lustre/ldlm/ldlm_flock.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Hewlett-Packard Development Company LP.
   5  *   Developed under the sponsorship of the US Government under
   6  *   Subcontract No. B514193
   7  *
   8  *   This file is part of the Lustre file system, http://www.lustre.org
   9  *   Lustre is a trademark of Cluster File Systems, Inc.
  10  *
  11  *   You may have signed or agreed to another license before downloading
  12  *   this software.  If so, you are bound by the terms and conditions
  13  *   of that agreement, and the following does not apply to you.  See the
  14  *   LICENSE file included with this distribution for more information.
  15  *
  16  *   If you did not agree to a different license, then this copy of Lustre
  17  *   is open source software; you can redistribute it and/or modify it
  18  *   under the terms of version 2 of the GNU General Public License as
  19  *   published by the Free Software Foundation.
  20  *
  21  *   In either case, Lustre is distributed in the hope that it will be
  22  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  23  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  24  *   license text for more details.
  25  */
  26
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28
  29 #ifdef __KERNEL__
  30 #include <lustre_dlm.h>
  31 #include <obd_support.h>
  32 #include <obd_class.h>
  33 #include <lustre_lib.h>
  34 #include <libcfs/list.h>
  35 #else
  36 #include <liblustre.h>
  37 #include <obd_class.h>
  38 #endif
  39
  40 #include "ldlm_internal.h"
  41
  42 #define l_flock_waitq   l_lru
  43
  44 static struct list_head ldlm_flock_waitq = CFS_LIST_HEAD_INIT(ldlm_flock_waitq);
  45
  46 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
  47                             void *data, int flag);
  48
  49 /**
  50  * list_for_remaining_safe - iterate over the remaining entries in a list
  51  *              and safeguard against removal of a list entry.
  52  * @pos:        the &struct list_head to use as a loop counter. pos MUST
  53  *              have been initialized prior to using it in this macro.
  54  * @n:          another &struct list_head to use as temporary storage
  55  * @head:       the head for your list.
  56  */
  57 #define list_for_remaining_safe(pos, n, head) \
  58         for (n = pos->next; pos != (head); pos = n, n = pos->next)
  59
  60 static inline int
  61 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
  62 {
  63         return((new->l_policy_data.l_flock.pid ==
  64                 lock->l_policy_data.l_flock.pid) &&
  65                (new->l_export == lock->l_export));
  66 }
  67
  68 static inline int
  69 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
  70 {
  71         return((new->l_policy_data.l_flock.start <=
  72                 lock->l_policy_data.l_flock.end) &&
  73                (new->l_policy_data.l_flock.end >=
  74                 lock->l_policy_data.l_flock.start));
  75 }
  76
  77 static inline void
  78 ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, int flags)
  79 {
  80         ENTRY;
  81
  82         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%x)",
  83                    mode, flags);
  84
  85         LASSERT(list_empty(&lock->l_flock_waitq));
  86
  87         list_del_init(&lock->l_res_link);
  88         if (flags == LDLM_FL_WAIT_NOREPROC) {
  89                 /* client side - set a flag to prevent sending a CANCEL */
  90                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
  91                 ldlm_lock_decref_internal(lock, mode);
  92         }
  93
  94         ldlm_lock_destroy_nolock(lock);
  95         EXIT;
  96 }
  97
  98 static int
  99 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *blocking_lock)
 100 {
 101         struct obd_export *req_export = req->l_export;
 102         struct obd_export *blocking_export = blocking_lock->l_export;
 103         pid_t req_pid = req->l_policy_data.l_flock.pid;
 104         pid_t blocking_pid = blocking_lock->l_policy_data.l_flock.pid;
 105         struct ldlm_lock *lock;
 106
 107 restart:
 108         list_for_each_entry(lock, &ldlm_flock_waitq, l_flock_waitq) {
 109                 if ((lock->l_policy_data.l_flock.pid != blocking_pid) ||
 110                     (lock->l_export != blocking_export))
 111                         continue;
 112
 113                 blocking_pid = lock->l_policy_data.l_flock.blocking_pid;
 114                 blocking_export = (struct obd_export *)(long)
 115                         lock->l_policy_data.l_flock.blocking_export;
 116                 if (blocking_pid == req_pid && blocking_export == req_export)
 117                         return 1;
 118
 119                 goto restart;
 120         }
 121
 122         return 0;
 123 }
 124
 125 int
 126 ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
 127                         ldlm_error_t *err, struct list_head *work_list)
 128 {
 129         struct ldlm_resource *res = req->l_resource;
 130         struct ldlm_namespace *ns = res->lr_namespace;
 131         struct list_head *tmp;
 132         struct list_head *ownlocks = NULL;
 133         struct ldlm_lock *lock = NULL;
 134         struct ldlm_lock *new = req;
 135         struct ldlm_lock *new2 = NULL;
 136         ldlm_mode_t mode = req->l_req_mode;
 137         int local = ns_is_client(ns);
 138         int added = (mode == LCK_NL);
 139         int overlaps = 0;
 140         ENTRY;
 141
 142         CDEBUG(D_DLMTRACE, "flags %#x pid %u mode %u start "LPU64" end "LPU64
 143                "\n", *flags, new->l_policy_data.l_flock.pid, mode,
 144                req->l_policy_data.l_flock.start,
 145                req->l_policy_data.l_flock.end);
 146
 147         *err = ELDLM_OK;
 148
 149         if (local) {
 150                 /* No blocking ASTs are sent to the clients for
 151                  * Posix file & record locks */
 152                 req->l_blocking_ast = NULL;
 153         } else {
 154                 /* Called on the server for lock cancels. */
 155                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 156         }
 157
 158         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 159                 /* This loop determines where this processes locks start
 160                  * in the resource lr_granted list. */
 161                 list_for_each(tmp, &res->lr_granted) {
 162                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 163                         if (ldlm_same_flock_owner(lock, req)) {
 164                                 ownlocks = tmp;
 165                                 break;
 166                         }
 167                 }
 168         } else {
 169                 lockmode_verify(mode);
 170
 171                 /* This loop determines if there are existing locks
 172                  * that conflict with the new lock request. */
 173                 list_for_each(tmp, &res->lr_granted) {
 174                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 175
 176                         if (ldlm_same_flock_owner(lock, req)) {
 177                                 if (!ownlocks)
 178                                         ownlocks = tmp;
 179                                 continue;
 180                         }
 181
 182                         /* locks are compatible, overlap doesn't matter */
 183                         if (lockmode_compat(lock->l_granted_mode, mode))
 184                                 continue;
 185
 186                         if (!ldlm_flocks_overlap(lock, req))
 187                                 continue;
 188
 189                         if (!first_enq)
 190                                 RETURN(LDLM_ITER_CONTINUE);
 191
 192                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 193                                 ldlm_flock_destroy(req, mode, *flags);
 194                                 *err = -EAGAIN;
 195                                 RETURN(LDLM_ITER_STOP);
 196                         }
 197
 198                         if (*flags & LDLM_FL_TEST_LOCK) {
 199                                 ldlm_flock_destroy(req, mode, *flags);
 200                                 req->l_req_mode = lock->l_granted_mode;
 201                                 req->l_policy_data.l_flock.pid =
 202                                         lock->l_policy_data.l_flock.pid;
 203                                 req->l_policy_data.l_flock.start =
 204                                         lock->l_policy_data.l_flock.start;
 205                                 req->l_policy_data.l_flock.end =
 206                                         lock->l_policy_data.l_flock.end;
 207                                 *flags |= LDLM_FL_LOCK_CHANGED;
 208                                 RETURN(LDLM_ITER_STOP);
 209                         }
 210
 211                         if (ldlm_flock_deadlock(req, lock)) {
 212                                 ldlm_flock_destroy(req, mode, *flags);
 213                                 *err = -EDEADLK;
 214                                 RETURN(LDLM_ITER_STOP);
 215                         }
 216
 217                         req->l_policy_data.l_flock.blocking_pid =
 218                                 lock->l_policy_data.l_flock.pid;
 219                         req->l_policy_data.l_flock.blocking_export =
 220                                 (long)(void *)lock->l_export;
 221
 222                         LASSERT(list_empty(&req->l_flock_waitq));
 223                         list_add_tail(&req->l_flock_waitq, &ldlm_flock_waitq);
 224
 225                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 226                         *flags |= LDLM_FL_BLOCK_GRANTED;
 227                         RETURN(LDLM_ITER_STOP);
 228                 }
 229         }
 230
 231         if (*flags & LDLM_FL_TEST_LOCK) {
 232                 ldlm_flock_destroy(req, mode, *flags);
 233                 req->l_req_mode = LCK_NL;
 234                 *flags |= LDLM_FL_LOCK_CHANGED;
 235                 RETURN(LDLM_ITER_STOP);
 236         }
 237
 238         /* In case we had slept on this lock request take it off of the
 239          * deadlock detection waitq. */
 240         list_del_init(&req->l_flock_waitq);
 241
 242         /* Scan the locks owned by this process that overlap this request.
 243          * We may have to merge or split existing locks. */
 244
 245         if (!ownlocks)
 246                 ownlocks = &res->lr_granted;
 247
 248         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 249                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 250
 251                 if (!ldlm_same_flock_owner(lock, new))
 252                         break;
 253
 254                 if (lock->l_granted_mode == mode) {
 255                         /* If the modes are the same then we need to process
 256                          * locks that overlap OR adjoin the new lock. The extra
 257                          * logic condition is necessary to deal with arithmetic
 258                          * overflow and underflow. */
 259                         if ((new->l_policy_data.l_flock.start >
 260                              (lock->l_policy_data.l_flock.end + 1))
 261                             && (lock->l_policy_data.l_flock.end !=
 262                                 OBD_OBJECT_EOF))
 263                                 continue;
 264
 265                         if ((new->l_policy_data.l_flock.end <
 266                              (lock->l_policy_data.l_flock.start - 1))
 267                             && (lock->l_policy_data.l_flock.start != 0))
 268                                 break;
 269
 270                         if (new->l_policy_data.l_flock.start <
 271                             lock->l_policy_data.l_flock.start) {
 272                                 lock->l_policy_data.l_flock.start =
 273                                         new->l_policy_data.l_flock.start;
 274                         } else {
 275                                 new->l_policy_data.l_flock.start =
 276                                         lock->l_policy_data.l_flock.start;
 277                         }
 278
 279                         if (new->l_policy_data.l_flock.end >
 280                             lock->l_policy_data.l_flock.end) {
 281                                 lock->l_policy_data.l_flock.end =
 282                                         new->l_policy_data.l_flock.end;
 283                         } else {
 284                                 new->l_policy_data.l_flock.end =
 285                                         lock->l_policy_data.l_flock.end;
 286                         }
 287
 288                         if (added) {
 289                                 ldlm_flock_destroy(lock, mode, *flags);
 290                         } else {
 291                                 new = lock;
 292                                 added = 1;
 293                         }
 294                         continue;
 295                 }
 296
 297                 if (new->l_policy_data.l_flock.start >
 298                     lock->l_policy_data.l_flock.end)
 299                         continue;
 300
 301                 if (new->l_policy_data.l_flock.end <
 302                     lock->l_policy_data.l_flock.start)
 303                         break;
 304
 305                 ++overlaps;
 306
 307                 if (new->l_policy_data.l_flock.start <=
 308                     lock->l_policy_data.l_flock.start) {
 309                         if (new->l_policy_data.l_flock.end <
 310                             lock->l_policy_data.l_flock.end) {
 311                                 lock->l_policy_data.l_flock.start =
 312                                         new->l_policy_data.l_flock.end + 1;
 313                                 break;
 314                         }
 315                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 316                         continue;
 317                 }
 318                 if (new->l_policy_data.l_flock.end >=
 319                     lock->l_policy_data.l_flock.end) {
 320                         lock->l_policy_data.l_flock.end =
 321                                 new->l_policy_data.l_flock.start - 1;
 322                         continue;
 323                 }
 324
 325                 /* split the existing lock into two locks */
 326
 327                 /* if this is an F_UNLCK operation then we could avoid
 328                  * allocating a new lock and use the req lock passed in
 329                  * with the request but this would complicate the reply
 330                  * processing since updates to req get reflected in the
 331                  * reply. The client side replays the lock request so
 332                  * it must see the original lock data in the reply. */
 333
 334                 /* XXX - if ldlm_lock_new() can sleep we should
 335                  * release the ns_lock, allocate the new lock,
 336                  * and restart processing this lock. */
 337                 new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
 338                                         lock->l_granted_mode, NULL, NULL, NULL,
 339                                         NULL, 0);
 340                 if (!new2) {
 341                         ldlm_flock_destroy(req, lock->l_granted_mode, *flags);
 342                         *err = -ENOLCK;
 343                         RETURN(LDLM_ITER_STOP);
 344                 }
 345
 346                 new2->l_granted_mode = lock->l_granted_mode;
 347                 new2->l_policy_data.l_flock.pid =
 348                         new->l_policy_data.l_flock.pid;
 349                 new2->l_policy_data.l_flock.start =
 350                         lock->l_policy_data.l_flock.start;
 351                 new2->l_policy_data.l_flock.end =
 352                         new->l_policy_data.l_flock.start - 1;
 353                 lock->l_policy_data.l_flock.start =
 354                         new->l_policy_data.l_flock.end + 1;
 355                 new2->l_conn_export = lock->l_conn_export;
 356                 if (lock->l_export != NULL) {
 357                         new2->l_export = class_export_get(lock->l_export);
 358                         spin_lock(&new2->l_export->exp_ldlm_data.led_lock);
 359                         list_add(&new2->l_export_chain,
 360                                  &new2->l_export->exp_ldlm_data.led_held_locks);
 361                         spin_unlock(&new2->l_export->exp_ldlm_data.led_lock);
 362                 }
 363                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 364                         ldlm_lock_addref_internal(new2, lock->l_granted_mode);
 365
 366                 /* insert new2 at lock */
 367                 ldlm_resource_add_lock(res, ownlocks, new2);
 368                 LDLM_LOCK_PUT(new2);
 369                 break;
 370         }
 371
 372         /* At this point we're granting the lock request. */
 373         req->l_granted_mode = req->l_req_mode;
 374
 375         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 376         if (!added) {
 377                 list_del_init(&req->l_res_link);
 378                 /* insert new lock before ownlocks in list. */
 379                 ldlm_resource_add_lock(res, ownlocks, req);
 380         }
 381
 382         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 383                 if (first_enq) {
 384                         /* If this is an unlock, reprocess the waitq and
 385                          * send completions ASTs for locks that can now be
 386                          * granted. The only problem with doing this
 387                          * reprocessing here is that the completion ASTs for
 388                          * newly granted locks will be sent before the unlock
 389                          * completion is sent. It shouldn't be an issue. Also
 390                          * note that ldlm_process_flock_lock() will recurse,
 391                          * but only once because first_enq will be false from
 392                          * ldlm_reprocess_queue. */
 393                         if ((mode == LCK_NL) && overlaps) {
 394                                 struct list_head rpc_list
 395                                                     = CFS_LIST_HEAD_INIT(rpc_list);
 396                                 int rc;
 397 restart:
 398                                 ldlm_reprocess_queue(res, &res->lr_waiting,
 399                                                      &rpc_list);
 400
 401                                 unlock_res(res);
 402                                 rc = ldlm_run_bl_ast_work(&rpc_list);
 403                                 lock_res(res);
 404                                 if (rc == -ERESTART)
 405                                         GOTO(restart, -ERESTART);
 406                        }
 407                 } else {
 408                         LASSERT(req->l_completion_ast);
 409                         ldlm_add_ast_work_item(req, NULL, work_list);
 410                 }
 411         }
 412
 413         /* In case we're reprocessing the requested lock we can't destroy
 414          * it until after calling ldlm_ast_work_item() above so that lawi()
 415          * can bump the reference count on req. Otherwise req could be freed
 416          * before the completion AST can be sent.  */
 417         if (added)
 418                 ldlm_flock_destroy(req, mode, *flags);
 419
 420         ldlm_resource_dump(D_OTHER, res);
 421         RETURN(LDLM_ITER_CONTINUE);
 422 }
 423
 424 struct ldlm_flock_wait_data {
 425         struct ldlm_lock *fwd_lock;
 426         int               fwd_generation;
 427 };
 428
 429 static void
 430 ldlm_flock_interrupted_wait(void *data)
 431 {
 432         struct ldlm_lock *lock;
 433         struct lustre_handle lockh;
 434         ENTRY;
 435
 436         lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
 437
 438         /* take lock off the deadlock detection waitq. */
 439         list_del_init(&lock->l_flock_waitq);
 440
 441         /* client side - set flag to prevent lock from being put on lru list */
 442         lock->l_flags |= LDLM_FL_CBPENDING;
 443
 444         ldlm_lock_decref_internal(lock, lock->l_req_mode);
 445         ldlm_lock2handle(lock, &lockh);
 446         ldlm_cli_cancel(&lockh);
 447         EXIT;
 448 }
 449
 450 int
 451 ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 452 {
 453         struct ldlm_namespace *ns;
 454         cfs_flock_t *getlk = lock->l_ast_data;
 455         struct ldlm_flock_wait_data fwd;
 456         struct obd_device *obd;
 457         struct obd_import *imp = NULL;
 458         ldlm_error_t err;
 459         int rc = 0;
 460         struct l_wait_info lwi;
 461         ENTRY;
 462
 463         CDEBUG(D_DLMTRACE, "flags: 0x%x data: %p getlk: %p\n",
 464                flags, data, getlk);
 465
 466         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 467
 468         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 469                        LDLM_FL_BLOCK_CONV)))
 470                 goto  granted;
 471
 472         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 473                    "sleeping");
 474
 475         fwd.fwd_lock = lock;
 476         obd = class_exp2obd(lock->l_conn_export);
 477
 478         /* if this is a local lock, then there is no import */
 479         if (obd != NULL)
 480                 imp = obd->u.cli.cl_import;
 481
 482         if (imp != NULL) {
 483                 spin_lock(&imp->imp_lock);
 484                 fwd.fwd_generation = imp->imp_generation;
 485                 spin_unlock(&imp->imp_lock);
 486         }
 487
 488         lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
 489
 490         /* Go to sleep until the lock is granted. */
 491         rc = l_wait_event(lock->l_waitq,
 492                           ((lock->l_req_mode == lock->l_granted_mode) ||
 493                            lock->l_destroyed), &lwi);
 494
 495         LDLM_DEBUG(lock, "client-side enqueue waking up: rc = %d", rc);
 496         RETURN(rc);
 497
 498 granted:
 499
 500         LDLM_DEBUG(lock, "client-side enqueue granted");
 501         ns = lock->l_resource->lr_namespace;
 502         lock_res(lock->l_resource);
 503
 504         /* take lock off the deadlock detection waitq. */
 505         list_del_init(&lock->l_flock_waitq);
 506
 507         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 508         list_del_init(&lock->l_res_link);
 509
 510         if (flags & LDLM_FL_TEST_LOCK) {
 511                 /* fcntl(F_GETLK) request */
 512                 /* The old mode was saved in getlk->fl_type so that if the mode
 513                  * in the lock changes we can decref the approprate refcount. */
 514                 ldlm_flock_destroy(lock, cfs_flock_type(getlk), LDLM_FL_WAIT_NOREPROC);
 515                 switch (lock->l_granted_mode) {
 516                 case LCK_PR:
 517                         cfs_flock_set_type(getlk, F_RDLCK);
 518                         break;
 519                 case LCK_PW:
 520                         cfs_flock_set_type(getlk, F_WRLCK);
 521                         break;
 522                 default:
 523                         cfs_flock_set_type(getlk, F_UNLCK);
 524                 }
 525                 cfs_flock_set_pid(getlk, (pid_t)lock->l_policy_data.l_flock.pid);
 526                 cfs_flock_set_start(getlk, (loff_t)lock->l_policy_data.l_flock.start);
 527                 cfs_flock_set_end(getlk, (loff_t)lock->l_policy_data.l_flock.end);
 528         } else {
 529                 int noreproc = LDLM_FL_WAIT_NOREPROC;
 530
 531                 /* We need to reprocess the lock to do merges or splits
 532                  * with existing locks owned by this process. */
 533                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 534                 if (flags == 0)
 535                         cfs_waitq_signal(&lock->l_waitq);
 536         }
 537         unlock_res(lock->l_resource);
 538         RETURN(0);
 539 }
 540 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 541
 542 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 543                             void *data, int flag)
 544 {
 545         struct ldlm_namespace *ns;
 546         ENTRY;
 547
 548         LASSERT(lock);
 549         LASSERT(flag == LDLM_CB_CANCELING);
 550
 551         ns = lock->l_resource->lr_namespace;
 552
 553         /* take lock off the deadlock detection waitq. */
 554         lock_res_and_lock(lock);
 555         list_del_init(&lock->l_flock_waitq);
 556         unlock_res_and_lock(lock);
 557         RETURN(0);
 558 }