db2/lock/lock.c

   1 /*-
   2  * See the file LICENSE for redistribution information.
   3  *
   4  * Copyright (c) 1996, 1997
   5  *      Sleepycat Software.  All rights reserved.
   6  */
   7
   8 #include "config.h"
   9
  10 #ifndef lint
  11 static const char sccsid[] = "@(#)lock.c        10.31 (Sleepycat) 8/17/97";
  12 #endif /* not lint */
  13
  14 #ifndef NO_SYSTEM_INCLUDES
  15 #include <sys/types.h>
  16 #include <sys/mman.h>
  17 #include <sys/stat.h>
  18
  19 #include <errno.h>
  20 #include <fcntl.h>
  21 #include <stddef.h>
  22 #include <stdio.h>
  23 #include <stdlib.h>
  24 #include <string.h>
  25 #include <unistd.h>
  26 #endif
  27
  28 #include "db_int.h"
  29 #include "shqueue.h"
  30 #include "db_page.h"
  31 #include "db_shash.h"
  32 #include "lock.h"
  33 #include "common_ext.h"
  34 #include "db_am.h"
  35
  36 static void __lock_checklocker __P((DB_LOCKTAB *, struct __db_lock *, int));
  37 static int  __lock_count_locks __P((DB_LOCKREGION *));
  38 static int  __lock_count_objs __P((DB_LOCKREGION *));
  39 static int  __lock_create __P((const char *, int, DB_ENV *));
  40 static void __lock_freeobj __P((DB_LOCKTAB *, DB_LOCKOBJ *));
  41 static int  __lock_get_internal __P((DB_LOCKTAB *, u_int32_t, int, const DBT *,
  42     db_lockmode_t, struct __db_lock **));
  43 static int  __lock_grow_region __P((DB_LOCKTAB *, int, size_t));
  44 static int  __lock_put_internal __P((DB_LOCKTAB *, struct __db_lock *, int));
  45 static void __lock_remove_waiter
  46     __P((DB_LOCKTAB *, DB_LOCKOBJ *, struct __db_lock *, db_status_t));
  47 static void __lock_reset_region __P((DB_LOCKTAB *));
  48 static int  __lock_validate_region __P((DB_LOCKTAB *));
  49 #ifdef DEBUG
  50 static void __lock_dump_locker __P((DB_LOCKTAB *, DB_LOCKOBJ *));
  51 static void __lock_dump_object __P((DB_LOCKTAB *, DB_LOCKOBJ *));
  52 static void __lock_printlock __P((DB_LOCKTAB *, struct __db_lock *, int));
  53 #endif
  54
  55 /*
  56  * Create and initialize a lock region in shared memory.
  57  */
  58
  59 /*
  60  * __lock_create --
  61  *      Create the lock region.  Returns an errno.  In most cases,
  62  * the errno should be that returned by __db_ropen, in which case
  63  * an EAGAIN means that we should retry, and an EEXIST means that
  64  * the region exists and we didn't need to create it.  Any other
  65  * sort of errno should be treated as a system error, leading to a
  66  * failure of the original interface call.
  67  */
  68 static int
  69 __lock_create(path, mode, dbenv)
  70         const char *path;
  71         int mode;
  72         DB_ENV *dbenv;
  73 {
  74         struct __db_lock *lp;
  75         struct lock_header *tq_head;
  76         struct obj_header *obj_head;
  77         DB_LOCKOBJ *op;
  78         DB_LOCKREGION *lrp;
  79         u_int maxlocks;
  80         u_int32_t i;
  81         int fd, lock_modes, nelements, ret;
  82         u_int8_t *conflicts, *curaddr;
  83
  84         maxlocks = dbenv == NULL || dbenv->lk_max == 0 ?
  85             DB_LOCK_DEFAULT_N : dbenv->lk_max;
  86         lock_modes = dbenv == NULL || dbenv->lk_modes == 0 ?
  87             DB_LOCK_RW_N : dbenv->lk_modes;
  88         conflicts = dbenv == NULL || dbenv->lk_conflicts == NULL ?
  89             (u_int8_t *)db_rw_conflicts : dbenv->lk_conflicts;
  90
  91         if ((ret =
  92             __db_rcreate(dbenv, DB_APP_NONE, path, DB_DEFAULT_LOCK_FILE, mode,
  93             LOCK_REGION_SIZE(lock_modes, maxlocks, __db_tablesize(maxlocks)),
  94             &fd, &lrp)) != 0)
  95                 return (ret);
  96
  97         /* Region exists; now initialize it. */
  98         lrp->table_size = __db_tablesize(maxlocks);
  99         lrp->magic = DB_LOCKMAGIC;
 100         lrp->version = DB_LOCKVERSION;
 101         lrp->id = 0;
 102         lrp->maxlocks = maxlocks;
 103         lrp->need_dd = 0;
 104         lrp->detect = DB_LOCK_NORUN;
 105         lrp->numobjs = maxlocks;
 106         lrp->nlockers = 0;
 107         lrp->mem_bytes = ALIGN(STRING_SIZE(maxlocks), sizeof(size_t));
 108         lrp->increment = lrp->hdr.size / 2;
 109         lrp->nmodes = lock_modes;
 110         lrp->nconflicts = 0;
 111         lrp->nrequests = 0;
 112         lrp->nreleases = 0;
 113         lrp->ndeadlocks = 0;
 114
 115         /*
 116          * As we write the region, we've got to maintain the alignment
 117          * for the structures that follow each chunk.  This information
 118          * ends up being encapsulated both in here as well as in the
 119          * lock.h file for the XXX_SIZE macros.
 120          */
 121         /* Initialize conflict matrix. */
 122         curaddr = (u_int8_t *)lrp + sizeof(DB_LOCKREGION);
 123         memcpy(curaddr, conflicts, lock_modes * lock_modes);
 124         curaddr += lock_modes * lock_modes;
 125
 126         /*
 127          * Initialize hash table.
 128          */
 129         curaddr = (u_int8_t *)ALIGNP(curaddr, LOCK_HASH_ALIGN);
 130         lrp->hash_off = curaddr - (u_int8_t *)lrp;
 131         nelements = lrp->table_size;
 132         __db_hashinit(curaddr, nelements);
 133         curaddr += nelements * sizeof(DB_HASHTAB);
 134
 135         /*
 136          * Initialize locks onto a free list. Since locks contains mutexes,
 137          * we need to make sure that each lock is aligned on a MUTEX_ALIGNMENT
 138          * boundary.
 139          */
 140         curaddr = (u_int8_t *)ALIGNP(curaddr, MUTEX_ALIGNMENT);
 141         tq_head = &lrp->free_locks;
 142         SH_TAILQ_INIT(tq_head);
 143
 144         for (i = 0; i++ < maxlocks;
 145             curaddr += ALIGN(sizeof(struct __db_lock), MUTEX_ALIGNMENT)) {
 146                 lp = (struct __db_lock *)curaddr;
 147                 lp->status = DB_LSTAT_FREE;
 148                 SH_TAILQ_INSERT_HEAD(tq_head, lp, links, __db_lock);
 149         }
 150
 151         /* Initialize objects onto a free list.  */
 152         obj_head = &lrp->free_objs;
 153         SH_TAILQ_INIT(obj_head);
 154
 155         for (i = 0; i++ < maxlocks; curaddr += sizeof(DB_LOCKOBJ)) {
 156                 op = (DB_LOCKOBJ *)curaddr;
 157                 SH_TAILQ_INSERT_HEAD(obj_head, op, links, __db_lockobj);
 158         }
 159
 160         /*
 161          * Initialize the string space; as for all shared memory allocation
 162          * regions, this requires size_t alignment, since we store the
 163          * lengths of malloc'd areas in the area..
 164          */
 165         curaddr = (u_int8_t *)ALIGNP(curaddr, sizeof(size_t));
 166         lrp->mem_off = curaddr - (u_int8_t *)lrp;
 167         __db_shalloc_init(curaddr, lrp->mem_bytes);
 168
 169         /* Release the lock. */
 170         (void)__db_mutex_unlock(&lrp->hdr.lock, fd);
 171
 172         /* Now unmap the region. */
 173         if ((ret = __db_rclose(dbenv, fd, lrp)) != 0) {
 174                 (void)lock_unlink(path, 1 /* force */, dbenv);
 175                 return (ret);
 176         }
 177
 178         return (0);
 179 }
 180
 181 int
 182 lock_open(path, flags, mode, dbenv, ltp)
 183         const char *path;
 184         int flags, mode;
 185         DB_ENV *dbenv;
 186         DB_LOCKTAB **ltp;
 187 {
 188         DB_LOCKTAB *lt;
 189         int ret, retry_cnt;
 190
 191         /* Validate arguments. */
 192 #ifdef HAVE_SPINLOCKS
 193 #define OKFLAGS (DB_CREATE | DB_THREAD)
 194 #else
 195 #define OKFLAGS (DB_CREATE)
 196 #endif
 197         if ((ret = __db_fchk(dbenv, "lock_open", flags, OKFLAGS)) != 0)
 198                 return (ret);
 199
 200         /*
 201          * Create the lock table structure.
 202          */
 203         if ((lt = (DB_LOCKTAB *)calloc(1, sizeof(DB_LOCKTAB))) == NULL) {
 204                 __db_err(dbenv, "%s", strerror(errno));
 205                 return (ENOMEM);
 206         }
 207         lt->dbenv = dbenv;
 208
 209         /*
 210          * Now, create the lock region if it doesn't already exist.
 211          */
 212         retry_cnt = 0;
 213 retry:  if (LF_ISSET(DB_CREATE) &&
 214             (ret = __lock_create(path, mode, dbenv)) != 0)
 215                 if (ret == EAGAIN && ++retry_cnt < 3) {
 216                         (void)__db_sleep(1, 0);
 217                         goto retry;
 218                 } else if (ret == EEXIST) /* We did not create the region */
 219                         LF_CLR(DB_CREATE);
 220                 else
 221                         goto out;
 222
 223         /*
 224          * Finally, open the region, map it in, and increment the
 225          * reference count.
 226          */
 227         retry_cnt = 0;
 228 retry1: if ((ret = __db_ropen(dbenv, DB_APP_NONE, path, DB_DEFAULT_LOCK_FILE,
 229             LF_ISSET(~(DB_CREATE | DB_THREAD)), &lt->fd, &lt->region)) != 0) {
 230                 if (ret == EAGAIN && ++retry_cnt < 3) {
 231                         (void)__db_sleep(1, 0);
 232                         goto retry1;
 233                 }
 234                 goto out;
 235          }
 236
 237         if (lt->region->magic != DB_LOCKMAGIC) {
 238                 __db_err(dbenv, "lock_open: Bad magic number");
 239                 ret = EINVAL;
 240                 goto out;
 241         }
 242
 243         /* Check for automatic deadlock detection. */
 244         if (dbenv->lk_detect != DB_LOCK_NORUN) {
 245                 if (lt->region->detect != DB_LOCK_NORUN &&
 246                     dbenv->lk_detect != DB_LOCK_DEFAULT &&
 247                     lt->region->detect != dbenv->lk_detect) {
 248                         __db_err(dbenv,
 249                             "lock_open: incompatible deadlock detector mode");
 250                         ret = EINVAL;
 251                         goto out;
 252                 }
 253                 if (lt->region->detect == DB_LOCK_NORUN)
 254                         lt->region->detect = dbenv->lk_detect;
 255         }
 256
 257         /* Set up remaining pointers into region. */
 258         lt->conflicts = (u_int8_t *)lt->region + sizeof(DB_LOCKREGION);
 259         lt->hashtab =
 260             (DB_HASHTAB *)((u_int8_t *)lt->region + lt->region->hash_off);
 261         lt->mem = (void *)((u_int8_t *)lt->region + lt->region->mem_off);
 262         lt->reg_size = lt->region->hdr.size;
 263
 264         *ltp = lt;
 265         return (0);
 266
 267 /* Error handling. */
 268 out:    if (lt->region != NULL)
 269                 (void)__db_rclose(lt->dbenv, lt->fd, lt->region);
 270         if (LF_ISSET(DB_CREATE))
 271                 (void)lock_unlink(path, 1, lt->dbenv);
 272         free(lt);
 273         return (ret);
 274 }
 275
 276 int
 277 lock_id (lt, idp)
 278         DB_LOCKTAB *lt;
 279         u_int32_t *idp;
 280 {
 281         u_int32_t id;
 282
 283         LOCK_LOCKREGION(lt);
 284         if (lt->region->id >= DB_LOCK_MAXID)
 285                 lt->region->id = 0;
 286         id = ++lt->region->id;
 287         UNLOCK_LOCKREGION(lt);
 288
 289         *idp = id;
 290         return (0);
 291 }
 292
 293 int
 294 lock_vec(lt, locker, flags, list, nlist, elistp)
 295         DB_LOCKTAB *lt;
 296         u_int32_t locker;
 297         int flags, nlist;
 298         DB_LOCKREQ *list, **elistp;
 299 {
 300         struct __db_lock *lp;
 301         DB_LOCKOBJ *sh_obj, *sh_locker;
 302         int i, ret, run_dd;
 303
 304         /* Validate arguments. */
 305         if ((ret =
 306             __db_fchk(lt->dbenv, "lock_vec", flags, DB_LOCK_NOWAIT)) != 0)
 307                 return (ret);
 308
 309         LOCK_LOCKREGION(lt);
 310
 311         if ((ret = __lock_validate_region(lt)) != 0) {
 312                 UNLOCK_LOCKREGION(lt);
 313                 return (ret);
 314         }
 315
 316         ret = 0;
 317         for (i = 0; i < nlist && ret == 0; i++) {
 318                 switch (list[i].op) {
 319                 case DB_LOCK_GET:
 320                         ret = __lock_get_internal(lt, locker, flags,
 321                             list[i].obj, list[i].mode, &lp);
 322                         if (ret == 0)
 323                                 list[i].lock = LOCK_TO_OFFSET(lt, lp);
 324                         break;
 325                 case DB_LOCK_PUT:
 326                         lp = OFFSET_TO_LOCK(lt, list[i].lock);
 327                         if (lp->holder != locker) {
 328                                 ret = DB_LOCK_NOTHELD;
 329                                 break;
 330                         }
 331                         list[i].mode = lp->mode;
 332
 333                         /* XXX Need to copy the object. ??? */
 334                         ret = __lock_put_internal(lt, lp, 0);
 335                         break;
 336                 case DB_LOCK_PUT_ALL:
 337                         /* Find the locker. */
 338                         if ((ret = __lock_getobj(lt, locker,
 339                             NULL, DB_LOCK_LOCKER, &sh_locker)) != 0)
 340                                 break;
 341
 342                         for (lp = SH_LIST_FIRST(&sh_locker->heldby, __db_lock);
 343                             lp != NULL;
 344                             lp = SH_LIST_FIRST(&sh_locker->heldby, __db_lock)) {
 345                                 if ((ret = __lock_put_internal(lt, lp, 0)) != 0)
 346                                         break;
 347                         }
 348                         __lock_freeobj(lt, sh_locker);
 349                         lt->region->nlockers--;
 350                         break;
 351                 case DB_LOCK_PUT_OBJ:
 352
 353                         /* Look up the object in the hash table. */
 354                         __db_hashlookup(lt->hashtab, __db_lockobj, links,
 355                             list[i].obj, sh_obj, lt->region->table_size,
 356                             __lock_ohash, __lock_cmp);
 357                         if (sh_obj == NULL) {
 358                                 ret = EINVAL;
 359                                 break;
 360                         }
 361                         /*
 362                          * Release waiters first, because they won't cause
 363                          * anyone else to be awakened.  If we release the
 364                          * lockers first, all the waiters get awakened
 365                          * needlessly.
 366                          */
 367                         for (lp = SH_TAILQ_FIRST(&sh_obj->waiters, __db_lock);
 368                             lp != NULL;
 369                             lp = SH_TAILQ_FIRST(&sh_obj->waiters, __db_lock)) {
 370                                 lt->region->nreleases += lp->refcount;
 371                                 __lock_remove_waiter(lt, sh_obj, lp,
 372                                     DB_LSTAT_NOGRANT);
 373                                 __lock_checklocker(lt, lp, 1);
 374                         }
 375
 376                         for (lp = SH_TAILQ_FIRST(&sh_obj->holders, __db_lock);
 377                             lp != NULL;
 378                             lp = SH_TAILQ_FIRST(&sh_obj->holders, __db_lock)) {
 379
 380                                 lt->region->nreleases += lp->refcount;
 381                                 SH_LIST_REMOVE(lp, locker_links, __db_lock);
 382                                 SH_TAILQ_REMOVE(&sh_obj->holders, lp, links,
 383                                     __db_lock);
 384                                 lp->status = DB_LSTAT_FREE;
 385                                 SH_TAILQ_INSERT_HEAD(&lt->region->free_locks,
 386                                     lp, links, __db_lock);
 387                         }
 388
 389                         /* Now free the object. */
 390                         __lock_freeobj(lt, sh_obj);
 391                         break;
 392 #ifdef DEBUG
 393                 case DB_LOCK_DUMP:
 394                         /* Find the locker. */
 395                         if ((ret = __lock_getobj(lt, locker,
 396                             NULL, DB_LOCK_LOCKER, &sh_locker)) != 0)
 397                                 break;
 398
 399                         for (lp = SH_LIST_FIRST(&sh_locker->heldby, __db_lock);
 400                             lp != NULL;
 401                             lp = SH_LIST_NEXT(lp, locker_links, __db_lock)) {
 402                                 __lock_printlock(lt, lp, 1);
 403                                 ret = EINVAL;
 404                         }
 405                         if (ret == 0) {
 406                                 __lock_freeobj(lt, sh_locker);
 407                                 lt->region->nlockers--;
 408                         }
 409                         break;
 410 #endif
 411                 default:
 412                         ret = EINVAL;
 413                         break;
 414                 }
 415         }
 416
 417         if (lt->region->need_dd && lt->region->detect != DB_LOCK_NORUN) {
 418                 run_dd = 1;
 419                 lt->region->need_dd = 0;
 420         } else
 421                 run_dd = 0;
 422
 423         UNLOCK_LOCKREGION(lt);
 424
 425         if (ret == 0 && run_dd)
 426                 lock_detect(lt, 0, lt->region->detect);
 427
 428         if (elistp && ret != 0)
 429                 *elistp = &list[i - 1];
 430         return (ret);
 431 }
 432
 433 int
 434 lock_get(lt, locker, flags, obj, lock_mode, lock)
 435         DB_LOCKTAB *lt;
 436         u_int32_t locker;
 437         int flags;
 438         const DBT *obj;
 439         db_lockmode_t lock_mode;
 440         DB_LOCK *lock;
 441 {
 442         struct __db_lock *lockp;
 443         int ret;
 444
 445         /* Validate arguments. */
 446         if ((ret =
 447             __db_fchk(lt->dbenv, "lock_get", flags, DB_LOCK_NOWAIT)) != 0)
 448                 return (ret);
 449
 450         LOCK_LOCKREGION(lt);
 451
 452         ret = __lock_validate_region(lt);
 453         if (ret == 0 && (ret = __lock_get_internal(lt,
 454             locker, flags, obj, lock_mode, &lockp)) == 0) {
 455                 *lock = LOCK_TO_OFFSET(lt, lockp);
 456                 lt->region->nrequests++;
 457         }
 458
 459         UNLOCK_LOCKREGION(lt);
 460         return (ret);
 461 }
 462
 463 int
 464 lock_put(lt, lock)
 465         DB_LOCKTAB *lt;
 466         DB_LOCK lock;
 467 {
 468         struct __db_lock *lockp;
 469         int ret, run_dd;
 470
 471         LOCK_LOCKREGION(lt);
 472
 473         if ((ret = __lock_validate_region(lt)) != 0)
 474                 return (ret);
 475         else {
 476                 lockp = OFFSET_TO_LOCK(lt, lock);
 477                 ret = __lock_put_internal(lt, lockp, 0);
 478         }
 479
 480         __lock_checklocker(lt, lockp, 0);
 481
 482         if (lt->region->need_dd && lt->region->detect != DB_LOCK_NORUN) {
 483                 run_dd = 1;
 484                 lt->region->need_dd = 0;
 485         } else
 486                 run_dd = 0;
 487
 488         UNLOCK_LOCKREGION(lt);
 489
 490         if (ret == 0 && run_dd)
 491                 lock_detect(lt, 0, lt->region->detect);
 492
 493         return (ret);
 494 }
 495
 496 int
 497 lock_close(lt)
 498         DB_LOCKTAB *lt;
 499 {
 500         int ret;
 501
 502         if ((ret = __db_rclose(lt->dbenv, lt->fd, lt->region)) != 0)
 503                 return (ret);
 504
 505         /* Free lock table. */
 506         free(lt);
 507         return (0);
 508 }
 509
 510 int
 511 lock_unlink(path, force, dbenv)
 512         const char *path;
 513         int force;
 514         DB_ENV *dbenv;
 515 {
 516         return (__db_runlink(dbenv,
 517             DB_APP_NONE, path, DB_DEFAULT_LOCK_FILE, force));
 518 }
 519
 520 /*
 521  * XXX This looks like it could be void, but I'm leaving it returning
 522  * an int because I think it will have to when we go through and add
 523  * the appropriate error checking for the EINTR on mutexes.
 524  */
 525 static int
 526 __lock_put_internal(lt, lockp, do_all)
 527         DB_LOCKTAB *lt;
 528         struct __db_lock *lockp;
 529         int do_all;
 530 {
 531         struct __db_lock *lp_w, *lp_h, *next_waiter;
 532         DB_LOCKOBJ *sh_obj;
 533         int state_changed;
 534
 535         if (lockp->refcount == 0 || (lockp->status != DB_LSTAT_HELD &&
 536             lockp->status != DB_LSTAT_WAITING) || lockp->obj == 0) {
 537                 __db_err(lt->dbenv, "lock_put: invalid lock %lu",
 538                     (u_long)((u_int8_t *)lockp - (u_int8_t *)lt->region));
 539                 return (EINVAL);
 540         }
 541
 542         if (do_all)
 543                 lt->region->nreleases += lockp->refcount;
 544         else
 545                 lt->region->nreleases++;
 546         if (do_all == 0 && lockp->refcount > 1) {
 547                 lockp->refcount--;
 548                 return (0);
 549         }
 550
 551         /* Get the object associated with this lock. */
 552         sh_obj = (DB_LOCKOBJ *)((u_int8_t *)lockp + lockp->obj);
 553
 554         /* Remove lock from locker list. */
 555         SH_LIST_REMOVE(lockp, locker_links, __db_lock);
 556
 557         /* Remove this lock from its holders/waitlist. */
 558         if (lockp->status != DB_LSTAT_HELD)
 559                 __lock_remove_waiter(lt, sh_obj, lockp, DB_LSTAT_FREE);
 560         else
 561                 SH_TAILQ_REMOVE(&sh_obj->holders, lockp, links, __db_lock);
 562
 563         /*
 564          * We need to do lock promotion.  We also need to determine if
 565          * we're going to need to run the deadlock detector again.  If
 566          * we release locks, and there are waiters, but no one gets promoted,
 567          * then we haven't fundamentally changed the lockmgr state, so
 568          * we may still have a deadlock and we have to run again.  However,
 569          * if there were no waiters, or we actually promoted someone, then
 570          * we are OK and we don't have to run it immediately.
 571          */
 572         for (lp_w = SH_TAILQ_FIRST(&sh_obj->waiters, __db_lock),
 573             state_changed = lp_w == NULL;
 574             lp_w != NULL;
 575             lp_w = next_waiter) {
 576                 next_waiter = SH_TAILQ_NEXT(lp_w, links, __db_lock);
 577                 for (lp_h = SH_TAILQ_FIRST(&sh_obj->holders, __db_lock);
 578                     lp_h != NULL;
 579                     lp_h = SH_TAILQ_NEXT(lp_h, links, __db_lock)) {
 580                         if (CONFLICTS(lt, lp_h->mode, lp_w->mode) &&
 581                             lp_h->holder != lp_w->holder)
 582                                 break;
 583                 }
 584                 if (lp_h != NULL)       /* Found a conflict. */
 585                         break;
 586
 587                 /* No conflict, promote the waiting lock. */
 588                 SH_TAILQ_REMOVE(&sh_obj->waiters, lp_w, links, __db_lock);
 589                 lp_w->status = DB_LSTAT_PENDING;
 590                 SH_TAILQ_INSERT_TAIL(&sh_obj->holders, lp_w, links);
 591
 592                 /* Wake up waiter. */
 593                 (void)__db_mutex_unlock(&lp_w->mutex, lt->fd);
 594                 state_changed = 1;
 595         }
 596
 597         /* Check if object should be reclaimed. */
 598         if (SH_TAILQ_FIRST(&sh_obj->holders, __db_lock) == NULL) {
 599                 __db_hashremove_el(lt->hashtab, __db_lockobj, links, sh_obj,
 600                     lt->region->table_size, __lock_lhash);
 601                 __db_shalloc_free(lt->mem, SH_DBT_PTR(&sh_obj->lockobj));
 602                 SH_TAILQ_INSERT_HEAD(&lt->region->free_objs, sh_obj, links,
 603                     __db_lockobj);
 604                 state_changed = 1;
 605         }
 606
 607         /* Free lock. */
 608         lockp->status = DB_LSTAT_FREE;
 609         SH_TAILQ_INSERT_HEAD(&lt->region->free_locks, lockp, links, __db_lock);
 610
 611         /*
 612          * If we did not promote anyone; we need to run the deadlock
 613          * detector again.
 614          */
 615         if (state_changed == 0)
 616                 lt->region->need_dd = 1;
 617
 618         return (0);
 619 }
 620
 621 static int
 622 __lock_get_internal(lt, locker, flags, obj, lock_mode, lockp)
 623         DB_LOCKTAB *lt;
 624         u_int32_t locker;
 625         int flags;
 626         const DBT *obj;
 627         db_lockmode_t lock_mode;
 628         struct __db_lock **lockp;
 629 {
 630         struct __db_lock *newl, *lp;
 631         DB_LOCKOBJ *sh_obj, *sh_locker;
 632         DB_LOCKREGION *lrp;
 633         size_t newl_off;
 634         int ret;
 635
 636         ret = 0;
 637         /*
 638          * Check that lock mode is valid.
 639          */
 640
 641         lrp = lt->region;
 642         if ((u_int32_t)lock_mode >= lrp->nmodes) {
 643                 __db_err(lt->dbenv,
 644                     "lock_get: invalid lock mode %lu\n", (u_long)lock_mode);
 645                 return (EINVAL);
 646         }
 647
 648         /* Allocate a new lock.  Optimize for the common case of a grant. */
 649         if ((newl = SH_TAILQ_FIRST(&lrp->free_locks, __db_lock)) == NULL) {
 650                 if ((ret = __lock_grow_region(lt, DB_LOCK_LOCK, 0)) != 0)
 651                         return (ret);
 652                 lrp = lt->region;
 653                 newl = SH_TAILQ_FIRST(&lrp->free_locks, __db_lock);
 654         }
 655         newl_off = LOCK_TO_OFFSET(lt, newl);
 656
 657         /* Optimize for common case of granting a lock. */
 658         SH_TAILQ_REMOVE(&lrp->free_locks, newl, links, __db_lock);
 659
 660         newl->mode = lock_mode;
 661         newl->status = DB_LSTAT_HELD;
 662         newl->holder = locker;
 663         newl->refcount = 1;
 664
 665         if ((ret =
 666             __lock_getobj(lt, 0, (DBT *)obj, DB_LOCK_OBJTYPE, &sh_obj)) != 0)
 667                 return (ret);
 668
 669         lrp = lt->region;                       /* getobj might have grown */
 670         newl = OFFSET_TO_LOCK(lt, newl_off);
 671
 672         /* Now make new lock point to object */
 673         newl->obj = SH_PTR_TO_OFF(newl, sh_obj);
 674
 675         /*
 676          * Now we have a lock and an object and we need to see if we should
 677          * grant the lock.  We use a FIFO ordering so we can only grant a
 678          * new lock if it does not conflict with anyone on the holders list
 679          * OR anyone on the waiters list.  In case of conflict, we put the
 680          * new lock on the end of the waiters list.
 681          */
 682         for (lp = SH_TAILQ_FIRST(&sh_obj->holders, __db_lock);
 683             lp != NULL;
 684             lp = SH_TAILQ_NEXT(lp, links, __db_lock)) {
 685                 if (CONFLICTS(lt, lp->mode, lock_mode) &&
 686                     locker != lp->holder)
 687                         break;
 688                 else if (lp->holder == locker && lp->mode == lock_mode &&
 689                     lp->status == DB_LSTAT_HELD) {
 690                         /* Lock is already held, just inc the ref count. */
 691                         lp->refcount++;
 692                         SH_TAILQ_INSERT_HEAD(&lrp->free_locks, newl, links,
 693                             __db_lock);
 694                         *lockp = lp;
 695                         return (0);
 696                 }
 697         }
 698
 699         if (lp == NULL)
 700                 for (lp = SH_TAILQ_FIRST(&sh_obj->waiters, __db_lock);
 701                     lp != NULL;
 702                     lp = SH_TAILQ_NEXT(lp, links, __db_lock)) {
 703                         if (CONFLICTS(lt, lp->mode, lock_mode) &&
 704                             locker != lp->holder)
 705                                 break;
 706                 }
 707         if (lp == NULL)
 708                 SH_TAILQ_INSERT_TAIL(&sh_obj->holders, newl, links);
 709         else if (!(flags & DB_LOCK_NOWAIT))
 710                 SH_TAILQ_INSERT_TAIL(&sh_obj->waiters, newl, links);
 711         else {
 712                 /* Free the lock and return an error. */
 713                 newl->status = DB_LSTAT_FREE;
 714                 SH_TAILQ_INSERT_HEAD(&lrp->free_locks, newl, links, __db_lock);
 715                 return (DB_LOCK_NOTGRANTED);
 716         }
 717
 718         /*
 719          * This is really a blocker for the process, so initialize it
 720          * set.  That way the current process will block when it tries
 721          * to get it and the waking process will release it.
 722          */
 723         (void)__db_mutex_init(&newl->mutex,
 724             MUTEX_LOCK_OFFSET(lt->region, &newl->mutex));
 725         (void)__db_mutex_lock(&newl->mutex, lt->fd,
 726             lt->dbenv == NULL ? NULL : lt->dbenv->db_yield);
 727
 728         /*
 729          * Now, insert the lock onto its locker's list.
 730          */
 731         if ((ret =
 732             __lock_getobj(lt, locker, NULL, DB_LOCK_LOCKER, &sh_locker)) != 0)
 733                 return (ret);
 734
 735         lrp = lt->region;
 736         SH_LIST_INSERT_HEAD(&sh_locker->heldby, newl, locker_links, __db_lock);
 737
 738         if (lp != NULL) {
 739                 newl->status = DB_LSTAT_WAITING;
 740                 lrp->nconflicts++;
 741                 /*
 742                  * We are about to wait; must release the region mutex.
 743                  * Then, when we wakeup, we need to reacquire the region
 744                  * mutex before continuing.
 745                  */
 746                 if (lrp->detect == DB_LOCK_NORUN)
 747                         lt->region->need_dd = 1;
 748                 UNLOCK_LOCKREGION(lt);
 749
 750                 /*
 751                  * We are about to wait; before waiting, see if the deadlock
 752                  * detector should be run.
 753                  */
 754                 if (lrp->detect != DB_LOCK_NORUN)
 755                         ret = lock_detect(lt, 0, lrp->detect);
 756
 757                 (void)__db_mutex_lock(&newl->mutex,
 758                     lt->fd, lt->dbenv == NULL ? NULL : lt->dbenv->db_yield);
 759
 760                 LOCK_LOCKREGION(lt);
 761                 if (newl->status != DB_LSTAT_PENDING) {
 762                         /* Return to free list. */
 763                         __lock_checklocker(lt, newl, 0);
 764                         SH_TAILQ_INSERT_HEAD(&lrp->free_locks, newl, links,
 765                             __db_lock);
 766                         switch (newl->status) {
 767                                 case DB_LSTAT_ABORTED:
 768                                         ret = DB_LOCK_DEADLOCK;
 769                                         break;
 770                                 case DB_LSTAT_NOGRANT:
 771                                         ret = DB_LOCK_NOTGRANTED;
 772                                         break;
 773                                 default:
 774                                         ret = EINVAL;
 775                                         break;
 776                         }
 777                         newl->status = DB_LSTAT_FREE;
 778                         newl = NULL;
 779                 } else
 780                         newl->status = DB_LSTAT_HELD;
 781         }
 782
 783         *lockp = newl;
 784         return (ret);
 785 }
 786
 787 /*
 788  * This is called at every interface to verify if the region
 789  * has changed size, and if so, to remap the region in and
 790  * reset the process pointers.
 791  */
 792 static int
 793 __lock_validate_region(lt)
 794         DB_LOCKTAB *lt;
 795 {
 796         int ret;
 797
 798         if (lt->reg_size == lt->region->hdr.size)
 799                 return (0);
 800
 801         /* Grow the region. */
 802         if ((ret = __db_rremap(lt->dbenv, lt->region,
 803             lt->reg_size, lt->region->hdr.size, lt->fd, &lt->region)) != 0)
 804                 return (ret);
 805
 806         __lock_reset_region(lt);
 807
 808         return (0);
 809 }
 810
 811 /*
 812  * We have run out of space; time to grow the region.
 813  */
 814 static int
 815 __lock_grow_region(lt, which, howmuch)
 816         DB_LOCKTAB *lt;
 817         int which;
 818         size_t howmuch;
 819 {
 820         struct __db_lock *newl;
 821         struct lock_header *lock_head;
 822         struct obj_header *obj_head;
 823         DB_LOCKOBJ *op;
 824         DB_LOCKREGION *lrp;
 825         float lock_ratio, obj_ratio;
 826         size_t incr, oldsize, used;
 827         u_int32_t i, newlocks, newmem, newobjs;
 828         int ret, usedlocks, usedmem, usedobjs;
 829         u_int8_t *curaddr;
 830
 831         lrp = lt->region;
 832         oldsize = lrp->hdr.size;
 833         incr = lrp->increment;
 834
 835         /* Figure out how much of each sort of space we have. */
 836         usedmem = lrp->mem_bytes - __db_shalloc_count(lt->mem);
 837         usedobjs = lrp->numobjs - __lock_count_objs(lrp);
 838         usedlocks = lrp->maxlocks - __lock_count_locks(lrp);
 839
 840         /*
 841          * Figure out what fraction of the used space belongs to each
 842          * different type of "thing" in the region.  Then partition the
 843          * new space up according to this ratio.
 844          */
 845         used = usedmem +
 846             usedlocks * ALIGN(sizeof(struct __db_lock), MUTEX_ALIGNMENT) +
 847             usedobjs * sizeof(DB_LOCKOBJ);
 848
 849         lock_ratio = usedlocks *
 850             ALIGN(sizeof(struct __db_lock), MUTEX_ALIGNMENT) / (float)used;
 851         obj_ratio = usedobjs * sizeof(DB_LOCKOBJ) / (float)used;
 852
 853         newlocks = (u_int32_t)(lock_ratio *
 854             incr / ALIGN(sizeof(struct __db_lock), MUTEX_ALIGNMENT));
 855         newobjs = (u_int32_t)(obj_ratio * incr / sizeof(DB_LOCKOBJ));
 856         newmem = incr -
 857             (newobjs * sizeof(DB_LOCKOBJ) +
 858             newlocks * ALIGN(sizeof(struct __db_lock), MUTEX_ALIGNMENT));
 859
 860         /*
 861          * Make sure we allocate enough memory for the object being
 862          * requested.
 863          */
 864         switch (which) {
 865                 case DB_LOCK_LOCK:
 866                         if (newlocks == 0) {
 867                                 newlocks = 10;
 868                                 incr += newlocks * sizeof(struct __db_lock);
 869                         }
 870                         break;
 871                 case DB_LOCK_OBJ:
 872                         if (newobjs == 0) {
 873                                 newobjs = 10;
 874                                 incr += newobjs * sizeof(DB_LOCKOBJ);
 875                         }
 876                         break;
 877                 case DB_LOCK_MEM:
 878                         if (newmem < howmuch * 2) {
 879                                 incr += howmuch * 2 - newmem;
 880                                 newmem = howmuch * 2;
 881                         }
 882                         break;
 883         }
 884
 885         newmem += ALIGN(incr, sizeof(size_t)) - incr;
 886         incr = ALIGN(incr, sizeof(size_t));
 887
 888         /*
 889          * Since we are going to be allocating locks at the beginning of the
 890          * new chunk, we need to make sure that the chunk is MUTEX_ALIGNMENT
 891          * aligned.  We did not guarantee this when we created the region, so
 892          * we may need to pad the old region by extra bytes to ensure this
 893          * alignment.
 894          */
 895         incr += ALIGN(oldsize, MUTEX_ALIGNMENT) - oldsize;
 896
 897         __db_err(lt->dbenv,
 898             "Growing lock region: %lu locks %lu objs %lu bytes",
 899             (u_long)newlocks, (u_long)newobjs, (u_long)newmem);
 900
 901         if ((ret = __db_rgrow(lt->dbenv, lt->fd, incr)) != 0)
 902                 return (ret);
 903         if ((ret = __db_rremap(lt->dbenv,
 904             lt->region, oldsize, oldsize + incr, lt->fd, &lt->region)) != 0)
 905                 return (ret);
 906         __lock_reset_region(lt);
 907
 908         /* Update region parameters. */
 909         lrp = lt->region;
 910         lrp->increment = incr << 1;
 911         lrp->maxlocks += newlocks;
 912         lrp->numobjs += newobjs;
 913         lrp->mem_bytes += newmem;
 914
 915         curaddr = (u_int8_t *)lrp + oldsize;
 916         curaddr = (u_int8_t *)ALIGNP(curaddr, MUTEX_ALIGNMENT);
 917
 918         /* Put new locks onto the free list. */
 919         lock_head = &lrp->free_locks;
 920         for (i = 0; i++ < newlocks;
 921             curaddr += ALIGN(sizeof(struct __db_lock), MUTEX_ALIGNMENT)) {
 922                 newl = (struct __db_lock *)curaddr;
 923                 SH_TAILQ_INSERT_HEAD(lock_head, newl, links, __db_lock);
 924         }
 925
 926         /* Put new objects onto the free list.  */
 927         obj_head = &lrp->free_objs;
 928         for (i = 0; i++ < newobjs; curaddr += sizeof(DB_LOCKOBJ)) {
 929                 op = (DB_LOCKOBJ *)curaddr;
 930                 SH_TAILQ_INSERT_HEAD(obj_head, op, links, __db_lockobj);
 931         }
 932
 933         *((size_t *)curaddr) = newmem - sizeof(size_t);
 934         curaddr += sizeof(size_t);
 935         __db_shalloc_free(lt->mem, curaddr);
 936
 937         return (0);
 938 }
 939
 940 #ifdef DEBUG
 941 void
 942 __lock_dump_region(lt, flags)
 943         DB_LOCKTAB *lt;
 944         unsigned long flags;
 945 {
 946         struct __db_lock *lp;
 947         DB_LOCKOBJ *op;
 948         DB_LOCKREGION *lrp;
 949         u_int32_t i, j;
 950
 951         lrp = lt->region;
 952
 953         printf("Lock region parameters\n");
 954         printf("%s:0x%x\t%s:%lu\t%s:%lu\t%s:%lu\n%s:%lu\t%s:%lu\t%s:%lu\t\n",
 955             "magic      ", lrp->magic,
 956             "version    ", (u_long)lrp->version,
 957             "processes  ", (u_long)lrp->hdr.refcnt,
 958             "maxlocks   ", (u_long)lrp->maxlocks,
 959             "table size ", (u_long)lrp->table_size,
 960             "nmodes     ", (u_long)lrp->nmodes,
 961             "numobjs    ", (u_long)lrp->numobjs);
 962         printf("%s:%lu\t%s:%lu\t%s:%lu\n%s:%lu\t%s:%lu\t%s:%lu\n",
 963             "size       ", (u_long)lrp->hdr.size,
 964             "nlockers   ", (u_long)lrp->nlockers,
 965             "hash_off   ", (u_long)lrp->hash_off,
 966             "increment  ", (u_long)lrp->increment,
 967             "mem_off    ", (u_long)lrp->mem_off,
 968             "mem_bytes  ", (u_long)lrp->mem_bytes);
 969 #ifndef HAVE_SPINLOCKS
 970         printf("Mutex: off %lu", (u_long)lrp->hdr.lock.off);
 971 #endif
 972 #ifdef MUTEX_STATISTICS
 973         printf(" waits %lu nowaits %lu",
 974             (u_long)lrp->hdr.lock.mutex_set_wait,
 975             (u_long)lrp->hdr.lock.mutex_set_nowait);
 976 #endif
 977         printf("\n%s:%lu\t%s:%lu\t%s:%lu\t%s:%lu\n",
 978             "nconflicts ", (u_long)lrp->nconflicts,
 979             "nrequests  ", (u_long)lrp->nrequests,
 980             "nreleases  ", (u_long)lrp->nreleases,
 981             "ndeadlocks ", (u_long)lrp->ndeadlocks);
 982         printf("need_dd    %lu\n", (u_long)lrp->need_dd);
 983         if (flags & LOCK_DEBUG_CONF) {
 984                 printf("\nConflict matrix\n");
 985
 986                 for (i = 0; i < lrp->nmodes; i++) {
 987                         for (j = 0; j < lrp->nmodes; j++)
 988                                 printf("%lu\t",
 989                                     (u_long)lt->conflicts[i * lrp->nmodes + j]);
 990                         printf("\n");
 991                 }
 992         }
 993
 994         for (i = 0; i < lrp->table_size; i++) {
 995                 op = SH_TAILQ_FIRST(&lt->hashtab[i], __db_lockobj);
 996                 if (op != NULL && flags & LOCK_DEBUG_BUCKET)
 997                         printf("Bucket %lu:\n", (unsigned long)i);
 998                 while (op != NULL) {
 999                         if (op->type == DB_LOCK_LOCKER &&
1000                             flags & LOCK_DEBUG_LOCKERS)
1001                                 __lock_dump_locker(lt, op);
1002                         else if (flags & LOCK_DEBUG_OBJECTS &&
1003                             op->type == DB_LOCK_OBJTYPE)
1004                                 __lock_dump_object(lt, op);
1005                         op = SH_TAILQ_NEXT(op, links, __db_lockobj);
1006                 }
1007         }
1008
1009         if (flags & LOCK_DEBUG_LOCK) {
1010                 printf("\nLock Free List\n");
1011                 for (lp = SH_TAILQ_FIRST(&lrp->free_locks, __db_lock);
1012                     lp != NULL;
1013                     lp = SH_TAILQ_NEXT(lp, links, __db_lock)) {
1014                         printf("0x%x: %lu\t%lu\t%lu\t0x%x\n", (u_int)lp,
1015                             (u_long)lp->holder, (u_long)lp->mode,
1016                             (u_long)lp->status, (u_int)lp->obj);
1017                 }
1018         }
1019
1020         if (flags & LOCK_DEBUG_LOCK) {
1021                 printf("\nObject Free List\n");
1022                 for (op = SH_TAILQ_FIRST(&lrp->free_objs, __db_lockobj);
1023                     op != NULL;
1024                     op = SH_TAILQ_NEXT(op, links, __db_lockobj))
1025                         printf("0x%x\n", (u_int)op);
1026         }
1027
1028         if (flags & LOCK_DEBUG_MEM) {
1029                 printf("\nMemory Free List\n");
1030                 __db_shalloc_dump(stdout, lt->mem);
1031         }
1032 }
1033
1034 static void
1035 __lock_dump_locker(lt, op)
1036         DB_LOCKTAB *lt;
1037         DB_LOCKOBJ *op;
1038 {
1039         struct __db_lock *lp;
1040         u_int32_t locker;
1041         void *ptr;
1042
1043         ptr = SH_DBT_PTR(&op->lockobj);
1044         memcpy(&locker, ptr, sizeof(u_int32_t));
1045         printf("L %lu", (u_long)locker);
1046
1047         lp = SH_LIST_FIRST(&op->heldby, __db_lock);
1048         if (lp == NULL) {
1049                 printf("\n");
1050                 return;
1051         }
1052         for (; lp != NULL; lp = SH_LIST_NEXT(lp, locker_links, __db_lock))
1053                 __lock_printlock(lt, lp, 0);
1054 }
1055
1056 static void
1057 __lock_dump_object(lt, op)
1058         DB_LOCKTAB *lt;
1059         DB_LOCKOBJ *op;
1060 {
1061         struct __db_lock *lp;
1062         u_int32_t j;
1063         char *ptr;
1064
1065         ptr = SH_DBT_PTR(&op->lockobj);
1066         for (j = 0; j < op->lockobj.size; ptr++, j++)
1067                 printf("%c", (int)*ptr);
1068         printf("\n");
1069
1070         printf("H:");
1071         for (lp =
1072             SH_TAILQ_FIRST(&op->holders, __db_lock);
1073             lp != NULL;
1074             lp = SH_TAILQ_NEXT(lp, links, __db_lock))
1075                 __lock_printlock(lt, lp, 0);
1076         lp = SH_TAILQ_FIRST(&op->waiters, __db_lock);
1077         if (lp != NULL) {
1078                 printf("\nW:");
1079                 for (; lp != NULL; lp = SH_TAILQ_NEXT(lp, links, __db_lock))
1080                         __lock_printlock(lt, lp, 0);
1081         }
1082 }
1083
1084 int
1085 __lock_is_locked(lt, locker, dbt, mode)
1086         DB_LOCKTAB *lt;
1087         u_int32_t locker;
1088         DBT *dbt;
1089         db_lockmode_t mode;
1090 {
1091         struct __db_lock *lp;
1092         DB_LOCKOBJ *sh_obj;
1093         DB_LOCKREGION *lrp;
1094
1095         lrp = lt->region;
1096
1097         /* Look up the object in the hash table. */
1098         __db_hashlookup(lt->hashtab, __db_lockobj, links,
1099             dbt, sh_obj, lrp->table_size, __lock_ohash, __lock_cmp);
1100         if (sh_obj == NULL)
1101                 return (0);
1102
1103         for (lp = SH_TAILQ_FIRST(&sh_obj->holders, __db_lock);
1104             lp != NULL;
1105             lp = SH_TAILQ_FIRST(&sh_obj->holders, __db_lock)) {
1106                 if (lp->holder == locker && lp->mode == mode)
1107                         return (1);
1108         }
1109
1110         return (0);
1111 }
1112
1113 static void
1114 __lock_printlock(lt, lp, ispgno)
1115         DB_LOCKTAB *lt;
1116         struct __db_lock *lp;
1117         int ispgno;
1118 {
1119         DB_LOCKOBJ *lockobj;
1120         db_pgno_t pgno;
1121         size_t obj;
1122         u_int8_t *ptr;
1123         char *mode, *stat;
1124
1125         switch (lp->mode) {
1126         case DB_LOCK_IREAD:
1127                 mode = "IREAD";
1128                 break;
1129         case DB_LOCK_IWR:
1130                 mode = "IWR";
1131                 break;
1132         case DB_LOCK_IWRITE:
1133                 mode = "IWRITE";
1134                 break;
1135         case DB_LOCK_NG:
1136                 mode = "NG";
1137                 break;
1138         case DB_LOCK_READ:
1139                 mode = "READ";
1140                 break;
1141         case DB_LOCK_WRITE:
1142                 mode = "WRITE";
1143                 break;
1144         default:
1145                 mode = "UNKNOWN";
1146                 break;
1147         }
1148         switch (lp->status) {
1149         case DB_LSTAT_ABORTED:
1150                 stat = "ABORT";
1151                 break;
1152         case DB_LSTAT_ERR:
1153                 stat = "ERROR";
1154                 break;
1155         case DB_LSTAT_FREE:
1156                 stat = "FREE";
1157                 break;
1158         case DB_LSTAT_HELD:
1159                 stat = "HELD";
1160                 break;
1161         case DB_LSTAT_NOGRANT:
1162                 stat = "NONE";
1163                 break;
1164         case DB_LSTAT_WAITING:
1165                 stat = "WAIT";
1166                 break;
1167         case DB_LSTAT_PENDING:
1168                 stat = "PENDING";
1169                 break;
1170         default:
1171                 stat = "UNKNOWN";
1172                 break;
1173         }
1174         printf("\t%lu\t%s\t%lu\t%s\t",
1175             (u_long)lp->holder, mode, (u_long)lp->refcount, stat);
1176
1177         lockobj = (DB_LOCKOBJ *)((u_int8_t *)lp + lp->obj);
1178         ptr = SH_DBT_PTR(&lockobj->lockobj);
1179         if (ispgno) {
1180                 /* Assume this is a DBT lock. */
1181                 memcpy(&pgno, ptr, sizeof(db_pgno_t));
1182                 printf("page %lu\n", (u_long)pgno);
1183         } else {
1184                 obj = (u_int8_t *)lp + lp->obj - (u_int8_t *)lt->region;
1185                 printf("0x%lx ", (u_long)obj);
1186                 __db_pr(ptr, lockobj->lockobj.size);
1187                 printf("\n");
1188         }
1189 }
1190
1191 #endif
1192
1193 static int
1194 __lock_count_locks(lrp)
1195         DB_LOCKREGION *lrp;
1196 {
1197         struct __db_lock *newl;
1198         int count;
1199
1200         count = 0;
1201         for (newl = SH_TAILQ_FIRST(&lrp->free_locks, __db_lock);
1202             newl != NULL;
1203             newl = SH_TAILQ_NEXT(newl, links, __db_lock))
1204                 count++;
1205
1206         return (count);
1207 }
1208
1209 static int
1210 __lock_count_objs(lrp)
1211         DB_LOCKREGION *lrp;
1212 {
1213         DB_LOCKOBJ *obj;
1214         int count;
1215
1216         count = 0;
1217         for (obj = SH_TAILQ_FIRST(&lrp->free_objs, __db_lockobj);
1218             obj != NULL;
1219             obj = SH_TAILQ_NEXT(obj, links, __db_lockobj))
1220                 count++;
1221
1222         return (count);
1223 }
1224
1225 /*
1226  * PUBLIC: int __lock_getobj  __P((DB_LOCKTAB *,
1227  * PUBLIC:     u_int32_t, DBT *, u_int32_t type, DB_LOCKOBJ **));
1228  */
1229 int
1230 __lock_getobj(lt, locker, dbt, type, objp)
1231         DB_LOCKTAB *lt;
1232         u_int32_t locker, type;
1233         DBT *dbt;
1234         DB_LOCKOBJ **objp;
1235 {
1236         DB_LOCKREGION *lrp;
1237         DB_LOCKOBJ *sh_obj;
1238         u_int32_t obj_size;
1239         int ret;
1240         void *p, *src;
1241
1242         lrp = lt->region;
1243
1244         /* Look up the object in the hash table. */
1245         if (type == DB_LOCK_OBJTYPE) {
1246                 __db_hashlookup(lt->hashtab, __db_lockobj, links, dbt, sh_obj,
1247                     lrp->table_size, __lock_ohash, __lock_cmp);
1248                 obj_size = dbt->size;
1249         } else {
1250                 __db_hashlookup(lt->hashtab, __db_lockobj, links, locker,
1251                     sh_obj, lrp->table_size, __lock_locker_hash,
1252                     __lock_locker_cmp);
1253                 obj_size = sizeof(locker);
1254         }
1255
1256         /*
1257          * If we found the object, then we can just return it.  If
1258          * we didn't find the object, then we need to create it.
1259          */
1260         if (sh_obj == NULL) {
1261                 /* Create new object and then insert it into hash table. */
1262                 if ((sh_obj = SH_TAILQ_FIRST(&lrp->free_objs, __db_lockobj))
1263                     == NULL) {
1264                         if ((ret = __lock_grow_region(lt, DB_LOCK_OBJ, 0)) != 0)
1265                                 return (ret);
1266                         lrp = lt->region;
1267                         sh_obj = SH_TAILQ_FIRST(&lrp->free_objs, __db_lockobj);
1268                 }
1269                 if ((ret = __db_shalloc(lt->mem, obj_size, 0, &p)) != 0) {
1270                         if ((ret = __lock_grow_region(lt,
1271                             DB_LOCK_MEM, obj_size)) != 0)
1272                                 return (ret);
1273                         lrp = lt->region;
1274                         /* Reacquire the head of the list. */
1275                         sh_obj = SH_TAILQ_FIRST(&lrp->free_objs, __db_lockobj);
1276                         (void)__db_shalloc(lt->mem, obj_size, 0, &p);
1277                 }
1278                 sh_obj->type = type;
1279                 src = type == DB_LOCK_OBJTYPE ? dbt->data : (void *)&locker;
1280                 memcpy(p, src, obj_size);
1281                 SH_TAILQ_REMOVE(&lrp->free_objs, sh_obj, links, __db_lockobj);
1282
1283                 SH_TAILQ_INIT(&sh_obj->waiters);
1284                 if (type == DB_LOCK_LOCKER)
1285                         SH_LIST_INIT(&sh_obj->heldby);
1286                 else
1287                         SH_TAILQ_INIT(&sh_obj->holders);
1288                 sh_obj->lockobj.size = obj_size;
1289                 sh_obj->lockobj.off = SH_PTR_TO_OFF(&sh_obj->lockobj, p);
1290
1291                 __db_hashinsert(lt->hashtab, __db_lockobj, links, sh_obj,
1292                     lrp->table_size, __lock_lhash);
1293
1294                 if (type == DB_LOCK_LOCKER)
1295                         lrp->nlockers++;
1296         }
1297
1298         *objp = sh_obj;
1299         return (0);
1300 }
1301
1302 /*
1303  * Any lock on the waitlist has a process waiting for it.  Therefore, we
1304  * can't return the lock to the freelist immediately.  Instead, we can
1305  * remove the lock from the list of waiters, set the status field of the
1306  * lock, and then let the process waking up return the lock to the
1307  * free list.
1308  */
1309 static void
1310 __lock_remove_waiter(lt, sh_obj, lockp, status)
1311         DB_LOCKTAB *lt;
1312         DB_LOCKOBJ *sh_obj;
1313         struct __db_lock *lockp;
1314         db_status_t status;
1315 {
1316         SH_TAILQ_REMOVE(&sh_obj->waiters, lockp, links, __db_lock);
1317         lockp->status = status;
1318
1319         /* Wake whoever is waiting on this lock. */
1320         (void)__db_mutex_unlock(&lockp->mutex, lt->fd);
1321 }
1322
1323 static void
1324 __lock_freeobj(lt, obj)
1325         DB_LOCKTAB *lt;
1326         DB_LOCKOBJ *obj;
1327 {
1328         __db_hashremove_el(lt->hashtab, __db_lockobj, links,
1329             obj, lt->region->table_size, __lock_lhash);
1330         __db_shalloc_free(lt->mem, SH_DBT_PTR(&obj->lockobj));
1331         SH_TAILQ_INSERT_HEAD(&lt->region->free_objs, obj, links, __db_lockobj);
1332 }
1333
1334 static void
1335 __lock_checklocker(lt, lockp, do_remove)
1336         DB_LOCKTAB *lt;
1337         struct __db_lock *lockp;
1338         int do_remove;
1339 {
1340         DB_LOCKOBJ *sh_locker;
1341
1342         if (do_remove)
1343                 SH_LIST_REMOVE(lockp, locker_links, __db_lock);
1344
1345         /* if the locker list is NULL, free up the object. */
1346         if (__lock_getobj(lt, lockp->holder, NULL, DB_LOCK_LOCKER, &sh_locker)
1347             == 0 && SH_LIST_FIRST(&sh_locker->heldby, __db_lock) == NULL) {
1348                 __lock_freeobj(lt, sh_locker);
1349                 lt->region->nlockers--;
1350         }
1351 }
1352
1353 static void
1354 __lock_reset_region(lt)
1355         DB_LOCKTAB *lt;
1356 {
1357         lt->conflicts = (u_int8_t *)lt->region + sizeof(DB_LOCKREGION);
1358         lt->hashtab =
1359             (DB_HASHTAB *)((u_int8_t *)lt->region + lt->region->hash_off);
1360         lt->mem = (void *)((u_int8_t *)lt->region + lt->region->mem_off);
1361         lt->reg_size = lt->region->hdr.size;
1362 }