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 splitted = 0;
 299         const struct ldlm_callback_suite null_cbs = { NULL };
 300 #ifdef HAVE_SERVER_SUPPORT
 301         struct list_head *grant_work = (intention == LDLM_PROCESS_ENQUEUE ?
 302                                         NULL : work_list);
 303 #endif
 304         ENTRY;
 305
 306         CDEBUG(D_DLMTRACE, "flags %#llx owner %llu pid %u mode %u start "
 307                "%llu end %llu\n", *flags,
 308                new->l_policy_data.l_flock.owner,
 309                new->l_policy_data.l_flock.pid, mode,
 310                req->l_policy_data.l_flock.start,
 311                req->l_policy_data.l_flock.end);
 312
 313         *err = ELDLM_OK;
 314
 315         if (local) {
 316                 /* No blocking ASTs are sent to the clients for
 317                  * Posix file & record locks
 318                  */
 319                 req->l_blocking_ast = NULL;
 320         } else {
 321                 /* Called on the server for lock cancels. */
 322                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 323         }
 324
 325 reprocess:
 326         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 327                 /* This loop determines where this processes locks start
 328                  * in the resource lr_granted list.
 329                  */
 330                 list_for_each(tmp, &res->lr_granted) {
 331                         lock = list_entry(tmp, struct ldlm_lock,
 332                                           l_res_link);
 333                         if (ldlm_same_flock_owner(lock, req)) {
 334                                 ownlocks = tmp;
 335                                 break;
 336                         }
 337                 }
 338         }
 339 #ifdef HAVE_SERVER_SUPPORT
 340         else {
 341                 int reprocess_failed = 0;
 342                 lockmode_verify(mode);
 343
 344                 /* This loop determines if there are existing locks
 345                  * that conflict with the new lock request.
 346                  */
 347                 list_for_each(tmp, &res->lr_granted) {
 348                         lock = list_entry(tmp, struct ldlm_lock,
 349                                           l_res_link);
 350
 351                         if (ldlm_same_flock_owner(lock, req)) {
 352                                 if (!ownlocks)
 353                                         ownlocks = tmp;
 354                                 continue;
 355                         }
 356
 357                         if (req->l_req_mode == LCK_PR &&
 358                             lock->l_granted_mode == LCK_PR &&
 359                             lock->l_policy_data.l_flock.start <=
 360                                 req->l_policy_data.l_flock.start &&
 361                             lock->l_policy_data.l_flock.end >=
 362                                 req->l_policy_data.l_flock.end) {
 363                                 /* there can't be granted WR lock */
 364                                 break;
 365                         }
 366                         /* locks are compatible, overlap doesn't matter */
 367                         if (lockmode_compat(lock->l_granted_mode, mode))
 368                                 continue;
 369
 370                         if (!ldlm_flocks_overlap(lock, req))
 371                                 continue;
 372
 373                         if (intention != LDLM_PROCESS_ENQUEUE) {
 374                                 if (ldlm_flock_deadlock(req, lock)) {
 375                                         ldlm_flock_cancel_on_deadlock(
 376                                                 req, grant_work);
 377                                         RETURN(LDLM_ITER_CONTINUE);
 378                                 }
 379                                 reprocess_failed = 1;
 380                                 break;
 381                         }
 382
 383                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 384                                 ldlm_flock_destroy(req, mode, *flags);
 385                                 *err = -EAGAIN;
 386                                 RETURN(LDLM_ITER_STOP);
 387                         }
 388
 389                         if (*flags & LDLM_FL_TEST_LOCK) {
 390                                 ldlm_flock_destroy(req, mode, *flags);
 391                                 req->l_req_mode = lock->l_granted_mode;
 392                                 req->l_policy_data.l_flock.pid =
 393                                         lock->l_policy_data.l_flock.pid;
 394                                 req->l_policy_data.l_flock.start =
 395                                         lock->l_policy_data.l_flock.start;
 396                                 req->l_policy_data.l_flock.end =
 397                                         lock->l_policy_data.l_flock.end;
 398                                 *flags |= LDLM_FL_LOCK_CHANGED;
 399                                 RETURN(LDLM_ITER_STOP);
 400                         }
 401
 402                         /* add lock to blocking list before deadlock
 403                          * check to prevent race
 404                          */
 405                         ldlm_flock_blocking_link(req, lock);
 406
 407                         if (ldlm_flock_deadlock(req, lock)) {
 408                                 ldlm_flock_blocking_unlink(req);
 409                                 ldlm_flock_destroy(req, mode, *flags);
 410                                 *err = -EDEADLK;
 411                                 RETURN(LDLM_ITER_STOP);
 412                         }
 413
 414                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 415                         *flags |= LDLM_FL_BLOCK_GRANTED;
 416                         RETURN(LDLM_ITER_STOP);
 417                 }
 418                 if (reprocess_failed)
 419                         RETURN(LDLM_ITER_CONTINUE);
 420         }
 421
 422         if (*flags & LDLM_FL_TEST_LOCK) {
 423                 ldlm_flock_destroy(req, mode, *flags);
 424                 req->l_req_mode = LCK_NL;
 425                 *flags |= LDLM_FL_LOCK_CHANGED;
 426                 RETURN(LDLM_ITER_STOP);
 427         }
 428
 429         /* In case we had slept on this lock request take it off of the
 430          * deadlock detection hash list.
 431          */
 432         ldlm_flock_blocking_unlink(req);
 433 #endif /* HAVE_SERVER_SUPPORT */
 434
 435         /* Scan the locks owned by this process that overlap this request.
 436          * We may have to merge or split existing locks.
 437          */
 438         if (!ownlocks)
 439                 ownlocks = &res->lr_granted;
 440
 441         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 442                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 443
 444                 if (!ldlm_same_flock_owner(lock, new))
 445                         break;
 446
 447                 if (lock->l_granted_mode == mode) {
 448                         /* If the modes are the same then we need to process
 449                          * locks that overlap OR adjoin the new lock. The extra
 450                          * logic condition is necessary to deal with arithmetic
 451                          * overflow and underflow.
 452                          */
 453                         if ((new->l_policy_data.l_flock.start >
 454                              (lock->l_policy_data.l_flock.end + 1))
 455                             && (lock->l_policy_data.l_flock.end !=
 456                                 OBD_OBJECT_EOF))
 457                                 continue;
 458
 459                         if ((new->l_policy_data.l_flock.end <
 460                              (lock->l_policy_data.l_flock.start - 1))
 461                             && (lock->l_policy_data.l_flock.start != 0))
 462                                 break;
 463
 464                         if (new->l_policy_data.l_flock.start <
 465                             lock->l_policy_data.l_flock.start) {
 466                                 lock->l_policy_data.l_flock.start =
 467                                         new->l_policy_data.l_flock.start;
 468                         } else {
 469                                 new->l_policy_data.l_flock.start =
 470                                         lock->l_policy_data.l_flock.start;
 471                         }
 472
 473                         if (new->l_policy_data.l_flock.end >
 474                             lock->l_policy_data.l_flock.end) {
 475                                 lock->l_policy_data.l_flock.end =
 476                                         new->l_policy_data.l_flock.end;
 477                         } else {
 478                                 new->l_policy_data.l_flock.end =
 479                                         lock->l_policy_data.l_flock.end;
 480                         }
 481
 482                         if (added) {
 483                                 ldlm_flock_destroy(lock, mode, *flags);
 484                         } else {
 485                                 new = lock;
 486                                 added = 1;
 487                         }
 488                         continue;
 489                 }
 490
 491                 if (new->l_policy_data.l_flock.start >
 492                     lock->l_policy_data.l_flock.end)
 493                         continue;
 494
 495                 if (new->l_policy_data.l_flock.end <
 496                     lock->l_policy_data.l_flock.start)
 497                         break;
 498
 499                 res->lr_flock_node.lfn_needs_reprocess = true;
 500
 501                 if (new->l_policy_data.l_flock.start <=
 502                     lock->l_policy_data.l_flock.start) {
 503                         if (new->l_policy_data.l_flock.end <
 504                             lock->l_policy_data.l_flock.end) {
 505                                 lock->l_policy_data.l_flock.start =
 506                                         new->l_policy_data.l_flock.end + 1;
 507                                 break;
 508                         }
 509                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 510                         continue;
 511                 }
 512                 if (new->l_policy_data.l_flock.end >=
 513                     lock->l_policy_data.l_flock.end) {
 514                         lock->l_policy_data.l_flock.end =
 515                                 new->l_policy_data.l_flock.start - 1;
 516                         continue;
 517                 }
 518
 519                 /* split the existing lock into two locks */
 520
 521                 /* if this is an F_UNLCK operation then we could avoid
 522                  * allocating a new lock and use the req lock passed in
 523                  * with the request but this would complicate the reply
 524                  * processing since updates to req get reflected in the
 525                  * reply. The client side replays the lock request so
 526                  * it must see the original lock data in the reply.
 527                  */
 528
 529                 /* XXX - if ldlm_lock_new() can sleep we should
 530                  * release the lr_lock, allocate the new lock,
 531                  * and restart processing this lock.
 532                  */
 533                 if (new2 == NULL) {
 534                         unlock_res_and_lock(req);
 535                         new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
 536                                                 lock->l_granted_mode, &null_cbs,
 537                                                 NULL, 0, LVB_T_NONE);
 538                         lock_res_and_lock(req);
 539                         if (IS_ERR(new2)) {
 540                                 ldlm_flock_destroy(req, lock->l_granted_mode,
 541                                                    *flags);
 542                                 *err = PTR_ERR(new2);
 543                                 RETURN(LDLM_ITER_STOP);
 544                         }
 545                         goto reprocess;
 546                 }
 547
 548                 splitted = 1;
 549
 550                 new2->l_granted_mode = lock->l_granted_mode;
 551                 new2->l_policy_data.l_flock.pid =
 552                         new->l_policy_data.l_flock.pid;
 553                 new2->l_policy_data.l_flock.owner =
 554                         new->l_policy_data.l_flock.owner;
 555                 new2->l_policy_data.l_flock.start =
 556                         lock->l_policy_data.l_flock.start;
 557                 new2->l_policy_data.l_flock.end =
 558                         new->l_policy_data.l_flock.start - 1;
 559                 lock->l_policy_data.l_flock.start =
 560                         new->l_policy_data.l_flock.end + 1;
 561                 new2->l_conn_export = lock->l_conn_export;
 562                 if (lock->l_export != NULL) {
 563                         new2->l_export = class_export_lock_get(lock->l_export,
 564                                                                new2);
 565                         if (new2->l_export->exp_lock_hash &&
 566                             hlist_unhashed(&new2->l_exp_hash))
 567                                 cfs_hash_add(new2->l_export->exp_lock_hash,
 568                                              &new2->l_remote_handle,
 569                                              &new2->l_exp_hash);
 570                 }
 571                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 572                         ldlm_lock_addref_internal_nolock(new2,
 573                                                          lock->l_granted_mode);
 574
 575                 /* insert new2 at lock */
 576                 ldlm_resource_add_lock(res, ownlocks, new2);
 577                 LDLM_LOCK_RELEASE(new2);
 578                 break;
 579         }
 580
 581         /* if new2 is created but never used, destroy it*/
 582         if (splitted == 0 && new2 != NULL)
 583                 ldlm_lock_destroy_nolock(new2);
 584
 585         /* At this point we're granting the lock request. */
 586         req->l_granted_mode = req->l_req_mode;
 587
 588         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 589         if (!added) {
 590                 list_del_init(&req->l_res_link);
 591                 /* insert new lock before ownlocks in list. */
 592                 ldlm_resource_add_lock(res, ownlocks, req);
 593         }
 594
 595         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 596 #ifdef HAVE_SERVER_SUPPORT
 597                 if (intention == LDLM_PROCESS_ENQUEUE) {
 598                         /* If this is an unlock, reprocess the waitq and
 599                          * send completions ASTs for locks that can now be
 600                          * granted. The only problem with doing this
 601                          * reprocessing here is that the completion ASTs for
 602                          * newly granted locks will be sent before the unlock
 603                          * completion is sent. It shouldn't be an issue. Also
 604                          * note that ldlm_process_flock_lock() will recurse,
 605                          * but only once because 'intention' won't be
 606                          * LDLM_PROCESS_ENQUEUE from ldlm_reprocess_queue.
 607                          */
 608                         struct ldlm_flock_node *fn = &res->lr_flock_node;
 609 restart:
 610                         if (mode == LCK_NL && fn->lfn_needs_reprocess &&
 611                             atomic_read(&fn->lfn_unlock_pending) == 0) {
 612                                 LIST_HEAD(rpc_list);
 613                                 int rc;
 614
 615                                 ldlm_reprocess_queue(res, &res->lr_waiting,
 616                                                      &rpc_list,
 617                                                      LDLM_PROCESS_RESCAN, 0);
 618                                 fn->lfn_needs_reprocess = false;
 619                                 unlock_res_and_lock(req);
 620                                 rc = ldlm_run_ast_work(ns, &rpc_list,
 621                                                        LDLM_WORK_CP_AST);
 622                                 lock_res_and_lock(req);
 623                                 if (rc == -ERESTART) {
 624                                         fn->lfn_needs_reprocess = true;
 625                                         GOTO(restart, rc);
 626                                 }
 627                         }
 628                 } else {
 629                         LASSERT(req->l_completion_ast);
 630                         ldlm_add_ast_work_item(req, NULL, grant_work);
 631                 }
 632 #else /* !HAVE_SERVER_SUPPORT */
 633                 /* The only one possible case for client-side calls flock
 634                  * policy function is ldlm_flock_completion_ast inside which
 635                  * carries LDLM_FL_WAIT_NOREPROC flag.
 636                  */
 637                 CERROR("Illegal parameter for client-side-only module.\n");
 638                 LBUG();
 639 #endif /* HAVE_SERVER_SUPPORT */
 640         }
 641
 642         /* In case we're reprocessing the requested lock we can't destroy
 643          * it until after calling ldlm_add_ast_work_item() above so that laawi()
 644          * can bump the reference count on \a req. Otherwise \a req
 645          * could be freed before the completion AST can be sent.
 646          */
 647         if (added)
 648                 ldlm_flock_destroy(req, mode, *flags);
 649
 650         ldlm_resource_dump(D_INFO, res);
 651         RETURN(LDLM_ITER_CONTINUE);
 652 }
 653
 654 /**
 655  * Flock completion callback function.
 656  *
 657  * \param lock [in,out]: A lock to be handled
 658  * \param flags    [in]: flags
 659  * \param *data    [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
 660  *
 661  * \retval 0    : success
 662  * \retval <0   : failure
 663  */
 664 int
 665 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 666 {
 667         struct file_lock *getlk = lock->l_ast_data;
 668         struct obd_device *obd;
 669         enum ldlm_error err;
 670         int rc = 0;
 671         ENTRY;
 672
 673         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT2, 4);
 674         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT3)) {
 675                 lock_res_and_lock(lock);
 676                 lock->l_flags |= LDLM_FL_FAIL_LOC;
 677                 unlock_res_and_lock(lock);
 678                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT3, 4);
 679         }
 680         CDEBUG(D_DLMTRACE, "flags: %#llx data: %p getlk: %p\n",
 681                flags, data, getlk);
 682
 683         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 684
 685         if (flags & LDLM_FL_FAILED)
 686                 goto granted;
 687
 688         if (!(flags & LDLM_FL_BLOCKED_MASK)) {
 689                 if (NULL == data)
 690                         /* mds granted the lock in the reply */
 691                         goto granted;
 692                 /* CP AST RPC: lock get granted, wake it up */
 693                 wake_up(&lock->l_waitq);
 694                 RETURN(0);
 695         }
 696
 697         LDLM_DEBUG(lock,
 698                    "client-side enqueue returned a blocked lock, sleeping");
 699         obd = class_exp2obd(lock->l_conn_export);
 700
 701         /* Go to sleep until the lock is granted. */
 702         rc = l_wait_event_abortable(lock->l_waitq,
 703                                     is_granted_or_cancelled(lock));
 704         if (rc < 0) {
 705                 /* take lock off the deadlock detection hash list. */
 706                 lock_res_and_lock(lock);
 707                 ldlm_flock_blocking_unlink(lock);
 708
 709                 /* client side - set flag to prevent lock from being
 710                  * put on LRU list
 711                  */
 712                 ldlm_set_cbpending(lock);
 713                 unlock_res_and_lock(lock);
 714
 715                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 716                            rc);
 717                 RETURN(rc);
 718         }
 719
 720 granted:
 721         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
 722
 723         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT4)) {
 724                 lock_res_and_lock(lock);
 725                 /* DEADLOCK is always set with CBPENDING */
 726                 lock->l_flags |= LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING;
 727                 unlock_res_and_lock(lock);
 728                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT4, 4);
 729         }
 730         if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_CP_CB_WAIT5)) {
 731                 lock_res_and_lock(lock);
 732                 /* DEADLOCK is always set with CBPENDING */
 733                 lock->l_flags |= (LDLM_FL_FAIL_LOC |
 734                                   LDLM_FL_FLOCK_DEADLOCK | LDLM_FL_CBPENDING);
 735                 unlock_res_and_lock(lock);
 736                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT5, 4);
 737         }
 738
 739         lock_res_and_lock(lock);
 740
 741
 742         /* Protect against race where lock could have been just destroyed
 743          * due to overlap in ldlm_process_flock_lock().
 744          */
 745         if (ldlm_is_destroyed(lock)) {
 746                 unlock_res_and_lock(lock);
 747                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 748
 749                 /* An error is still to be returned, to propagate it up to
 750                  * ldlm_cli_enqueue_fini() caller. */
 751                 RETURN(-EIO);
 752         }
 753
 754         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 755         ldlm_resource_unlink_lock(lock);
 756
 757         /* Import invalidation. We need to actually release the lock
 758          * references being held, so that it can go away. No point in
 759          * holding the lock even if app still believes it has it, since
 760          * server already dropped it anyway. Only for granted locks too.
 761          */
 762         /* Do the same for DEADLOCK'ed locks. */
 763         if (ldlm_is_failed(lock) || ldlm_is_flock_deadlock(lock)) {
 764                 int mode;
 765
 766                 if (flags & LDLM_FL_TEST_LOCK)
 767                         LASSERT(ldlm_is_test_lock(lock));
 768
 769                 if (ldlm_is_test_lock(lock) || ldlm_is_flock_deadlock(lock))
 770                         mode = getlk->fl_type;
 771                 else
 772                         mode = lock->l_req_mode;
 773
 774                 if (ldlm_is_flock_deadlock(lock)) {
 775                         LDLM_DEBUG(lock, "client-side enqueue deadlock "
 776                                    "received");
 777                         rc = -EDEADLK;
 778                 }
 779                 ldlm_flock_destroy(lock, mode, LDLM_FL_WAIT_NOREPROC);
 780                 unlock_res_and_lock(lock);
 781
 782                 /* Need to wake up the waiter if we were evicted */
 783                 wake_up(&lock->l_waitq);
 784
 785                 /* An error is still to be returned, to propagate it up to
 786                  * ldlm_cli_enqueue_fini() caller.
 787                  */
 788                 RETURN(rc ? : -EIO);
 789         }
 790
 791         LDLM_DEBUG(lock, "client-side enqueue granted");
 792
 793         if (flags & LDLM_FL_TEST_LOCK) {
 794                 /*
 795                  * fcntl(F_GETLK) request
 796                  * The old mode was saved in getlk->fl_type so that if the mode
 797                  * in the lock changes we can decref the appropriate refcount.
 798                  */
 799                 LASSERT(ldlm_is_test_lock(lock));
 800                 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
 801                 switch (lock->l_granted_mode) {
 802                 case LCK_PR:
 803                         getlk->fl_type = F_RDLCK;
 804                         break;
 805                 case LCK_PW:
 806                         getlk->fl_type = F_WRLCK;
 807                         break;
 808                 default:
 809                         getlk->fl_type = F_UNLCK;
 810                 }
 811                 getlk->fl_pid = (pid_t)lock->l_policy_data.l_flock.pid;
 812                 getlk->fl_start = (loff_t)lock->l_policy_data.l_flock.start;
 813                 getlk->fl_end = (loff_t)lock->l_policy_data.l_flock.end;
 814         } else {
 815                 __u64 noreproc = LDLM_FL_WAIT_NOREPROC;
 816
 817                 /* We need to reprocess the lock to do merges or splits
 818                  * with existing locks owned by this process.
 819                  */
 820                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 821         }
 822         unlock_res_and_lock(lock);
 823         RETURN(rc);
 824 }
 825 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 826
 827 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 828                             void *data, int flag)
 829 {
 830         ENTRY;
 831
 832         LASSERT(lock);
 833         LASSERT(flag == LDLM_CB_CANCELING);
 834
 835         /* take lock off the deadlock detection hash list. */
 836         lock_res_and_lock(lock);
 837         ldlm_flock_blocking_unlink(lock);
 838         unlock_res_and_lock(lock);
 839         RETURN(0);
 840 }
 841
 842 void ldlm_flock_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
 843                                      union ldlm_policy_data *lpolicy)
 844 {
 845         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
 846         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
 847         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
 848         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
 849 }
 850
 851 void ldlm_flock_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
 852                                      union ldlm_wire_policy_data *wpolicy)
 853 {
 854         memset(wpolicy, 0, sizeof(*wpolicy));
 855         wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
 856         wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
 857         wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
 858         wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
 859 }
 860
 861 /*
 862  * Export handle<->flock hash operations.
 863  */
 864 static unsigned
 865 ldlm_export_flock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
 866 {
 867         return cfs_hash_u64_hash(*(__u64 *)key, mask);
 868 }
 869
 870 static void *
 871 ldlm_export_flock_key(struct hlist_node *hnode)
 872 {
 873         struct ldlm_lock *lock;
 874
 875         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 876         return &lock->l_policy_data.l_flock.owner;
 877 }
 878
 879 static int
 880 ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode)
 881 {
 882         return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
 883 }
 884
 885 static void *
 886 ldlm_export_flock_object(struct hlist_node *hnode)
 887 {
 888         return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 889 }
 890
 891 static void
 892 ldlm_export_flock_get(struct cfs_hash *hs, struct hlist_node *hnode)
 893 {
 894         struct ldlm_lock *lock;
 895         struct ldlm_flock *flock;
 896
 897         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 898         LDLM_LOCK_GET(lock);
 899
 900         flock = &lock->l_policy_data.l_flock;
 901         LASSERT(flock->blocking_export != NULL);
 902         class_export_get(flock->blocking_export);
 903         atomic_inc(&flock->blocking_refs);
 904 }
 905
 906 static void
 907 ldlm_export_flock_put(struct cfs_hash *hs, struct hlist_node *hnode)
 908 {
 909         struct ldlm_lock *lock;
 910         struct ldlm_flock *flock;
 911
 912         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 913
 914         flock = &lock->l_policy_data.l_flock;
 915         LASSERT(flock->blocking_export != NULL);
 916         class_export_put(flock->blocking_export);
 917         if (atomic_dec_and_test(&flock->blocking_refs)) {
 918                 flock->blocking_owner = 0;
 919                 flock->blocking_export = NULL;
 920         }
 921         LDLM_LOCK_RELEASE(lock);
 922 }
 923
 924 static struct cfs_hash_ops ldlm_export_flock_ops = {
 925         .hs_hash        = ldlm_export_flock_hash,
 926         .hs_key         = ldlm_export_flock_key,
 927         .hs_keycmp      = ldlm_export_flock_keycmp,
 928         .hs_object      = ldlm_export_flock_object,
 929         .hs_get         = ldlm_export_flock_get,
 930         .hs_put         = ldlm_export_flock_put,
 931         .hs_put_locked  = ldlm_export_flock_put,
 932 };
 933
 934 int ldlm_init_flock_export(struct obd_export *exp)
 935 {
 936         if( strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_MDT_NAME) != 0)
 937                 RETURN(0);
 938
 939         exp->exp_flock_hash =
 940                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
 941                                 HASH_EXP_LOCK_CUR_BITS,
 942                                 HASH_EXP_LOCK_MAX_BITS,
 943                                 HASH_EXP_LOCK_BKT_BITS, 0,
 944                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 945                                 &ldlm_export_flock_ops,
 946                                 CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE);
 947         if (!exp->exp_flock_hash)
 948                 RETURN(-ENOMEM);
 949
 950         RETURN(0);
 951 }
 952
 953 void ldlm_destroy_flock_export(struct obd_export *exp)
 954 {
 955         ENTRY;
 956         if (exp->exp_flock_hash) {
 957                 cfs_hash_putref(exp->exp_flock_hash);
 958                 exp->exp_flock_hash = NULL;
 959         }
 960         EXIT;
 961 }