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