Pileus Git - ~andy/linux/blob - drivers/staging/lustre/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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003 Hewlett-Packard Development Company LP.
  28  * Developed under the sponsorship of the US Government under
  29  * Subcontract No. B514193
  30  *
  31  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  32  * Use is subject to license terms.
  33  *
  34  * Copyright (c) 2010, 2012, Intel Corporation.
  35  */
  36 /*
  37  * This file is part of Lustre, http://www.lustre.org/
  38  * Lustre is a trademark of Sun Microsystems, Inc.
  39  */
  40
  41 /**
  42  * This file implements POSIX lock type for Lustre.
  43  * Its policy properties are start and end of extent and PID.
  44  *
  45  * These locks are only done through MDS due to POSIX semantics requiring
  46  * e.g. that locks could be only partially released and as such split into
  47  * two parts, and also that two adjacent locks from the same process may be
  48  * merged into a single wider lock.
  49  *
  50  * Lock modes are mapped like this:
  51  * PR and PW for READ and WRITE locks
  52  * NL to request a releasing of a portion of the lock
  53  *
  54  * These flock locks never timeout.
  55  */
  56
  57 #define DEBUG_SUBSYSTEM S_LDLM
  58
  59 #include <lustre_dlm.h>
  60 #include <obd_support.h>
  61 #include <obd_class.h>
  62 #include <lustre_lib.h>
  63 #include <linux/list.h>
  64
  65 #include "ldlm_internal.h"
  66
  67 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
  68                             void *data, int flag);
  69
  70 /**
  71  * list_for_remaining_safe - iterate over the remaining entries in a list
  72  *            and safeguard against removal of a list entry.
  73  * \param pos   the &struct list_head to use as a loop counter. pos MUST
  74  *            have been initialized prior to using it in this macro.
  75  * \param n     another &struct list_head to use as temporary storage
  76  * \param head  the head for your list.
  77  */
  78 #define list_for_remaining_safe(pos, n, head) \
  79         for (n = pos->next; pos != (head); pos = n, n = pos->next)
  80
  81 static inline int
  82 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
  83 {
  84         return((new->l_policy_data.l_flock.owner ==
  85                 lock->l_policy_data.l_flock.owner) &&
  86                (new->l_export == lock->l_export));
  87 }
  88
  89 static inline int
  90 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
  91 {
  92         return((new->l_policy_data.l_flock.start <=
  93                 lock->l_policy_data.l_flock.end) &&
  94                (new->l_policy_data.l_flock.end >=
  95                 lock->l_policy_data.l_flock.start));
  96 }
  97
  98 static inline int ldlm_flock_blocking_link(struct ldlm_lock *req,
  99                                            struct ldlm_lock *lock)
 100 {
 101         int rc = 0;
 102
 103         /* For server only */
 104         if (req->l_export == NULL)
 105                 return 0;
 106
 107         if (unlikely(req->l_export->exp_flock_hash == NULL)) {
 108                 rc = ldlm_init_flock_export(req->l_export);
 109                 if (rc)
 110                         goto error;
 111         }
 112
 113         LASSERT(hlist_unhashed(&req->l_exp_flock_hash));
 114
 115         req->l_policy_data.l_flock.blocking_owner =
 116                 lock->l_policy_data.l_flock.owner;
 117         req->l_policy_data.l_flock.blocking_export =
 118                 lock->l_export;
 119         req->l_policy_data.l_flock.blocking_refs = 0;
 120
 121         cfs_hash_add(req->l_export->exp_flock_hash,
 122                      &req->l_policy_data.l_flock.owner,
 123                      &req->l_exp_flock_hash);
 124 error:
 125         return rc;
 126 }
 127
 128 static inline void ldlm_flock_blocking_unlink(struct ldlm_lock *req)
 129 {
 130         /* For server only */
 131         if (req->l_export == NULL)
 132                 return;
 133
 134         check_res_locked(req->l_resource);
 135         if (req->l_export->exp_flock_hash != NULL &&
 136             !hlist_unhashed(&req->l_exp_flock_hash))
 137                 cfs_hash_del(req->l_export->exp_flock_hash,
 138                              &req->l_policy_data.l_flock.owner,
 139                              &req->l_exp_flock_hash);
 140 }
 141
 142 static inline void
 143 ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, __u64 flags)
 144 {
 145         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%llx)",
 146                    mode, flags);
 147
 148         /* Safe to not lock here, since it should be empty anyway */
 149         LASSERT(hlist_unhashed(&lock->l_exp_flock_hash));
 150
 151         list_del_init(&lock->l_res_link);
 152         if (flags == LDLM_FL_WAIT_NOREPROC &&
 153             !(lock->l_flags & LDLM_FL_FAILED)) {
 154                 /* client side - set a flag to prevent sending a CANCEL */
 155                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
 156
 157                 /* when reaching here, it is under lock_res_and_lock(). Thus,
 158                    need call the nolock version of ldlm_lock_decref_internal*/
 159                 ldlm_lock_decref_internal_nolock(lock, mode);
 160         }
 161
 162         ldlm_lock_destroy_nolock(lock);
 163 }
 164
 165 /**
 166  * POSIX locks deadlock detection code.
 167  *
 168  * Given a new lock \a req and an existing lock \a bl_lock it conflicts
 169  * with, we need to iterate through all blocked POSIX locks for this
 170  * export and see if there is a deadlock condition arising. (i.e. when
 171  * one client holds a lock on something and want a lock on something
 172  * else and at the same time another client has the opposite situation).
 173  */
 174 static int
 175 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *bl_lock)
 176 {
 177         struct obd_export *req_exp = req->l_export;
 178         struct obd_export *bl_exp = bl_lock->l_export;
 179         __u64 req_owner = req->l_policy_data.l_flock.owner;
 180         __u64 bl_owner = bl_lock->l_policy_data.l_flock.owner;
 181
 182         /* For server only */
 183         if (req_exp == NULL)
 184                 return 0;
 185
 186         class_export_get(bl_exp);
 187         while (1) {
 188                 struct obd_export *bl_exp_new;
 189                 struct ldlm_lock *lock = NULL;
 190                 struct ldlm_flock *flock;
 191
 192                 if (bl_exp->exp_flock_hash != NULL)
 193                         lock = cfs_hash_lookup(bl_exp->exp_flock_hash,
 194                                                &bl_owner);
 195                 if (lock == NULL)
 196                         break;
 197
 198                 LASSERT(req != lock);
 199                 flock = &lock->l_policy_data.l_flock;
 200                 LASSERT(flock->owner == bl_owner);
 201                 bl_owner = flock->blocking_owner;
 202                 bl_exp_new = class_export_get(flock->blocking_export);
 203                 class_export_put(bl_exp);
 204
 205                 cfs_hash_put(bl_exp->exp_flock_hash, &lock->l_exp_flock_hash);
 206                 bl_exp = bl_exp_new;
 207
 208                 if (bl_owner == req_owner && bl_exp == req_exp) {
 209                         class_export_put(bl_exp);
 210                         return 1;
 211                 }
 212         }
 213         class_export_put(bl_exp);
 214
 215         return 0;
 216 }
 217
 218 /**
 219  * Process a granting attempt for flock lock.
 220  * Must be called under ns lock held.
 221  *
 222  * This function looks for any conflicts for \a lock in the granted or
 223  * waiting queues. The lock is granted if no conflicts are found in
 224  * either queue.
 225  *
 226  * It is also responsible for splitting a lock if a portion of the lock
 227  * is released.
 228  *
 229  * If \a first_enq is 0 (ie, called from ldlm_reprocess_queue):
 230  *   - blocking ASTs have already been sent
 231  *
 232  * If \a first_enq is 1 (ie, called from ldlm_lock_enqueue):
 233  *   - blocking ASTs have not been sent yet, so list of conflicting locks
 234  *     would be collected and ASTs sent.
 235  */
 236 int
 237 ldlm_process_flock_lock(struct ldlm_lock *req, __u64 *flags, int first_enq,
 238                         ldlm_error_t *err, struct list_head *work_list)
 239 {
 240         struct ldlm_resource *res = req->l_resource;
 241         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
 242         struct list_head *tmp;
 243         struct list_head *ownlocks = NULL;
 244         struct ldlm_lock *lock = NULL;
 245         struct ldlm_lock *new = req;
 246         struct ldlm_lock *new2 = NULL;
 247         ldlm_mode_t mode = req->l_req_mode;
 248         int local = ns_is_client(ns);
 249         int added = (mode == LCK_NL);
 250         int overlaps = 0;
 251         int splitted = 0;
 252         const struct ldlm_callback_suite null_cbs = { NULL };
 253         int rc;
 254
 255         CDEBUG(D_DLMTRACE, "flags %#llx owner "LPU64" pid %u mode %u start "
 256                LPU64" end "LPU64"\n", *flags,
 257                new->l_policy_data.l_flock.owner,
 258                new->l_policy_data.l_flock.pid, mode,
 259                req->l_policy_data.l_flock.start,
 260                req->l_policy_data.l_flock.end);
 261
 262         *err = ELDLM_OK;
 263
 264         if (local) {
 265                 /* No blocking ASTs are sent to the clients for
 266                  * Posix file & record locks */
 267                 req->l_blocking_ast = NULL;
 268         } else {
 269                 /* Called on the server for lock cancels. */
 270                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 271         }
 272
 273 reprocess:
 274         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 275                 /* This loop determines where this processes locks start
 276                  * in the resource lr_granted list. */
 277                 list_for_each(tmp, &res->lr_granted) {
 278                         lock = list_entry(tmp, struct ldlm_lock,
 279                                               l_res_link);
 280                         if (ldlm_same_flock_owner(lock, req)) {
 281                                 ownlocks = tmp;
 282                                 break;
 283                         }
 284                 }
 285         } else {
 286                 lockmode_verify(mode);
 287
 288                 /* This loop determines if there are existing locks
 289                  * that conflict with the new lock request. */
 290                 list_for_each(tmp, &res->lr_granted) {
 291                         lock = list_entry(tmp, struct ldlm_lock,
 292                                               l_res_link);
 293
 294                         if (ldlm_same_flock_owner(lock, req)) {
 295                                 if (!ownlocks)
 296                                         ownlocks = tmp;
 297                                 continue;
 298                         }
 299
 300                         /* locks are compatible, overlap doesn't matter */
 301                         if (lockmode_compat(lock->l_granted_mode, mode))
 302                                 continue;
 303
 304                         if (!ldlm_flocks_overlap(lock, req))
 305                                 continue;
 306
 307                         if (!first_enq)
 308                                 return LDLM_ITER_CONTINUE;
 309
 310                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 311                                 ldlm_flock_destroy(req, mode, *flags);
 312                                 *err = -EAGAIN;
 313                                 return LDLM_ITER_STOP;
 314                         }
 315
 316                         if (*flags & LDLM_FL_TEST_LOCK) {
 317                                 ldlm_flock_destroy(req, mode, *flags);
 318                                 req->l_req_mode = lock->l_granted_mode;
 319                                 req->l_policy_data.l_flock.pid =
 320                                         lock->l_policy_data.l_flock.pid;
 321                                 req->l_policy_data.l_flock.start =
 322                                         lock->l_policy_data.l_flock.start;
 323                                 req->l_policy_data.l_flock.end =
 324                                         lock->l_policy_data.l_flock.end;
 325                                 *flags |= LDLM_FL_LOCK_CHANGED;
 326                                 return LDLM_ITER_STOP;
 327                         }
 328
 329                         /* add lock to blocking list before deadlock
 330                          * check to prevent race */
 331                         rc = ldlm_flock_blocking_link(req, lock);
 332                         if (rc) {
 333                                 ldlm_flock_destroy(req, mode, *flags);
 334                                 *err = rc;
 335                                 return LDLM_ITER_STOP;
 336                         }
 337                         if (ldlm_flock_deadlock(req, lock)) {
 338                                 ldlm_flock_blocking_unlink(req);
 339                                 ldlm_flock_destroy(req, mode, *flags);
 340                                 *err = -EDEADLK;
 341                                 return LDLM_ITER_STOP;
 342                         }
 343
 344                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 345                         *flags |= LDLM_FL_BLOCK_GRANTED;
 346                         return LDLM_ITER_STOP;
 347                 }
 348         }
 349
 350         if (*flags & LDLM_FL_TEST_LOCK) {
 351                 ldlm_flock_destroy(req, mode, *flags);
 352                 req->l_req_mode = LCK_NL;
 353                 *flags |= LDLM_FL_LOCK_CHANGED;
 354                 return LDLM_ITER_STOP;
 355         }
 356
 357         /* In case we had slept on this lock request take it off of the
 358          * deadlock detection hash list. */
 359         ldlm_flock_blocking_unlink(req);
 360
 361         /* Scan the locks owned by this process that overlap this request.
 362          * We may have to merge or split existing locks. */
 363
 364         if (!ownlocks)
 365                 ownlocks = &res->lr_granted;
 366
 367         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 368                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 369
 370                 if (!ldlm_same_flock_owner(lock, new))
 371                         break;
 372
 373                 if (lock->l_granted_mode == mode) {
 374                         /* If the modes are the same then we need to process
 375                          * locks that overlap OR adjoin the new lock. The extra
 376                          * logic condition is necessary to deal with arithmetic
 377                          * overflow and underflow. */
 378                         if ((new->l_policy_data.l_flock.start >
 379                              (lock->l_policy_data.l_flock.end + 1))
 380                             && (lock->l_policy_data.l_flock.end !=
 381                                 OBD_OBJECT_EOF))
 382                                 continue;
 383
 384                         if ((new->l_policy_data.l_flock.end <
 385                              (lock->l_policy_data.l_flock.start - 1))
 386                             && (lock->l_policy_data.l_flock.start != 0))
 387                                 break;
 388
 389                         if (new->l_policy_data.l_flock.start <
 390                             lock->l_policy_data.l_flock.start) {
 391                                 lock->l_policy_data.l_flock.start =
 392                                         new->l_policy_data.l_flock.start;
 393                         } else {
 394                                 new->l_policy_data.l_flock.start =
 395                                         lock->l_policy_data.l_flock.start;
 396                         }
 397
 398                         if (new->l_policy_data.l_flock.end >
 399                             lock->l_policy_data.l_flock.end) {
 400                                 lock->l_policy_data.l_flock.end =
 401                                         new->l_policy_data.l_flock.end;
 402                         } else {
 403                                 new->l_policy_data.l_flock.end =
 404                                         lock->l_policy_data.l_flock.end;
 405                         }
 406
 407                         if (added) {
 408                                 ldlm_flock_destroy(lock, mode, *flags);
 409                         } else {
 410                                 new = lock;
 411                                 added = 1;
 412                         }
 413                         continue;
 414                 }
 415
 416                 if (new->l_policy_data.l_flock.start >
 417                     lock->l_policy_data.l_flock.end)
 418                         continue;
 419
 420                 if (new->l_policy_data.l_flock.end <
 421                     lock->l_policy_data.l_flock.start)
 422                         break;
 423
 424                 ++overlaps;
 425
 426                 if (new->l_policy_data.l_flock.start <=
 427                     lock->l_policy_data.l_flock.start) {
 428                         if (new->l_policy_data.l_flock.end <
 429                             lock->l_policy_data.l_flock.end) {
 430                                 lock->l_policy_data.l_flock.start =
 431                                         new->l_policy_data.l_flock.end + 1;
 432                                 break;
 433                         }
 434                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 435                         continue;
 436                 }
 437                 if (new->l_policy_data.l_flock.end >=
 438                     lock->l_policy_data.l_flock.end) {
 439                         lock->l_policy_data.l_flock.end =
 440                                 new->l_policy_data.l_flock.start - 1;
 441                         continue;
 442                 }
 443
 444                 /* split the existing lock into two locks */
 445
 446                 /* if this is an F_UNLCK operation then we could avoid
 447                  * allocating a new lock and use the req lock passed in
 448                  * with the request but this would complicate the reply
 449                  * processing since updates to req get reflected in the
 450                  * reply. The client side replays the lock request so
 451                  * it must see the original lock data in the reply. */
 452
 453                 /* XXX - if ldlm_lock_new() can sleep we should
 454                  * release the lr_lock, allocate the new lock,
 455                  * and restart processing this lock. */
 456                 if (!new2) {
 457                         unlock_res_and_lock(req);
 458                         new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
 459                                                 lock->l_granted_mode, &null_cbs,
 460                                                 NULL, 0, LVB_T_NONE);
 461                         lock_res_and_lock(req);
 462                         if (!new2) {
 463                                 ldlm_flock_destroy(req, lock->l_granted_mode,
 464                                                    *flags);
 465                                 *err = -ENOLCK;
 466                                 return LDLM_ITER_STOP;
 467                         }
 468                         goto reprocess;
 469                 }
 470
 471                 splitted = 1;
 472
 473                 new2->l_granted_mode = lock->l_granted_mode;
 474                 new2->l_policy_data.l_flock.pid =
 475                         new->l_policy_data.l_flock.pid;
 476                 new2->l_policy_data.l_flock.owner =
 477                         new->l_policy_data.l_flock.owner;
 478                 new2->l_policy_data.l_flock.start =
 479                         lock->l_policy_data.l_flock.start;
 480                 new2->l_policy_data.l_flock.end =
 481                         new->l_policy_data.l_flock.start - 1;
 482                 lock->l_policy_data.l_flock.start =
 483                         new->l_policy_data.l_flock.end + 1;
 484                 new2->l_conn_export = lock->l_conn_export;
 485                 if (lock->l_export != NULL) {
 486                         new2->l_export = class_export_lock_get(lock->l_export, new2);
 487                         if (new2->l_export->exp_lock_hash &&
 488                             hlist_unhashed(&new2->l_exp_hash))
 489                                 cfs_hash_add(new2->l_export->exp_lock_hash,
 490                                              &new2->l_remote_handle,
 491                                              &new2->l_exp_hash);
 492                 }
 493                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 494                         ldlm_lock_addref_internal_nolock(new2,
 495                                                          lock->l_granted_mode);
 496
 497                 /* insert new2 at lock */
 498                 ldlm_resource_add_lock(res, ownlocks, new2);
 499                 LDLM_LOCK_RELEASE(new2);
 500                 break;
 501         }
 502
 503         /* if new2 is created but never used, destroy it*/
 504         if (splitted == 0 && new2 != NULL)
 505                 ldlm_lock_destroy_nolock(new2);
 506
 507         /* At this point we're granting the lock request. */
 508         req->l_granted_mode = req->l_req_mode;
 509
 510         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 511         if (!added) {
 512                 list_del_init(&req->l_res_link);
 513                 /* insert new lock before ownlocks in list. */
 514                 ldlm_resource_add_lock(res, ownlocks, req);
 515         }
 516
 517         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 518                 /* The only one possible case for client-side calls flock
 519                  * policy function is ldlm_flock_completion_ast inside which
 520                  * carries LDLM_FL_WAIT_NOREPROC flag. */
 521                 CERROR("Illegal parameter for client-side-only module.\n");
 522                 LBUG();
 523         }
 524
 525         /* In case we're reprocessing the requested lock we can't destroy
 526          * it until after calling ldlm_add_ast_work_item() above so that laawi()
 527          * can bump the reference count on \a req. Otherwise \a req
 528          * could be freed before the completion AST can be sent.  */
 529         if (added)
 530                 ldlm_flock_destroy(req, mode, *flags);
 531
 532         ldlm_resource_dump(D_INFO, res);
 533         return LDLM_ITER_CONTINUE;
 534 }
 535
 536 struct ldlm_flock_wait_data {
 537         struct ldlm_lock *fwd_lock;
 538         int            fwd_generation;
 539 };
 540
 541 static void
 542 ldlm_flock_interrupted_wait(void *data)
 543 {
 544         struct ldlm_lock *lock;
 545
 546         lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
 547
 548         /* take lock off the deadlock detection hash list. */
 549         lock_res_and_lock(lock);
 550         ldlm_flock_blocking_unlink(lock);
 551
 552         /* client side - set flag to prevent lock from being put on LRU list */
 553         lock->l_flags |= LDLM_FL_CBPENDING;
 554         unlock_res_and_lock(lock);
 555 }
 556
 557 /**
 558  * Flock completion callback function.
 559  *
 560  * \param lock [in,out]: A lock to be handled
 561  * \param flags    [in]: flags
 562  * \param *data    [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
 563  *
 564  * \retval 0    : success
 565  * \retval <0   : failure
 566  */
 567 int
 568 ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 569 {
 570         struct file_lock                *getlk = lock->l_ast_data;
 571         struct obd_device             *obd;
 572         struct obd_import             *imp = NULL;
 573         struct ldlm_flock_wait_data     fwd;
 574         struct l_wait_info            lwi;
 575         ldlm_error_t                err;
 576         int                          rc = 0;
 577
 578         CDEBUG(D_DLMTRACE, "flags: 0x%llx data: %p getlk: %p\n",
 579                flags, data, getlk);
 580
 581         /* Import invalidation. We need to actually release the lock
 582          * references being held, so that it can go away. No point in
 583          * holding the lock even if app still believes it has it, since
 584          * server already dropped it anyway. Only for granted locks too. */
 585         if ((lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) ==
 586             (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) {
 587                 if (lock->l_req_mode == lock->l_granted_mode &&
 588                     lock->l_granted_mode != LCK_NL &&
 589                     NULL == data)
 590                         ldlm_lock_decref_internal(lock, lock->l_req_mode);
 591
 592                 /* Need to wake up the waiter if we were evicted */
 593                 wake_up(&lock->l_waitq);
 594                 return 0;
 595         }
 596
 597         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 598
 599         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 600                        LDLM_FL_BLOCK_CONV))) {
 601                 if (NULL == data)
 602                         /* mds granted the lock in the reply */
 603                         goto granted;
 604                 /* CP AST RPC: lock get granted, wake it up */
 605                 wake_up(&lock->l_waitq);
 606                 return 0;
 607         }
 608
 609         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 610                    "sleeping");
 611         fwd.fwd_lock = lock;
 612         obd = class_exp2obd(lock->l_conn_export);
 613
 614         /* if this is a local lock, there is no import */
 615         if (NULL != obd)
 616                 imp = obd->u.cli.cl_import;
 617
 618         if (NULL != imp) {
 619                 spin_lock(&imp->imp_lock);
 620                 fwd.fwd_generation = imp->imp_generation;
 621                 spin_unlock(&imp->imp_lock);
 622         }
 623
 624         lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
 625
 626         /* Go to sleep until the lock is granted. */
 627         rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
 628
 629         if (rc) {
 630                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 631                            rc);
 632                 return rc;
 633         }
 634
 635 granted:
 636         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
 637
 638         if (lock->l_flags & LDLM_FL_DESTROYED) {
 639                 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
 640                 return 0;
 641         }
 642
 643         if (lock->l_flags & LDLM_FL_FAILED) {
 644                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed");
 645                 return -EIO;
 646         }
 647
 648         if (rc) {
 649                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 650                            rc);
 651                 return rc;
 652         }
 653
 654         LDLM_DEBUG(lock, "client-side enqueue granted");
 655
 656         lock_res_and_lock(lock);
 657
 658         /* take lock off the deadlock detection hash list. */
 659         ldlm_flock_blocking_unlink(lock);
 660
 661         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 662         list_del_init(&lock->l_res_link);
 663
 664         if (flags & LDLM_FL_TEST_LOCK) {
 665                 /* fcntl(F_GETLK) request */
 666                 /* The old mode was saved in getlk->fl_type so that if the mode
 667                  * in the lock changes we can decref the appropriate refcount.*/
 668                 ldlm_flock_destroy(lock, flock_type(getlk),
 669                                    LDLM_FL_WAIT_NOREPROC);
 670                 switch (lock->l_granted_mode) {
 671                 case LCK_PR:
 672                         flock_set_type(getlk, F_RDLCK);
 673                         break;
 674                 case LCK_PW:
 675                         flock_set_type(getlk, F_WRLCK);
 676                         break;
 677                 default:
 678                         flock_set_type(getlk, F_UNLCK);
 679                 }
 680                 flock_set_pid(getlk, (pid_t)lock->l_policy_data.l_flock.pid);
 681                 flock_set_start(getlk,
 682                                 (loff_t)lock->l_policy_data.l_flock.start);
 683                 flock_set_end(getlk,
 684                               (loff_t)lock->l_policy_data.l_flock.end);
 685         } else {
 686                 __u64 noreproc = LDLM_FL_WAIT_NOREPROC;
 687
 688                 /* We need to reprocess the lock to do merges or splits
 689                  * with existing locks owned by this process. */
 690                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 691         }
 692         unlock_res_and_lock(lock);
 693         return 0;
 694 }
 695 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 696
 697 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 698                             void *data, int flag)
 699 {
 700         LASSERT(lock);
 701         LASSERT(flag == LDLM_CB_CANCELING);
 702
 703         /* take lock off the deadlock detection hash list. */
 704         lock_res_and_lock(lock);
 705         ldlm_flock_blocking_unlink(lock);
 706         unlock_res_and_lock(lock);
 707         return 0;
 708 }
 709
 710 void ldlm_flock_policy_wire18_to_local(const ldlm_wire_policy_data_t *wpolicy,
 711                                        ldlm_policy_data_t *lpolicy)
 712 {
 713         memset(lpolicy, 0, sizeof(*lpolicy));
 714         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
 715         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
 716         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
 717         /* Compat code, old clients had no idea about owner field and
 718          * relied solely on pid for ownership. Introduced in LU-104, 2.1,
 719          * April 2011 */
 720         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_pid;
 721 }
 722
 723
 724 void ldlm_flock_policy_wire21_to_local(const ldlm_wire_policy_data_t *wpolicy,
 725                                        ldlm_policy_data_t *lpolicy)
 726 {
 727         memset(lpolicy, 0, sizeof(*lpolicy));
 728         lpolicy->l_flock.start = wpolicy->l_flock.lfw_start;
 729         lpolicy->l_flock.end = wpolicy->l_flock.lfw_end;
 730         lpolicy->l_flock.pid = wpolicy->l_flock.lfw_pid;
 731         lpolicy->l_flock.owner = wpolicy->l_flock.lfw_owner;
 732 }
 733
 734 void ldlm_flock_policy_local_to_wire(const ldlm_policy_data_t *lpolicy,
 735                                      ldlm_wire_policy_data_t *wpolicy)
 736 {
 737         memset(wpolicy, 0, sizeof(*wpolicy));
 738         wpolicy->l_flock.lfw_start = lpolicy->l_flock.start;
 739         wpolicy->l_flock.lfw_end = lpolicy->l_flock.end;
 740         wpolicy->l_flock.lfw_pid = lpolicy->l_flock.pid;
 741         wpolicy->l_flock.lfw_owner = lpolicy->l_flock.owner;
 742 }
 743
 744 /*
 745  * Export handle<->flock hash operations.
 746  */
 747 static unsigned
 748 ldlm_export_flock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
 749 {
 750         return cfs_hash_u64_hash(*(__u64 *)key, mask);
 751 }
 752
 753 static void *
 754 ldlm_export_flock_key(struct hlist_node *hnode)
 755 {
 756         struct ldlm_lock *lock;
 757
 758         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 759         return &lock->l_policy_data.l_flock.owner;
 760 }
 761
 762 static int
 763 ldlm_export_flock_keycmp(const void *key, struct hlist_node *hnode)
 764 {
 765         return !memcmp(ldlm_export_flock_key(hnode), key, sizeof(__u64));
 766 }
 767
 768 static void *
 769 ldlm_export_flock_object(struct hlist_node *hnode)
 770 {
 771         return hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 772 }
 773
 774 static void
 775 ldlm_export_flock_get(struct cfs_hash *hs, struct hlist_node *hnode)
 776 {
 777         struct ldlm_lock *lock;
 778         struct ldlm_flock *flock;
 779
 780         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 781         LDLM_LOCK_GET(lock);
 782
 783         flock = &lock->l_policy_data.l_flock;
 784         LASSERT(flock->blocking_export != NULL);
 785         class_export_get(flock->blocking_export);
 786         flock->blocking_refs++;
 787 }
 788
 789 static void
 790 ldlm_export_flock_put(struct cfs_hash *hs, struct hlist_node *hnode)
 791 {
 792         struct ldlm_lock *lock;
 793         struct ldlm_flock *flock;
 794
 795         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_flock_hash);
 796         LDLM_LOCK_RELEASE(lock);
 797
 798         flock = &lock->l_policy_data.l_flock;
 799         LASSERT(flock->blocking_export != NULL);
 800         class_export_put(flock->blocking_export);
 801         if (--flock->blocking_refs == 0) {
 802                 flock->blocking_owner = 0;
 803                 flock->blocking_export = NULL;
 804         }
 805 }
 806
 807 static cfs_hash_ops_t ldlm_export_flock_ops = {
 808         .hs_hash        = ldlm_export_flock_hash,
 809         .hs_key  = ldlm_export_flock_key,
 810         .hs_keycmp      = ldlm_export_flock_keycmp,
 811         .hs_object      = ldlm_export_flock_object,
 812         .hs_get  = ldlm_export_flock_get,
 813         .hs_put  = ldlm_export_flock_put,
 814         .hs_put_locked  = ldlm_export_flock_put,
 815 };
 816
 817 int ldlm_init_flock_export(struct obd_export *exp)
 818 {
 819         exp->exp_flock_hash =
 820                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
 821                                 HASH_EXP_LOCK_CUR_BITS,
 822                                 HASH_EXP_LOCK_MAX_BITS,
 823                                 HASH_EXP_LOCK_BKT_BITS, 0,
 824                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
 825                                 &ldlm_export_flock_ops,
 826                                 CFS_HASH_DEFAULT | CFS_HASH_NBLK_CHANGE);
 827         if (!exp->exp_flock_hash)
 828                 return -ENOMEM;
 829
 830         return 0;
 831 }
 832 EXPORT_SYMBOL(ldlm_init_flock_export);
 833
 834 void ldlm_destroy_flock_export(struct obd_export *exp)
 835 {
 836         if (exp->exp_flock_hash) {
 837                 cfs_hash_putref(exp->exp_flock_hash);
 838                 exp->exp_flock_hash = NULL;
 839         }
 840 }
 841 EXPORT_SYMBOL(ldlm_destroy_flock_export);