock_export(struct obd_export *exp)
 948 {
 949         ENTRY;
 950         if (exp->exp_flock_hash) {
 951                 cfs_hash_putref(exp->exp_flock_hash);
 952                 exp->exp_flock_hash = NULL;
 953         }
 954         EXIT;
 955 }