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) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_LDLM
  25
  26 #ifdef __KERNEL__
  27 #include <linux/lustre_dlm.h>
  28 #include <linux/obd_support.h>
  29 #include <linux/obd_class.h>
  30 #include <linux/lustre_lib.h>
  31 #include <portals/list.h>
  32 #else
  33 #include <liblustre.h>
  34 #endif
  35
  36 #include "ldlm_internal.h"
  37
  38 #define l_flock_waitq   l_lru
  39
  40 static struct list_head ldlm_flock_waitq = LIST_HEAD_INIT(ldlm_flock_waitq);
  41
  42 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
  43                             void *data, int flag);
  44
  45 /**
  46  * list_for_remaining_safe - iterate over the remaining entries in a list
  47  *              and safeguard against removal of a list entry.
  48  * @pos:        the &struct list_head to use as a loop counter. pos MUST
  49  *              have been initialized prior to using it in this macro.
  50  * @n:          another &struct list_head to use as temporary storage
  51  * @head:       the head for your list.
  52  */
  53 #define list_for_remaining_safe(pos, n, head) \
  54         for (n = pos->next; pos != (head); pos = n, n = pos->next)
  55
  56 static inline int
  57 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
  58 {
  59         return((new->l_policy_data.l_flock.pid ==
  60                 lock->l_policy_data.l_flock.pid) &&
  61                (new->l_export == lock->l_export));
  62 }
  63
  64 static inline int
  65 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
  66 {
  67         return((new->l_policy_data.l_flock.start <=
  68                 lock->l_policy_data.l_flock.end) &&
  69                (new->l_policy_data.l_flock.end >=
  70                 lock->l_policy_data.l_flock.start));
  71 }
  72
  73 static inline void
  74 ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, int flags)
  75 {
  76         ENTRY;
  77
  78         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%x)",
  79                    mode, flags);
  80
  81         LASSERT(list_empty(&lock->l_flock_waitq));
  82
  83         list_del_init(&lock->l_res_link);
  84         if (flags == LDLM_FL_WAIT_NOREPROC) {
  85                 /* client side - set a flag to prevent sending a CANCEL */
  86                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
  87                 ldlm_lock_decref_internal(lock, mode);
  88         }
  89
  90         ldlm_lock_destroy(lock);
  91         EXIT;
  92 }
  93
  94 static int
  95 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *blocking_lock)
  96 {
  97         struct obd_export *req_export = req->l_export;
  98         struct obd_export *blocking_export = blocking_lock->l_export;
  99         pid_t req_pid = req->l_policy_data.l_flock.pid;
 100         pid_t blocking_pid = blocking_lock->l_policy_data.l_flock.pid;
 101         struct ldlm_lock *lock;
 102
 103 restart:
 104         list_for_each_entry(lock, &ldlm_flock_waitq, l_flock_waitq) {
 105                 if ((lock->l_policy_data.l_flock.pid != blocking_pid) ||
 106                     (lock->l_export != blocking_export))
 107                         continue;
 108
 109                 blocking_pid = lock->l_policy_data.l_flock.blocking_pid;
 110                 blocking_export = (struct obd_export *)(long)
 111                         lock->l_policy_data.l_flock.blocking_export;
 112                 if (blocking_pid == req_pid && blocking_export == req_export)
 113                         return 1;
 114
 115                 goto restart;
 116         }
 117
 118         return 0;
 119 }
 120
 121 int
 122 ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
 123                         ldlm_error_t *err)
 124 {
 125         struct ldlm_resource *res = req->l_resource;
 126         struct ldlm_namespace *ns = res->lr_namespace;
 127         struct list_head *tmp;
 128         struct list_head *ownlocks = NULL;
 129         struct ldlm_lock *lock = NULL;
 130         struct ldlm_lock *new = req;
 131         struct ldlm_lock *new2 = NULL;
 132         ldlm_mode_t mode = req->l_req_mode;
 133         int local = ns->ns_client;
 134         int added = (mode == LCK_NL);
 135         int overlaps = 0;
 136         ENTRY;
 137
 138         CDEBUG(D_DLMTRACE, "flags %#x pid "LPU64" mode %u start "LPU64" end "
 139                LPU64"\n", *flags, new->l_policy_data.l_flock.pid, mode,
 140                req->l_policy_data.l_flock.start,
 141                req->l_policy_data.l_flock.end);
 142
 143         *err = ELDLM_OK;
 144
 145         if (local) {
 146                 /* No blocking ASTs are sent to the clients for
 147                  * Posix file & record locks */
 148                 req->l_blocking_ast = NULL;
 149         } else {
 150                 /* Called on the server for lock cancels. */
 151                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 152         }
 153
 154         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 155                 /* This loop determines where this processes locks start
 156                  * in the resource lr_granted list. */
 157                 list_for_each(tmp, &res->lr_granted) {
 158                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 159                         if (ldlm_same_flock_owner(lock, req)) {
 160                                 ownlocks = tmp;
 161                                 break;
 162                         }
 163                 }
 164         } else {
 165                 lockmode_verify(mode);
 166
 167                 /* This loop determines if there are existing locks
 168                  * that conflict with the new lock request. */
 169                 list_for_each(tmp, &res->lr_granted) {
 170                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 171
 172                         if (ldlm_same_flock_owner(lock, req)) {
 173                                 if (!ownlocks)
 174                                         ownlocks = tmp;
 175                                 continue;
 176                         }
 177
 178                         /* locks are compatible, overlap doesn't matter */
 179                         if (lockmode_compat(lock->l_granted_mode, mode))
 180                                 continue;
 181
 182                         if (!ldlm_flocks_overlap(lock, req))
 183                                 continue;
 184
 185                         if (!first_enq)
 186                                 RETURN(LDLM_ITER_CONTINUE);
 187
 188                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 189                                 ldlm_flock_destroy(req, mode, *flags);
 190                                 *err = -EAGAIN;
 191                                 RETURN(LDLM_ITER_STOP);
 192                         }
 193
 194                         if (*flags & LDLM_FL_TEST_LOCK) {
 195                                 ldlm_flock_destroy(req, mode, *flags);
 196                                 req->l_req_mode = lock->l_granted_mode;
 197                                 req->l_policy_data.l_flock.pid =
 198                                         lock->l_policy_data.l_flock.pid;
 199                                 req->l_policy_data.l_flock.start =
 200                                         lock->l_policy_data.l_flock.start;
 201                                 req->l_policy_data.l_flock.end =
 202                                         lock->l_policy_data.l_flock.end;
 203                                 *flags |= LDLM_FL_LOCK_CHANGED;
 204                                 RETURN(LDLM_ITER_STOP);
 205                         }
 206
 207                         if (ldlm_flock_deadlock(req, lock)) {
 208                                 ldlm_flock_destroy(req, mode, *flags);
 209                                 *err = -EDEADLK;
 210                                 RETURN(LDLM_ITER_STOP);
 211                         }
 212
 213                         req->l_policy_data.l_flock.blocking_pid =
 214                                 lock->l_policy_data.l_flock.pid;
 215                         req->l_policy_data.l_flock.blocking_export =
 216                                 (long)(void *)lock->l_export;
 217
 218                         LASSERT(list_empty(&req->l_flock_waitq));
 219                         list_add_tail(&req->l_flock_waitq, &ldlm_flock_waitq);
 220
 221                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 222                         *flags |= LDLM_FL_BLOCK_GRANTED;
 223                         RETURN(LDLM_ITER_STOP);
 224                 }
 225         }
 226
 227         if (*flags & LDLM_FL_TEST_LOCK) {
 228                 ldlm_flock_destroy(req, mode, *flags);
 229                 req->l_req_mode = LCK_NL;
 230                 *flags |= LDLM_FL_LOCK_CHANGED;
 231                 RETURN(LDLM_ITER_STOP);
 232         }
 233
 234         /* In case we had slept on this lock request take it off of the
 235          * deadlock detection waitq. */
 236         list_del_init(&req->l_flock_waitq);
 237
 238         /* Scan the locks owned by this process that overlap this request.
 239          * We may have to merge or split existing locks. */
 240
 241         if (!ownlocks)
 242                 ownlocks = &res->lr_granted;
 243
 244         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 245                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 246
 247                 if (!ldlm_same_flock_owner(lock, new))
 248                         break;
 249
 250                 if (lock->l_granted_mode == mode) {
 251                         /* If the modes are the same then we need to process
 252                          * locks that overlap OR adjoin the new lock. The extra
 253                          * logic condition is necessary to deal with arithmetic
 254                          * overflow and underflow. */
 255                         if ((new->l_policy_data.l_flock.start >
 256                              (lock->l_policy_data.l_flock.end + 1))
 257                             && (lock->l_policy_data.l_flock.end != ~0))
 258                                 continue;
 259
 260                         if ((new->l_policy_data.l_flock.end <
 261                              (lock->l_policy_data.l_flock.start - 1))
 262                             && (lock->l_policy_data.l_flock.start != 0))
 263                                 break;
 264
 265                         if (new->l_policy_data.l_flock.start <
 266                             lock->l_policy_data.l_flock.start) {
 267                                 lock->l_policy_data.l_flock.start =
 268                                         new->l_policy_data.l_flock.start;
 269                         } else {
 270                                 new->l_policy_data.l_flock.start =
 271                                         lock->l_policy_data.l_flock.start;
 272                         }
 273
 274                         if (new->l_policy_data.l_flock.end >
 275                             lock->l_policy_data.l_flock.end) {
 276                                 lock->l_policy_data.l_flock.end =
 277                                         new->l_policy_data.l_flock.end;
 278                         } else {
 279                                 new->l_policy_data.l_flock.end =
 280                                         lock->l_policy_data.l_flock.end;
 281                         }
 282
 283                         if (added) {
 284                                 ldlm_flock_destroy(lock, mode, *flags);
 285                         } else {
 286                                 new = lock;
 287                                 added = 1;
 288                         }
 289                         continue;
 290                 }
 291
 292                 if (new->l_policy_data.l_flock.start >
 293                     lock->l_policy_data.l_flock.end)
 294                         continue;
 295
 296                 if (new->l_policy_data.l_flock.end <
 297                     lock->l_policy_data.l_flock.start)
 298                         break;
 299
 300                 ++overlaps;
 301
 302                 if (new->l_policy_data.l_flock.start <=
 303                     lock->l_policy_data.l_flock.start) {
 304                         if (new->l_policy_data.l_flock.end <
 305                             lock->l_policy_data.l_flock.end) {
 306                                 lock->l_policy_data.l_flock.start =
 307                                         new->l_policy_data.l_flock.end + 1;
 308                                 break;
 309                         }
 310                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 311                         continue;
 312                 }
 313                 if (new->l_policy_data.l_flock.end >=
 314                     lock->l_policy_data.l_flock.end) {
 315                         lock->l_policy_data.l_flock.end =
 316                                 new->l_policy_data.l_flock.start - 1;
 317                         continue;
 318                 }
 319
 320                 /* split the existing lock into two locks */
 321
 322                 /* if this is an F_UNLCK operation then we could avoid
 323                  * allocating a new lock and use the req lock passed in
 324                  * with the request but this would complicate the reply
 325                  * processing since updates to req get reflected in the
 326                  * reply. The client side replays the lock request so
 327                  * it must see the original lock data in the reply. */
 328
 329                 /* XXX - if ldlm_lock_new() can sleep we should
 330                  * release the ns_lock, allocate the new lock,
 331                  * and restart processing this lock. */
 332                 new2 = ldlm_lock_create(ns, NULL, res->lr_name, LDLM_FLOCK,
 333                                         lock->l_granted_mode, NULL, NULL, NULL,
 334                                         NULL, 0);
 335                 if (!new2) {
 336                         ldlm_flock_destroy(req, lock->l_granted_mode, *flags);
 337                         *err = -ENOLCK;
 338                         RETURN(LDLM_ITER_STOP);
 339                 }
 340
 341                 new2->l_granted_mode = lock->l_granted_mode;
 342                 new2->l_policy_data.l_flock.pid =
 343                         new->l_policy_data.l_flock.pid;
 344                 new2->l_policy_data.l_flock.start =
 345                         lock->l_policy_data.l_flock.start;
 346                 new2->l_policy_data.l_flock.end =
 347                         new->l_policy_data.l_flock.start - 1;
 348                 lock->l_policy_data.l_flock.start =
 349                         new->l_policy_data.l_flock.end + 1;
 350                 new2->l_conn_export = lock->l_conn_export;
 351                 if (lock->l_export != NULL) {
 352                         new2->l_export = class_export_get(lock->l_export);
 353                         list_add(&new2->l_export_chain,
 354                                  &new2->l_export->exp_ldlm_data.led_held_locks);
 355                 }
 356                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 357                         ldlm_lock_addref_internal(new2, lock->l_granted_mode);
 358
 359                 /* insert new2 at lock */
 360                 ldlm_resource_add_lock(res, ownlocks, new2);
 361                 LDLM_LOCK_PUT(new2);
 362                 break;
 363         }
 364
 365         /* At this point we're granting the lock request. */
 366         req->l_granted_mode = req->l_req_mode;
 367
 368         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 369         if (!added) {
 370                 list_del_init(&req->l_res_link);
 371                 /* insert new lock before ownlocks in list. */
 372                 ldlm_resource_add_lock(res, ownlocks, req);
 373         }
 374
 375         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 376                 if (first_enq) {
 377                         /* If this is an unlock, reprocess the waitq and
 378                          * send completions ASTs for locks that can now be
 379                          * granted. The only problem with doing this
 380                          * reprocessing here is that the completion ASTs for
 381                          * newly granted locks will be sent before the unlock
 382                          * completion is sent. It shouldn't be an issue. Also
 383                          * note that ldlm_process_flock_lock() will recurse,
 384                          * but only once because first_enq will be false from
 385                          * ldlm_reprocess_queue. */
 386                         if ((mode == LCK_NL) && overlaps) {
 387                                 struct list_head rpc_list
 388                                                     = LIST_HEAD_INIT(rpc_list);
 389                                 int rc;
 390 restart:
 391                                 res->lr_tmp = &rpc_list;
 392                                 ldlm_reprocess_queue(res, &res->lr_waiting);
 393                                 res->lr_tmp = NULL;
 394
 395                                 l_unlock(&ns->ns_lock);
 396                                 rc = ldlm_run_ast_work(res->lr_namespace,
 397                                                        &rpc_list);
 398                                 l_lock(&ns->ns_lock);
 399                                 if (rc == -ERESTART)
 400                                         GOTO(restart, -ERESTART);
 401                        }
 402                 } else {
 403                         LASSERT(req->l_completion_ast);
 404                         ldlm_add_ast_work_item(req, NULL, NULL, 0);
 405                 }
 406         }
 407
 408         /* In case we're reprocessing the requested lock we can't destroy
 409          * it until after calling ldlm_ast_work_item() above so that lawi()
 410          * can bump the reference count on req. Otherwise req could be freed
 411          * before the completion AST can be sent.  */
 412         if (added)
 413                 ldlm_flock_destroy(req, mode, *flags);
 414
 415         ldlm_resource_dump(res);
 416         RETURN(LDLM_ITER_CONTINUE);
 417 }
 418
 419 struct ldlm_flock_wait_data {
 420         struct ldlm_lock *fwd_lock;
 421         int               fwd_generation;
 422 };
 423
 424 static void
 425 ldlm_flock_interrupted_wait(void *data)
 426 {
 427         struct ldlm_lock *lock;
 428         struct lustre_handle lockh;
 429         int rc;
 430         ENTRY;
 431
 432         lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
 433
 434         /* take lock off the deadlock detection waitq. */
 435         list_del_init(&lock->l_flock_waitq);
 436
 437         ldlm_lock_decref_internal(lock, lock->l_req_mode);
 438         ldlm_lock2handle(lock, &lockh);
 439         rc = ldlm_cli_cancel(&lockh);
 440         EXIT;
 441 }
 442
 443 int
 444 ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 445 {
 446         struct ldlm_namespace *ns;
 447         struct file_lock *getlk = lock->l_ast_data;
 448         struct ldlm_flock_wait_data fwd;
 449         unsigned long irqflags;
 450         struct obd_device *obd;
 451         struct obd_import *imp = NULL;
 452         ldlm_error_t err;
 453         int rc = 0;
 454         struct l_wait_info lwi;
 455         ENTRY;
 456
 457         CDEBUG(D_DLMTRACE, "flags: 0x%x data: %p getlk: %p\n",
 458                flags, data, getlk);
 459
 460         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 461
 462         if (flags == 0) {
 463                 wake_up(&lock->l_waitq);
 464                 RETURN(0);
 465         }
 466
 467         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 468                        LDLM_FL_BLOCK_CONV)))
 469                 goto  granted;
 470
 471         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 472                    "sleeping");
 473
 474         ldlm_lock_dump(D_DLMTRACE, lock, 0);
 475
 476         fwd.fwd_lock = lock;
 477         obd = class_exp2obd(lock->l_conn_export);
 478
 479         /* if this is a local lock, then there is no import */
 480         if (obd != NULL)
 481                 imp = obd->u.cli.cl_import;
 482
 483         if (imp != NULL) {
 484                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 485                 fwd.fwd_generation = imp->imp_generation;
 486                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 487         }
 488
 489         lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
 490
 491         /* Go to sleep until the lock is granted. */
 492         rc = l_wait_event(lock->l_waitq,
 493                           ((lock->l_req_mode == lock->l_granted_mode) ||
 494                            lock->l_destroyed), &lwi);
 495
 496         if (rc) {
 497                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 498                            rc);
 499                 RETURN(rc);
 500         }
 501
 502         LASSERT(!(lock->l_destroyed));
 503
 504 granted:
 505
 506         LDLM_DEBUG(lock, "client-side enqueue waking up");
 507         ns = lock->l_resource->lr_namespace;
 508         l_lock(&ns->ns_lock);
 509
 510         /* take lock off the deadlock detection waitq. */
 511         list_del_init(&lock->l_flock_waitq);
 512
 513         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 514         list_del_init(&lock->l_res_link);
 515
 516         if (flags & LDLM_FL_TEST_LOCK) {
 517                 /* fcntl(F_GETLK) request */
 518                 /* The old mode was saved in getlk->fl_type so that if the mode
 519                  * in the lock changes we can decref the approprate refcount. */
 520                 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
 521                 switch (lock->l_granted_mode) {
 522                 case LCK_PR:
 523                         getlk->fl_type = F_RDLCK;
 524                         break;
 525                 case LCK_PW:
 526                         getlk->fl_type = F_WRLCK;
 527                         break;
 528                 default:
 529                         getlk->fl_type = F_UNLCK;
 530                 }
 531                 getlk->fl_pid = lock->l_policy_data.l_flock.pid;
 532                 getlk->fl_start = lock->l_policy_data.l_flock.start;
 533                 getlk->fl_end = lock->l_policy_data.l_flock.end;
 534         } else {
 535                 /* We need to reprocess the lock to do merges or splits
 536                  * with existing locks owned by this process. */
 537                 flags = LDLM_FL_WAIT_NOREPROC;
 538                 ldlm_process_flock_lock(lock, &flags, 1, &err);
 539         }
 540         l_unlock(&ns->ns_lock);
 541         RETURN(0);
 542 }
 543
 544 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 545                             void *data, int flag)
 546 {
 547         struct ldlm_namespace *ns;
 548         ENTRY;
 549
 550         LASSERT(lock);
 551         LASSERT(flag == LDLM_CB_CANCELING);
 552
 553         ns = lock->l_resource->lr_namespace;
 554
 555         /* take lock off the deadlock detection waitq. */
 556         l_lock(&ns->ns_lock);
 557         list_del_init(&lock->l_flock_waitq);
 558         l_unlock(&ns->ns_lock);
 559         RETURN(0);
 560 }