contrib/bind9/lib/dns/acache.c

   1 /*
   2  * Copyright (C) 2004-2006  Internet Systems Consortium, Inc. ("ISC")
   3  *
   4  * Permission to use, copy, modify, and distribute this software for any
   5  * purpose with or without fee is hereby granted, provided that the above
   6  * copyright notice and this permission notice appear in all copies.
   7  *
   8  * THE SOFTWARE IS PROVIDED "AS IS" AND ISC DISCLAIMS ALL WARRANTIES WITH
   9  * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
  10  * AND FITNESS.  IN NO EVENT SHALL ISC BE LIABLE FOR ANY SPECIAL, DIRECT,
  11  * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
  12  * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE
  13  * OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
  14  * PERFORMANCE OF THIS SOFTWARE.
  15  */
  16
  17 /* $Id: acache.c,v 1.3.2.16 2006/07/19 00:34:56 marka Exp $ */
  18
  19 #include <config.h>
  20
  21 #include <isc/atomic.h>
  22 #include <isc/event.h>
  23 #include <isc/hash.h>
  24 #include <isc/magic.h>
  25 #include <isc/mem.h>
  26 #include <isc/mutex.h>
  27 #include <isc/random.h>
  28 #include <isc/refcount.h>
  29 #include <isc/rwlock.h>
  30 #include <isc/task.h>
  31 #include <isc/time.h>
  32 #include <isc/timer.h>
  33
  34 #include <dns/acache.h>
  35 #include <dns/db.h>
  36 #include <dns/events.h>
  37 #include <dns/log.h>
  38 #include <dns/message.h>
  39 #include <dns/name.h>
  40 #include <dns/rdataset.h>
  41 #include <dns/result.h>
  42 #include <dns/zone.h>
  43
  44 #define ACACHE_MAGIC                    ISC_MAGIC('A', 'C', 'H', 'E')
  45 #define DNS_ACACHE_VALID(acache)        ISC_MAGIC_VALID(acache, ACACHE_MAGIC)
  46
  47 #define ACACHEENTRY_MAGIC               ISC_MAGIC('A', 'C', 'E', 'T')
  48 #define DNS_ACACHEENTRY_VALID(entry)    ISC_MAGIC_VALID(entry, ACACHEENTRY_MAGIC)
  49
  50 #define DBBUCKETS       67
  51
  52 #if 0
  53 #define ATRACE(m)       isc_log_write(dns_lctx, \
  54                                       DNS_LOGCATEGORY_DATABASE, \
  55                                       DNS_LOGMODULE_ACACHE, \
  56                                       ISC_LOG_DEBUG(3), \
  57                                       "acache %p: %s", acache, (m))
  58 #define AATRACE(a,m)    isc_log_write(dns_lctx, \
  59                                       DNS_LOGCATEGORY_DATABASE, \
  60                                       DNS_LOGMODULE_ACACHE, \
  61                                       ISC_LOG_DEBUG(3), \
  62                                       "acache %p: %s", (a), (m))
  63 #else
  64 #define ATRACE(m)
  65 #define AATRACE(a, m)
  66 #endif
  67
  68 /*
  69  * The following variables control incremental cleaning.
  70  * MINSIZE is how many bytes is the floor for dns_acache_setcachesize().
  71  * CLEANERINCREMENT is how many entries are examined in one pass.
  72  * (XXX simply derived from definitions in cache.c  There may be better
  73  *  constants here.)
  74  */
  75 #define DNS_ACACHE_MINSIZE              2097152 /* Bytes.  2097152 = 2 MB */
  76 #define DNS_ACACHE_CLEANERINCREMENT     1000    /* Number of entries. */
  77
  78 #define DEFAULT_ACACHE_ENTRY_LOCK_COUNT 1009    /*%< Should be prime. */
  79
  80 #if defined(ISC_RWLOCK_USEATOMIC) && defined(ISC_PLATFORM_HAVEATOMICSTORE)
  81 #define ACACHE_USE_RWLOCK 1
  82 #endif
  83
  84 #ifdef ACACHE_USE_RWLOCK
  85 #define ACACHE_INITLOCK(l)      isc_rwlock_init((l), 0, 0)
  86 #define ACACHE_DESTROYLOCK(l)   isc_rwlock_destroy(l)
  87 #define ACACHE_LOCK(l, t)       RWLOCK((l), (t))
  88 #define ACACHE_UNLOCK(l, t)     RWUNLOCK((l), (t))
  89
  90 #define acache_storetime(entry, t) \
  91         (isc_atomic_store((isc_int32_t *)&(entry)->lastused, (t)))
  92 #else
  93 #define ACACHE_INITLOCK(l)      isc_mutex_init(l)
  94 #define ACACHE_DESTROYLOCK(l)   DESTROYLOCK(l)
  95 #define ACACHE_LOCK(l, t)       LOCK(l)
  96 #define ACACHE_UNLOCK(l, t)     UNLOCK(l)
  97
  98 #define acache_storetime(entry, t) ((entry)->lastused = (t))
  99 #endif
 100
 101 /* Locked by acache lock */
 102 typedef struct dbentry {
 103         ISC_LINK(struct dbentry)        link;
 104
 105         dns_db_t                        *db;
 106         ISC_LIST(dns_acacheentry_t)     originlist;
 107         ISC_LIST(dns_acacheentry_t)     referlist;
 108 } dbentry_t;
 109
 110 typedef ISC_LIST(dbentry_t) dbentrylist_t;
 111
 112 typedef struct acache_cleaner acache_cleaner_t;
 113
 114 typedef enum {
 115         cleaner_s_idle, /* Waiting for cleaning-interval to expire. */
 116         cleaner_s_busy, /* Currently cleaning. */
 117         cleaner_s_done  /* Freed enough memory after being overmem. */
 118 } cleaner_state_t;
 119
 120 /*
 121  * Convenience macros for comprehensive assertion checking.
 122  */
 123 #define CLEANER_IDLE(c) ((c)->state == cleaner_s_idle && \
 124                          (c)->resched_event != NULL)
 125 #define CLEANER_BUSY(c) ((c)->state == cleaner_s_busy && \
 126                          (c)->resched_event == NULL)
 127
 128 struct acache_cleaner {
 129         isc_mutex_t             lock;
 130         /*
 131          * Locks overmem_event, overmem.  (See cache.c)
 132          */
 133
 134         dns_acache_t            *acache;
 135         unsigned int            cleaning_interval; /* The cleaning-interval
 136                                                       from named.conf,
 137                                                       in seconds. */
 138
 139         isc_stdtime_t           last_cleanup_time; /* The time when the last
 140                                                       cleanup task completed */
 141
 142         isc_timer_t             *cleaning_timer;
 143         isc_event_t             *resched_event; /* Sent by cleaner task to
 144                                                    itself to reschedule */
 145         isc_event_t             *overmem_event;
 146
 147         dns_acacheentry_t       *current_entry; /* The bookmark entry to
 148                                                    restart the cleaning.
 149                                                    Locked by acache lock. */
 150         int                     increment;      /* Number of entries to
 151                                                    clean in one increment */
 152
 153         unsigned long           ncleaned;       /* Number of entries cleaned
 154                                                    up (for logging purposes) */
 155         cleaner_state_t         state;          /* Idle/Busy/Done. */
 156         isc_boolean_t           overmem;        /* The acache is in an overmem
 157                                                    state. */
 158 };
 159
 160 struct dns_acachestats {
 161         unsigned int                    hits;
 162         unsigned int                    queries;
 163         unsigned int                    misses;
 164         unsigned int                    adds;
 165         unsigned int                    deleted;
 166         unsigned int                    cleaned;
 167         unsigned int                    cleaner_runs;
 168         unsigned int                    overmem;
 169         unsigned int                    overmem_nocreates;
 170         unsigned int                    nomem;
 171 };
 172
 173 /*
 174  * The actual acache object.
 175  */
 176
 177 struct dns_acache {
 178         unsigned int                    magic;
 179
 180         isc_mem_t                       *mctx;
 181         isc_refcount_t                  refs;
 182
 183 #ifdef ACACHE_USE_RWLOCK
 184         isc_rwlock_t                    *entrylocks;
 185 #else
 186         isc_mutex_t                     *entrylocks;
 187 #endif
 188
 189         isc_mutex_t                     lock;
 190
 191         int                             live_cleaners;
 192         acache_cleaner_t                cleaner;
 193         ISC_LIST(dns_acacheentry_t)     entries;
 194         unsigned int                    dbentries;
 195         dbentrylist_t                   dbbucket[DBBUCKETS];
 196
 197         isc_boolean_t                   shutting_down;
 198
 199         isc_task_t                      *task;
 200         isc_event_t                     cevent;
 201         isc_boolean_t                   cevent_sent;
 202
 203         dns_acachestats_t               stats;
 204 };
 205
 206 struct dns_acacheentry {
 207         unsigned int            magic;
 208
 209         unsigned int            locknum;
 210         isc_refcount_t          references;
 211
 212         dns_acache_t            *acache;
 213
 214         /* Data for Management of cache entries */
 215         ISC_LINK(dns_acacheentry_t) link;
 216         ISC_LINK(dns_acacheentry_t) olink;
 217         ISC_LINK(dns_acacheentry_t) rlink;
 218
 219         dns_db_t                *origdb; /* reference to the DB
 220                                             holding this entry */
 221
 222         /* Cache data */
 223         dns_zone_t              *zone;          /* zone this entry
 224                                                    belongs to */
 225         dns_db_t                *db;            /* DB this entry belongs to */
 226         dns_dbversion_t         *version;       /* the version of the DB */
 227         dns_dbnode_t            *node;          /* node this entry
 228                                                    belongs to */
 229         dns_name_t              *foundname;     /* corresponding DNS name
 230                                                    and rdataset */
 231
 232         /* Callback function and its argument */
 233         void                    (*callback)(dns_acacheentry_t *, void **);
 234         void                    *cbarg;
 235
 236         /* Timestamp of the last time this entry is referred to */
 237         isc_stdtime32_t         lastused;
 238 };
 239
 240 /*
 241  *      Internal functions (and prototypes).
 242  */
 243 static inline isc_boolean_t check_noentry(dns_acache_t *acache);
 244 static void destroy(dns_acache_t *acache);
 245 static void shutdown_entries(dns_acache_t *acache);
 246 static void shutdown_buckets(dns_acache_t *acache);
 247 static void destroy_entry(dns_acacheentry_t *ent);
 248 static inline void unlink_dbentries(dns_acache_t *acache,
 249                                     dns_acacheentry_t *ent);
 250 static inline isc_result_t finddbent(dns_acache_t *acache,
 251                                      dns_db_t *db, dbentry_t **dbentryp);
 252 static inline void clear_entry(dns_acache_t *acache, dns_acacheentry_t *entry);
 253 static isc_result_t acache_cleaner_init(dns_acache_t *acache,
 254                                         isc_timermgr_t *timermgr,
 255                                         acache_cleaner_t *cleaner);
 256 static void acache_cleaning_timer_action(isc_task_t *task, isc_event_t *event);
 257 static void acache_incremental_cleaning_action(isc_task_t *task,
 258                                                isc_event_t *event);
 259 static void acache_overmem_cleaning_action(isc_task_t *task,
 260                                            isc_event_t *event);
 261 static void acache_cleaner_shutdown_action(isc_task_t *task,
 262                                            isc_event_t *event);
 263
 264 /*
 265  * acache should be locked.  If it is not, the stats can get out of whack,
 266  * which is not a big deal for us since this is for debugging / stats
 267  */
 268 static void
 269 reset_stats(dns_acache_t *acache) {
 270         acache->stats.hits = 0;
 271         acache->stats.queries = 0;
 272         acache->stats.misses = 0;
 273         acache->stats.adds = 0;
 274         acache->stats.deleted = 0;
 275         acache->stats.cleaned = 0;
 276         acache->stats.overmem = 0;
 277         acache->stats.overmem_nocreates = 0;
 278         acache->stats.nomem = 0;
 279 }
 280
 281 /*
 282  * The acache must be locked before calling.
 283  */
 284 static inline isc_boolean_t
 285 check_noentry(dns_acache_t *acache) {
 286         if (ISC_LIST_EMPTY(acache->entries) && acache->dbentries == 0) {
 287                 return (ISC_TRUE);
 288         }
 289
 290         return (ISC_FALSE);
 291 }
 292
 293 /*
 294  * The acache must be locked before calling.
 295  */
 296 static void
 297 shutdown_entries(dns_acache_t *acache) {
 298         dns_acacheentry_t *entry, *entry_next;
 299
 300         REQUIRE(DNS_ACACHE_VALID(acache));
 301         INSIST(acache->shutting_down);
 302
 303         /*
 304          * Release the dependency of all entries, and detach them.
 305          */
 306         for (entry = ISC_LIST_HEAD(acache->entries);
 307              entry != NULL;
 308              entry = entry_next) {
 309                 entry_next = ISC_LIST_NEXT(entry, link);
 310
 311                 ACACHE_LOCK(&acache->entrylocks[entry->locknum],
 312                             isc_rwlocktype_write);
 313
 314                 /*
 315                  * If the cleaner holds this entry, it will be unlinked and
 316                  * freed in the cleaner later.
 317                  */
 318                 if (acache->cleaner.current_entry != entry)
 319                         ISC_LIST_UNLINK(acache->entries, entry, link);
 320                 unlink_dbentries(acache, entry);
 321                 if (entry->callback != NULL) {
 322                         (entry->callback)(entry, &entry->cbarg);
 323                         entry->callback = NULL;
 324                 }
 325
 326                 ACACHE_UNLOCK(&acache->entrylocks[entry->locknum],
 327                               isc_rwlocktype_write);
 328
 329                 if (acache->cleaner.current_entry != entry)
 330                         dns_acache_detachentry(&entry);
 331         }
 332 }
 333
 334 /*
 335  * The acache must be locked before calling.
 336  */
 337 static void
 338 shutdown_buckets(dns_acache_t *acache) {
 339         int i;
 340         dbentry_t *dbent;
 341
 342         REQUIRE(DNS_ACACHE_VALID(acache));
 343         INSIST(acache->shutting_down);
 344
 345         for (i = 0; i < DBBUCKETS; i++) {
 346                 while ((dbent = ISC_LIST_HEAD(acache->dbbucket[i])) != NULL) {
 347                         INSIST(ISC_LIST_EMPTY(dbent->originlist) &&
 348                                ISC_LIST_EMPTY(dbent->referlist));
 349                         ISC_LIST_UNLINK(acache->dbbucket[i], dbent, link);
 350
 351                         dns_db_detach(&dbent->db);
 352
 353                         isc_mem_put(acache->mctx, dbent, sizeof(*dbent));
 354
 355                         acache->dbentries--;
 356                 }
 357         }
 358
 359         INSIST(acache->dbentries == 0);
 360 }
 361
 362 static void
 363 shutdown_task(isc_task_t *task, isc_event_t *ev) {
 364         dns_acache_t *acache;
 365
 366         UNUSED(task);
 367
 368         acache = ev->ev_arg;
 369         INSIST(DNS_ACACHE_VALID(acache));
 370
 371         isc_event_free(&ev);
 372
 373         LOCK(&acache->lock);
 374
 375         shutdown_entries(acache);
 376         shutdown_buckets(acache);
 377
 378         UNLOCK(&acache->lock);
 379
 380         dns_acache_detach(&acache);
 381 }
 382
 383 /* The acache and the entry must be locked before calling. */
 384 static inline void
 385 unlink_dbentries(dns_acache_t *acache, dns_acacheentry_t *ent) {
 386         isc_result_t result;
 387         dbentry_t *dbent;
 388
 389         if (ISC_LINK_LINKED(ent, olink)) {
 390                 INSIST(ent->origdb != NULL);
 391                 dbent = NULL;
 392                 result = finddbent(acache, ent->origdb, &dbent);
 393                 INSIST(result == ISC_R_SUCCESS);
 394
 395                 ISC_LIST_UNLINK(dbent->originlist, ent, olink);
 396         }
 397         if (ISC_LINK_LINKED(ent, rlink)) {
 398                 INSIST(ent->db != NULL);
 399                 dbent = NULL;
 400                 result = finddbent(acache, ent->db, &dbent);
 401                 INSIST(result == ISC_R_SUCCESS);
 402
 403                 ISC_LIST_UNLINK(dbent->referlist, ent, rlink);
 404         }
 405 }
 406
 407 /* There must not be a reference to this entry. */
 408 static void
 409 destroy_entry(dns_acacheentry_t *entry) {
 410         dns_acache_t *acache;
 411
 412         REQUIRE(DNS_ACACHEENTRY_VALID(entry));
 413
 414         acache = entry->acache;
 415         REQUIRE(DNS_ACACHE_VALID(acache));
 416
 417         /*
 418          * Since there is no reference to this entry, it is safe to call
 419          * clear_entry() here.
 420          */
 421         clear_entry(acache, entry);
 422
 423         isc_mem_put(acache->mctx, entry, sizeof(*entry));
 424
 425         dns_acache_detach(&acache);
 426 }
 427
 428 static void
 429 destroy(dns_acache_t *acache) {
 430         int i;
 431
 432         REQUIRE(DNS_ACACHE_VALID(acache));
 433
 434         ATRACE("destroy");
 435
 436         isc_mem_setwater(acache->mctx, NULL, NULL, 0, 0);
 437
 438         if (acache->cleaner.overmem_event != NULL)
 439                 isc_event_free(&acache->cleaner.overmem_event);
 440
 441         if (acache->cleaner.resched_event != NULL)
 442                 isc_event_free(&acache->cleaner.resched_event);
 443
 444         if (acache->task != NULL)
 445                 isc_task_detach(&acache->task);
 446
 447         for (i = 0; i < DEFAULT_ACACHE_ENTRY_LOCK_COUNT; i++)
 448                 ACACHE_DESTROYLOCK(&acache->entrylocks[i]);
 449         isc_mem_put(acache->mctx, acache->entrylocks,
 450                     sizeof(*acache->entrylocks) *
 451                     DEFAULT_ACACHE_ENTRY_LOCK_COUNT);
 452
 453         DESTROYLOCK(&acache->cleaner.lock);
 454
 455         DESTROYLOCK(&acache->lock);
 456         acache->magic = 0;
 457
 458         isc_mem_putanddetach(&acache->mctx, acache, sizeof(*acache));
 459 }
 460
 461 static inline isc_result_t
 462 finddbent(dns_acache_t *acache, dns_db_t *db, dbentry_t **dbentryp) {
 463         int bucket;
 464         dbentry_t *dbentry;
 465
 466         REQUIRE(DNS_ACACHE_VALID(acache));
 467         REQUIRE(db != NULL);
 468         REQUIRE(dbentryp != NULL && *dbentryp == NULL);
 469
 470         /*
 471          * The caller must be holding the acache lock.
 472          */
 473
 474         bucket = isc_hash_calc((const unsigned char *)&db,
 475                                sizeof(db), ISC_TRUE) % DBBUCKETS;
 476
 477         for (dbentry = ISC_LIST_HEAD(acache->dbbucket[bucket]);
 478              dbentry != NULL;
 479              dbentry = ISC_LIST_NEXT(dbentry, link)) {
 480                 if (dbentry->db == db)
 481                         break;
 482         }
 483
 484         *dbentryp = dbentry;
 485
 486         if (dbentry == NULL)
 487                 return (ISC_R_NOTFOUND);
 488         else
 489                 return (ISC_R_SUCCESS);
 490 }
 491
 492 static inline void
 493 clear_entry(dns_acache_t *acache, dns_acacheentry_t *entry) {
 494         REQUIRE(DNS_ACACHE_VALID(acache));
 495         REQUIRE(DNS_ACACHEENTRY_VALID(entry));
 496
 497         /*
 498          * The caller must be holing the entry lock.
 499          */
 500
 501         if (entry->foundname) {
 502                 dns_rdataset_t *rdataset, *rdataset_next;
 503
 504                 for (rdataset = ISC_LIST_HEAD(entry->foundname->list);
 505                      rdataset != NULL;
 506                      rdataset = rdataset_next) {
 507                         rdataset_next = ISC_LIST_NEXT(rdataset, link);
 508                         ISC_LIST_UNLINK(entry->foundname->list,
 509                                         rdataset, link);
 510                         dns_rdataset_disassociate(rdataset);
 511                         isc_mem_put(acache->mctx, rdataset, sizeof(*rdataset));
 512                 }
 513                 if (dns_name_dynamic(entry->foundname))
 514                         dns_name_free(entry->foundname, acache->mctx);
 515                 isc_mem_put(acache->mctx, entry->foundname,
 516                             sizeof(*entry->foundname));
 517                 entry->foundname = NULL;
 518         }
 519
 520         if (entry->node != NULL) {
 521                 INSIST(entry->db != NULL);
 522                 dns_db_detachnode(entry->db, &entry->node);
 523         }
 524         if (entry->version != NULL) {
 525                 INSIST(entry->db != NULL);
 526                 dns_db_closeversion(entry->db, &entry->version, ISC_FALSE);
 527         }
 528         if (entry->db != NULL)
 529                 dns_db_detach(&entry->db);
 530         if (entry->zone != NULL)
 531                 dns_zone_detach(&entry->zone);
 532
 533         if (entry->origdb != NULL)
 534                 dns_db_detach(&entry->origdb);
 535 }
 536
 537 static isc_result_t
 538 acache_cleaner_init(dns_acache_t *acache, isc_timermgr_t *timermgr,
 539                     acache_cleaner_t *cleaner)
 540 {
 541         int result;
 542
 543         ATRACE("acache cleaner init");
 544
 545         result = isc_mutex_init(&cleaner->lock);
 546         if (result != ISC_R_SUCCESS)
 547                 goto fail;
 548
 549         cleaner->increment = DNS_ACACHE_CLEANERINCREMENT;
 550         cleaner->state = cleaner_s_idle;
 551         cleaner->acache = acache;
 552         cleaner->overmem = ISC_FALSE;
 553
 554         cleaner->cleaning_timer = NULL;
 555         cleaner->resched_event = NULL;
 556         cleaner->overmem_event = NULL;
 557         cleaner->current_entry = NULL;
 558
 559         if (timermgr != NULL) {
 560                 cleaner->acache->live_cleaners++;
 561
 562                 result = isc_task_onshutdown(acache->task,
 563                                              acache_cleaner_shutdown_action,
 564                                              acache);
 565                 if (result != ISC_R_SUCCESS) {
 566                         UNEXPECTED_ERROR(__FILE__, __LINE__,
 567                                          "acache cleaner: "
 568                                          "isc_task_onshutdown() failed: %s",
 569                                          dns_result_totext(result));
 570                         goto cleanup;
 571                 }
 572
 573                 cleaner->cleaning_interval = 0; /* Initially turned off. */
 574                 isc_stdtime_get(&cleaner->last_cleanup_time);
 575                 result = isc_timer_create(timermgr, isc_timertype_inactive,
 576                                           NULL, NULL,
 577                                           acache->task,
 578                                           acache_cleaning_timer_action,
 579                                           cleaner, &cleaner->cleaning_timer);
 580                 if (result != ISC_R_SUCCESS) {
 581                         UNEXPECTED_ERROR(__FILE__, __LINE__,
 582                                          "isc_timer_create() failed: %s",
 583                                          dns_result_totext(result));
 584                         result = ISC_R_UNEXPECTED;
 585                         goto cleanup;
 586                 }
 587
 588                 cleaner->resched_event =
 589                         isc_event_allocate(acache->mctx, cleaner,
 590                                            DNS_EVENT_ACACHECLEAN,
 591                                            acache_incremental_cleaning_action,
 592                                            cleaner, sizeof(isc_event_t));
 593                 if (cleaner->resched_event == NULL) {
 594                         result = ISC_R_NOMEMORY;
 595                         goto cleanup;
 596                 }
 597
 598                 cleaner->overmem_event =
 599                         isc_event_allocate(acache->mctx, cleaner,
 600                                            DNS_EVENT_ACACHEOVERMEM,
 601                                            acache_overmem_cleaning_action,
 602                                            cleaner, sizeof(isc_event_t));
 603                 if (cleaner->overmem_event == NULL) {
 604                         result = ISC_R_NOMEMORY;
 605                         goto cleanup;
 606                 }
 607         }
 608
 609         return (ISC_R_SUCCESS);
 610
 611  cleanup:
 612         if (cleaner->overmem_event != NULL)
 613                 isc_event_free(&cleaner->overmem_event);
 614         if (cleaner->resched_event != NULL)
 615                 isc_event_free(&cleaner->resched_event);
 616         if (cleaner->cleaning_timer != NULL)
 617                 isc_timer_detach(&cleaner->cleaning_timer);
 618         cleaner->acache->live_cleaners--;
 619         DESTROYLOCK(&cleaner->lock);
 620  fail:
 621         return (result);
 622 }
 623
 624 static void
 625 begin_cleaning(acache_cleaner_t *cleaner) {
 626         dns_acacheentry_t *head;
 627         dns_acache_t *acache = cleaner->acache;
 628
 629         /*
 630          * This function does not have to lock the cleaner, since critical
 631          * parameters (except current_entry, which is locked by acache lock,)
 632          * are only used in a single task context.
 633          */
 634
 635         REQUIRE(CLEANER_IDLE(cleaner));
 636         INSIST(DNS_ACACHE_VALID(acache));
 637         INSIST(cleaner->current_entry == NULL);
 638
 639         isc_log_write(dns_lctx, DNS_LOGCATEGORY_DATABASE,
 640                       DNS_LOGMODULE_ACACHE, ISC_LOG_DEBUG(1),
 641                       "begin acache cleaning, mem inuse %lu",
 642                       (unsigned long)isc_mem_inuse(cleaner->acache->mctx));
 643
 644         LOCK(&acache->lock);
 645
 646         head = ISC_LIST_HEAD(acache->entries);
 647         if (head != NULL)
 648                 dns_acache_attachentry(head, &cleaner->current_entry);
 649
 650         UNLOCK(&acache->lock);
 651
 652         if (cleaner->current_entry != NULL) {
 653                 cleaner->ncleaned = 0;
 654                 cleaner->state = cleaner_s_busy;
 655                 isc_task_send(acache->task, &cleaner->resched_event);
 656         }
 657
 658         return;
 659 }
 660
 661 static void
 662 end_cleaning(acache_cleaner_t *cleaner, isc_event_t *event) {
 663         dns_acache_t *acache = cleaner->acache;
 664
 665         REQUIRE(CLEANER_BUSY(cleaner));
 666         REQUIRE(event != NULL);
 667         REQUIRE(DNS_ACACHEENTRY_VALID(cleaner->current_entry));
 668
 669         /* No need to lock the cleaner (see begin_cleaning()). */
 670
 671         LOCK(&acache->lock);
 672
 673         /*
 674          * Even if the cleaner has the last reference to the entry, which means
 675          * the entry has been unused, it may still be linked if unlinking the
 676          * entry has been delayed due to the reference.
 677          */
 678         if (isc_refcount_current(&cleaner->current_entry->references) == 1) {
 679                 INSIST(cleaner->current_entry->callback == NULL);
 680
 681                 if (ISC_LINK_LINKED(cleaner->current_entry, link)) {
 682                         ISC_LIST_UNLINK(acache->entries,
 683                                         cleaner->current_entry, link);
 684                 }
 685         }
 686         dns_acache_detachentry(&cleaner->current_entry);
 687
 688         if (cleaner->overmem)
 689                 acache->stats.overmem++;
 690         acache->stats.cleaned += cleaner->ncleaned;
 691         acache->stats.cleaner_runs++;
 692
 693         isc_log_write(dns_lctx, DNS_LOGCATEGORY_DATABASE, DNS_LOGMODULE_ACACHE,
 694                       ISC_LOG_NOTICE,
 695                       "acache %p stats: hits=%d misses=%d queries=%d "
 696                       "adds=%d deleted=%d "
 697                       "cleaned=%d cleaner_runs=%d overmem=%d "
 698                       "overmem_nocreates=%d nomem=%d",
 699                       acache,
 700                       acache->stats.hits, acache->stats.misses,
 701                       acache->stats.queries,
 702                       acache->stats.adds, acache->stats.deleted,
 703                       acache->stats.cleaned, acache->stats.cleaner_runs,
 704                       acache->stats.overmem, acache->stats.overmem_nocreates,
 705                       acache->stats.nomem);
 706         reset_stats(acache);
 707
 708         isc_stdtime_get(&cleaner->last_cleanup_time);
 709
 710         UNLOCK(&acache->lock);
 711
 712         dns_acache_setcleaninginterval(cleaner->acache,
 713                                        cleaner->cleaning_interval);
 714
 715         isc_log_write(dns_lctx, DNS_LOGCATEGORY_DATABASE, DNS_LOGMODULE_ACACHE,
 716                       ISC_LOG_DEBUG(1), "end acache cleaning, "
 717                       "%lu entries cleaned, mem inuse %lu",
 718                       cleaner->ncleaned,
 719                       (unsigned long)isc_mem_inuse(cleaner->acache->mctx));
 720
 721         if (cleaner->overmem) {
 722                 isc_log_write(dns_lctx, DNS_LOGCATEGORY_DATABASE,
 723                               DNS_LOGMODULE_ACACHE, ISC_LOG_NOTICE,
 724                               "acache is still in overmem state "
 725                               "after cleaning");
 726         }
 727
 728         cleaner->ncleaned = 0;
 729         cleaner->state = cleaner_s_idle;
 730         cleaner->resched_event = event;
 731 }
 732
 733 /*
 734  * This is run once for every acache-cleaning-interval as defined
 735  * in named.conf.
 736  */
 737 static void
 738 acache_cleaning_timer_action(isc_task_t *task, isc_event_t *event) {
 739         acache_cleaner_t *cleaner = event->ev_arg;
 740
 741         UNUSED(task);
 742
 743         INSIST(event->ev_type == ISC_TIMEREVENT_TICK);
 744
 745         isc_log_write(dns_lctx, DNS_LOGCATEGORY_DATABASE, DNS_LOGMODULE_ACACHE,
 746                       ISC_LOG_DEBUG(1), "acache cleaning timer fired, "
 747                       "cleaner state = %d", cleaner->state);
 748
 749         if (cleaner->state == cleaner_s_idle)
 750                 begin_cleaning(cleaner);
 751
 752         isc_event_free(&event);
 753 }
 754
 755 /* The caller must hold entry lock. */
 756 static inline isc_boolean_t
 757 entry_stale(acache_cleaner_t *cleaner, dns_acacheentry_t *entry,
 758             isc_stdtime32_t now32, unsigned int interval)
 759 {
 760         /*
 761          * If the callback has been canceled, we definitely do not need the
 762          * entry.
 763          */
 764         if (entry->callback == NULL)
 765                 return (ISC_TRUE);
 766
 767         if (interval > cleaner->cleaning_interval)
 768                 interval = cleaner->cleaning_interval;
 769
 770         if (entry->lastused + interval < now32)
 771                 return (ISC_TRUE);
 772
 773         /*
 774          * If the acache is in the overmem state, probabilistically decide if
 775          * the entry should be purged, based on the time passed from its last
 776          * use and the cleaning interval.
 777          */
 778         if (cleaner->overmem) {
 779                 unsigned int passed =
 780                         now32 - entry->lastused; /* <= interval */
 781                 isc_uint32_t val;
 782
 783                 if (passed > interval / 2)
 784                         return (ISC_TRUE);
 785                 isc_random_get(&val);
 786                 if (passed > interval / 4)
 787                         return (ISC_TF(val % 4 == 0));
 788                 return (ISC_TF(val % 8 == 0));
 789         }
 790
 791         return (ISC_FALSE);
 792 }
 793
 794 /*
 795  * Do incremental cleaning.
 796  */
 797 static void
 798 acache_incremental_cleaning_action(isc_task_t *task, isc_event_t *event) {
 799         acache_cleaner_t *cleaner = event->ev_arg;
 800         dns_acache_t *acache = cleaner->acache;
 801         dns_acacheentry_t *entry, *next = NULL;
 802         int n_entries;
 803         isc_stdtime32_t now32, last32;
 804         isc_stdtime_t now;
 805         unsigned int interval;
 806
 807         INSIST(DNS_ACACHE_VALID(acache));
 808         INSIST(task == acache->task);
 809         INSIST(event->ev_type == DNS_EVENT_ACACHECLEAN);
 810
 811         if (cleaner->state == cleaner_s_done) {
 812                 cleaner->state = cleaner_s_busy;
 813                 end_cleaning(cleaner, event);
 814                 return;
 815         }
 816
 817         INSIST(CLEANER_BUSY(cleaner));
 818
 819         n_entries = cleaner->increment;
 820
 821         isc_stdtime_get(&now);
 822         isc_stdtime_convert32(now, &now32);
 823
 824         LOCK(&acache->lock);
 825
 826         entry = cleaner->current_entry;
 827         isc_stdtime_convert32(cleaner->last_cleanup_time, &last32);
 828         INSIST(now32 > last32);
 829         interval = now32 - last32;
 830
 831         while (n_entries-- > 0) {
 832                 isc_boolean_t is_stale = ISC_FALSE;
 833
 834                 INSIST(entry != NULL);
 835
 836                 next = ISC_LIST_NEXT(entry, link);
 837
 838                 ACACHE_LOCK(&acache->entrylocks[entry->locknum],
 839                             isc_rwlocktype_write);
 840
 841                 is_stale = entry_stale(cleaner, entry, now32, interval);
 842                 if (is_stale) {
 843                         ISC_LIST_UNLINK(acache->entries, entry, link);
 844                         unlink_dbentries(acache, entry);
 845                         if (entry->callback != NULL)
 846                                 (entry->callback)(entry, &entry->cbarg);
 847                         entry->callback = NULL;
 848
 849                         cleaner->ncleaned++;
 850                 }
 851
 852                 ACACHE_UNLOCK(&acache->entrylocks[entry->locknum],
 853                               isc_rwlocktype_write);
 854
 855                 if (is_stale)
 856                         dns_acache_detachentry(&entry);
 857
 858                 if (next == NULL) {
 859                         if (cleaner->overmem) {
 860                                 entry = ISC_LIST_HEAD(acache->entries);
 861                                 if (entry != NULL) {
 862                                         /*
 863                                          * If we are still in the overmem
 864                                          * state, keep cleaning.
 865                                          */
 866                                         isc_log_write(dns_lctx,
 867                                                       DNS_LOGCATEGORY_DATABASE,
 868                                                       DNS_LOGMODULE_ACACHE,
 869                                                       ISC_LOG_DEBUG(1),
 870                                                       "acache cleaner: "
 871                                                       "still overmem, "
 872                                                       "reset and try again");
 873                                         continue;
 874                                 }
 875                         }
 876
 877                         UNLOCK(&acache->lock);
 878                         end_cleaning(cleaner, event);
 879                         return;
 880                 }
 881
 882                 entry = next;
 883         }
 884
 885         /*
 886          * We have successfully performed a cleaning increment but have
 887          * not gone through the entire cache.  Remember the entry that will
 888          * be the starting point in the next clean-up, and reschedule another
 889          * batch.  If it fails, just try to continue anyway.
 890          */
 891         INSIST(next != NULL && next != cleaner->current_entry);
 892         dns_acache_detachentry(&cleaner->current_entry);
 893         dns_acache_attachentry(next, &cleaner->current_entry);
 894
 895         UNLOCK(&acache->lock);
 896
 897         isc_log_write(dns_lctx, DNS_LOGCATEGORY_DATABASE, DNS_LOGMODULE_ACACHE,
 898                       ISC_LOG_DEBUG(1), "acache cleaner: checked %d entries, "
 899                       "mem inuse %lu, sleeping", cleaner->increment,
 900                       (unsigned long)isc_mem_inuse(cleaner->acache->mctx));
 901
 902         isc_task_send(task, &event);
 903         INSIST(CLEANER_BUSY(cleaner));
 904
 905         return;
 906 }
 907
 908 /*
 909  * This is called when the acache either surpasses its upper limit
 910  * or shrinks beyond its lower limit.
 911  */
 912 static void
 913 acache_overmem_cleaning_action(isc_task_t *task, isc_event_t *event) {
 914         acache_cleaner_t *cleaner = event->ev_arg;
 915         isc_boolean_t want_cleaning = ISC_FALSE;
 916
 917         UNUSED(task);
 918
 919         INSIST(event->ev_type == DNS_EVENT_ACACHEOVERMEM);
 920         INSIST(cleaner->overmem_event == NULL);
 921
 922         isc_log_write(dns_lctx, DNS_LOGCATEGORY_DATABASE, DNS_LOGMODULE_ACACHE,
 923                       ISC_LOG_DEBUG(1), "overmem_cleaning_action called, "
 924                       "overmem = %d, state = %d", cleaner->overmem,
 925                       cleaner->state);
 926
 927         LOCK(&cleaner->lock);
 928
 929         if (cleaner->overmem) {
 930                 if (cleaner->state == cleaner_s_idle)
 931                         want_cleaning = ISC_TRUE;
 932         } else {
 933                 if (cleaner->state == cleaner_s_busy)
 934                         /*
 935                          * end_cleaning() can't be called here because
 936                          * then both cleaner->overmem_event and
 937                          * cleaner->resched_event will point to this
 938                          * event.  Set the state to done, and then
 939                          * when the acache_incremental_cleaning_action() event
 940                          * is posted, it will handle the end_cleaning.
 941                          */
 942                         cleaner->state = cleaner_s_done;
 943         }
 944
 945         cleaner->overmem_event = event;
 946
 947         UNLOCK(&cleaner->lock);
 948
 949         if (want_cleaning)
 950                 begin_cleaning(cleaner);
 951 }
 952
 953 static void
 954 water(void *arg, int mark) {
 955         dns_acache_t *acache = arg;
 956         isc_boolean_t overmem = ISC_TF(mark == ISC_MEM_HIWATER);
 957
 958         REQUIRE(DNS_ACACHE_VALID(acache));
 959
 960         isc_log_write(dns_lctx, DNS_LOGCATEGORY_DATABASE,
 961                       DNS_LOGMODULE_ACACHE, ISC_LOG_DEBUG(1),
 962                       "acache memory reaches %s watermark, mem inuse %lu",
 963                       overmem ? "high" : "low",
 964                       (unsigned long)isc_mem_inuse(acache->mctx));
 965
 966         LOCK(&acache->cleaner.lock);
 967
 968         acache->cleaner.overmem = overmem;
 969
 970         if (acache->cleaner.overmem_event != NULL)
 971                 isc_task_send(acache->task, &acache->cleaner.overmem_event);
 972
 973         UNLOCK(&acache->cleaner.lock);
 974 }
 975
 976 /*
 977  * The cleaner task is shutting down; do the necessary cleanup.
 978  */
 979 static void
 980 acache_cleaner_shutdown_action(isc_task_t *task, isc_event_t *event) {
 981         dns_acache_t *acache = event->ev_arg;
 982         isc_boolean_t should_free = ISC_FALSE;
 983
 984         INSIST(task == acache->task);
 985         INSIST(event->ev_type == ISC_TASKEVENT_SHUTDOWN);
 986         INSIST(DNS_ACACHE_VALID(acache));
 987
 988         ATRACE("acache cleaner shutdown");
 989
 990         if (CLEANER_BUSY(&acache->cleaner))
 991                 end_cleaning(&acache->cleaner, event);
 992         else
 993                 isc_event_free(&event);
 994
 995         LOCK(&acache->lock);
 996
 997         acache->live_cleaners--;
 998         INSIST(acache->live_cleaners == 0);
 999
1000         if (isc_refcount_current(&acache->refs) == 0) {
1001                 INSIST(check_noentry(acache) == ISC_TRUE);
1002                 should_free = ISC_TRUE;
1003         }
1004
1005         /*
1006          * By detaching the timer in the context of its task,
1007          * we are guaranteed that there will be no further timer
1008          * events.
1009          */
1010         if (acache->cleaner.cleaning_timer != NULL)
1011                 isc_timer_detach(&acache->cleaner.cleaning_timer);
1012
1013         /* Make sure we don't reschedule anymore. */
1014         (void)isc_task_purge(task, NULL, DNS_EVENT_ACACHECLEAN, NULL);
1015
1016         UNLOCK(&acache->lock);
1017
1018         if (should_free)
1019                 destroy(acache);
1020 }
1021
1022 /*
1023  *      Public functions.
1024  */
1025
1026 isc_result_t
1027 dns_acache_create(dns_acache_t **acachep, isc_mem_t *mctx,
1028                   isc_taskmgr_t *taskmgr, isc_timermgr_t *timermgr)
1029 {
1030         int i;
1031         isc_result_t result;
1032         dns_acache_t *acache;
1033
1034         REQUIRE(acachep != NULL && *acachep == NULL);
1035         REQUIRE(mctx != NULL);
1036         REQUIRE(taskmgr != NULL);
1037
1038         acache = isc_mem_get(mctx, sizeof(*acache));
1039         if (acache == NULL)
1040                 return (ISC_R_NOMEMORY);
1041
1042         ATRACE("create");
1043
1044         result = isc_refcount_init(&acache->refs, 1);
1045         if (result != ISC_R_SUCCESS) {
1046                 isc_mem_put(mctx, acache, sizeof(*acache));
1047                 return (result);
1048         }
1049
1050         result = isc_mutex_init(&acache->lock);
1051         if (result != ISC_R_SUCCESS) {
1052                 isc_refcount_decrement(&acache->refs, NULL);
1053                 isc_refcount_destroy(&acache->refs);
1054                 isc_mem_put(mctx, acache, sizeof(*acache));
1055                 return (result);
1056         }
1057
1058         acache->mctx = NULL;
1059         isc_mem_attach(mctx, &acache->mctx);
1060         ISC_LIST_INIT(acache->entries);
1061
1062         acache->shutting_down = ISC_FALSE;
1063
1064         acache->task = NULL;
1065         acache->entrylocks = NULL;
1066
1067         result = isc_task_create(taskmgr, 1, &acache->task);
1068         if (result != ISC_R_SUCCESS) {
1069                 UNEXPECTED_ERROR(__FILE__, __LINE__,
1070                                  "isc_task_create() failed(): %s",
1071                                  dns_result_totext(result));
1072                 result = ISC_R_UNEXPECTED;
1073                 goto cleanup;
1074         }
1075         isc_task_setname(acache->task, "acachetask", acache);
1076         ISC_EVENT_INIT(&acache->cevent, sizeof(acache->cevent), 0, NULL,
1077                        DNS_EVENT_ACACHECONTROL, shutdown_task, NULL,
1078                        NULL, NULL, NULL);
1079         acache->cevent_sent = ISC_FALSE;
1080
1081         acache->dbentries = 0;
1082         for (i = 0; i < DBBUCKETS; i++)
1083                 ISC_LIST_INIT(acache->dbbucket[i]);
1084
1085         acache->entrylocks = isc_mem_get(mctx, sizeof(*acache->entrylocks) *
1086                                          DEFAULT_ACACHE_ENTRY_LOCK_COUNT);
1087         if (acache->entrylocks == NULL) {
1088                 result = ISC_R_NOMEMORY;
1089                 goto cleanup;
1090         }
1091         for (i = 0; i < DEFAULT_ACACHE_ENTRY_LOCK_COUNT; i++) {
1092                 result = ACACHE_INITLOCK(&acache->entrylocks[i]);
1093                 if (result != ISC_R_SUCCESS) {
1094                         while (i-- > 0)
1095                                 ACACHE_DESTROYLOCK(&acache->entrylocks[i]);
1096                         isc_mem_put(mctx, acache->entrylocks,
1097                                     sizeof(*acache->entrylocks) *
1098                                     DEFAULT_ACACHE_ENTRY_LOCK_COUNT);
1099                         acache->entrylocks = NULL;
1100                         goto cleanup;
1101                 }
1102         }
1103
1104         acache->live_cleaners = 0;
1105         result = acache_cleaner_init(acache, timermgr, &acache->cleaner);
1106         if (result != ISC_R_SUCCESS)
1107                 goto cleanup;
1108
1109         acache->stats.cleaner_runs = 0;
1110         reset_stats(acache);
1111
1112         acache->magic = ACACHE_MAGIC;
1113
1114         *acachep = acache;
1115         return (ISC_R_SUCCESS);
1116
1117  cleanup:
1118         if (acache->task != NULL)
1119                 isc_task_detach(&acache->task);
1120         DESTROYLOCK(&acache->lock);
1121         isc_refcount_decrement(&acache->refs, NULL);
1122         isc_refcount_destroy(&acache->refs);
1123         if (acache->entrylocks != NULL) {
1124                 for (i = 0; i < DEFAULT_ACACHE_ENTRY_LOCK_COUNT; i++)
1125                         ACACHE_DESTROYLOCK(&acache->entrylocks[i]);
1126                 isc_mem_put(mctx, acache->entrylocks,
1127                             sizeof(*acache->entrylocks) *
1128                             DEFAULT_ACACHE_ENTRY_LOCK_COUNT);
1129         }
1130         isc_mem_put(mctx, acache, sizeof(*acache));
1131         isc_mem_detach(&mctx);
1132
1133         return (result);
1134 }
1135
1136 void
1137 dns_acache_attach(dns_acache_t *source, dns_acache_t **targetp) {
1138         REQUIRE(DNS_ACACHE_VALID(source));
1139         REQUIRE(targetp != NULL && *targetp == NULL);
1140
1141         AATRACE(source, "attach");
1142
1143         isc_refcount_increment(&source->refs, NULL);
1144
1145         *targetp = source;
1146 }
1147
1148 void
1149 dns_acache_countquerymiss(dns_acache_t *acache) {
1150         acache->stats.misses++;         /* XXXSK danger: unlocked! */
1151         acache->stats.queries++;        /* XXXSK danger: unlocked! */
1152 }
1153
1154 void
1155 dns_acache_detach(dns_acache_t **acachep) {
1156         dns_acache_t *acache;
1157         unsigned int refs;
1158         isc_boolean_t should_free = ISC_FALSE;
1159
1160         REQUIRE(acachep != NULL && DNS_ACACHE_VALID(*acachep));
1161         acache = *acachep;
1162
1163         ATRACE("detach");
1164
1165         isc_refcount_decrement(&acache->refs, &refs);
1166         if (refs == 0) {
1167                 INSIST(check_noentry(acache) == ISC_TRUE);
1168                 should_free = ISC_TRUE;
1169         }
1170
1171         *acachep = NULL;
1172
1173         /*
1174          * If we're exiting and the cleaner task exists, let it free the cache.
1175          */
1176         if (should_free && acache->live_cleaners > 0) {
1177                 isc_task_shutdown(acache->task);
1178                 should_free = ISC_FALSE;
1179         }
1180
1181         if (should_free)
1182                 destroy(acache);
1183 }
1184
1185 void
1186 dns_acache_shutdown(dns_acache_t *acache) {
1187         REQUIRE(DNS_ACACHE_VALID(acache));
1188
1189         LOCK(&acache->lock);
1190
1191         ATRACE("shutdown");
1192
1193         if (!acache->shutting_down) {
1194                 isc_event_t *event;
1195                 dns_acache_t *acache_evarg = NULL;
1196
1197                 INSIST(!acache->cevent_sent);
1198
1199                 acache->shutting_down = ISC_TRUE;
1200
1201                 isc_mem_setwater(acache->mctx, NULL, NULL, 0, 0);
1202
1203                 /*
1204                  * Self attach the object in order to prevent it from being
1205                  * destroyed while waiting for the event.
1206                  */
1207                 dns_acache_attach(acache, &acache_evarg);
1208                 event = &acache->cevent;
1209                 event->ev_arg = acache_evarg;
1210                 isc_task_send(acache->task, &event);
1211                 acache->cevent_sent = ISC_TRUE;
1212         }
1213
1214         UNLOCK(&acache->lock);
1215 }
1216
1217 isc_result_t
1218 dns_acache_setdb(dns_acache_t *acache, dns_db_t *db) {
1219         int bucket;
1220         dbentry_t *dbentry;
1221         isc_result_t result = ISC_R_SUCCESS;
1222
1223         REQUIRE(DNS_ACACHE_VALID(acache));
1224         REQUIRE(db != NULL);
1225
1226         ATRACE("setdb");
1227
1228         LOCK(&acache->lock);
1229
1230         dbentry = NULL;
1231         result = finddbent(acache, db, &dbentry);
1232         if (result == ISC_R_SUCCESS) {
1233                 result = ISC_R_EXISTS;
1234                 goto end;
1235         }
1236         result = ISC_R_SUCCESS;
1237
1238         dbentry = isc_mem_get(acache->mctx, sizeof(*dbentry));
1239         if (dbentry == NULL) {
1240                 result = ISC_R_NOMEMORY;
1241                 goto end;
1242         }
1243
1244         ISC_LINK_INIT(dbentry, link);
1245         ISC_LIST_INIT(dbentry->originlist);
1246         ISC_LIST_INIT(dbentry->referlist);
1247
1248         dbentry->db = NULL;
1249         dns_db_attach(db, &dbentry->db);
1250
1251         bucket = isc_hash_calc((const unsigned char *)&db,
1252                                sizeof(db), ISC_TRUE) % DBBUCKETS;
1253
1254         ISC_LIST_APPEND(acache->dbbucket[bucket], dbentry, link);
1255
1256         acache->dbentries++;
1257
1258  end:
1259         UNLOCK(&acache->lock);
1260
1261         return (result);
1262 }
1263
1264 isc_result_t
1265 dns_acache_putdb(dns_acache_t *acache, dns_db_t *db) {
1266         int bucket;
1267         isc_result_t result;
1268         dbentry_t *dbentry;
1269         dns_acacheentry_t *entry;
1270
1271         REQUIRE(DNS_ACACHE_VALID(acache));
1272         REQUIRE(db != NULL);
1273
1274         ATRACE("putdb");
1275
1276         LOCK(&acache->lock);
1277
1278         dbentry = NULL;
1279         result = finddbent(acache, db, &dbentry);
1280         if (result != ISC_R_SUCCESS) {
1281                 /*
1282                  * The entry may have not been created due to memory shortage.
1283                  */
1284                 UNLOCK(&acache->lock);
1285                 return (ISC_R_NOTFOUND);
1286         }
1287
1288         /*
1289          * Release corresponding cache entries: for each entry, release all
1290          * links the entry has, and then callback to the entry holder (if any).
1291          * If no other external references exist (this can happen if the
1292          * original holder has canceled callback,) destroy it here.
1293          */
1294         while ((entry = ISC_LIST_HEAD(dbentry->originlist)) != NULL) {
1295                 ACACHE_LOCK(&acache->entrylocks[entry->locknum],
1296                             isc_rwlocktype_write);
1297
1298                 /*
1299                  * Releasing olink first would avoid finddbent() in
1300                  * unlink_dbentries().
1301                  */
1302                 ISC_LIST_UNLINK(dbentry->originlist, entry, olink);
1303                 if (acache->cleaner.current_entry != entry)
1304                         ISC_LIST_UNLINK(acache->entries, entry, link);
1305                 unlink_dbentries(acache, entry);
1306
1307                 if (entry->callback != NULL)
1308                         (entry->callback)(entry, &entry->cbarg);
1309                 entry->callback = NULL;
1310
1311                 ACACHE_UNLOCK(&acache->entrylocks[entry->locknum],
1312                               isc_rwlocktype_write);
1313
1314                 if (acache->cleaner.current_entry != entry)
1315                         dns_acache_detachentry(&entry);
1316         }
1317         while ((entry = ISC_LIST_HEAD(dbentry->referlist)) != NULL) {
1318                 ACACHE_LOCK(&acache->entrylocks[entry->locknum],
1319                             isc_rwlocktype_write);
1320
1321                 ISC_LIST_UNLINK(dbentry->referlist, entry, rlink);
1322                 if (acache->cleaner.current_entry != entry)
1323                         ISC_LIST_UNLINK(acache->entries, entry, link);
1324                 unlink_dbentries(acache, entry);
1325
1326                 if (entry->callback != NULL)
1327                         (entry->callback)(entry, &entry->cbarg);
1328                 entry->callback = NULL;
1329
1330                 ACACHE_UNLOCK(&acache->entrylocks[entry->locknum],
1331                               isc_rwlocktype_write);
1332
1333                 if (acache->cleaner.current_entry != entry)
1334                         dns_acache_detachentry(&entry);
1335         }
1336
1337         INSIST(ISC_LIST_EMPTY(dbentry->originlist) &&
1338                ISC_LIST_EMPTY(dbentry->referlist));
1339
1340         bucket = isc_hash_calc((const unsigned char *)&db,
1341                                sizeof(db), ISC_TRUE) % DBBUCKETS;
1342         ISC_LIST_UNLINK(acache->dbbucket[bucket], dbentry, link);
1343         dns_db_detach(&dbentry->db);
1344
1345         isc_mem_put(acache->mctx, dbentry, sizeof(*dbentry));
1346
1347         acache->dbentries--;
1348
1349         acache->stats.deleted++;
1350
1351         UNLOCK(&acache->lock);
1352
1353         return (ISC_R_SUCCESS);
1354 }
1355
1356 isc_result_t
1357 dns_acache_createentry(dns_acache_t *acache, dns_db_t *origdb,
1358                        void (*callback)(dns_acacheentry_t *, void **),
1359                        void *cbarg, dns_acacheentry_t **entryp)
1360 {
1361         dns_acacheentry_t *newentry;
1362         isc_result_t result;
1363         isc_uint32_t r;
1364
1365         REQUIRE(DNS_ACACHE_VALID(acache));
1366         REQUIRE(entryp != NULL && *entryp == NULL);
1367         REQUIRE(origdb != NULL);
1368
1369         /*
1370          * Should we exceed our memory limit for some reason (for
1371          * example, if the cleaner does not run aggressively enough),
1372          * then we will not create additional entries.
1373          *
1374          * XXXSK: It might be better to lock the acache->cleaner->lock,
1375          * but locking may be an expensive bottleneck. If we misread
1376          * the value, we will occasionally refuse to create a few
1377          * cache entries, or create a few that we should not. I do not
1378          * expect this to happen often, and it will not have very bad
1379          * effects when it does. So no lock for now.
1380          */
1381         if (acache->cleaner.overmem) {
1382                 acache->stats.overmem_nocreates++; /* XXXSK danger: unlocked! */
1383                 return (ISC_R_NORESOURCES);
1384         }
1385
1386         newentry = isc_mem_get(acache->mctx, sizeof(*newentry));
1387         if (newentry == NULL) {
1388                 acache->stats.nomem++;  /* XXXMLG danger: unlocked! */
1389                 return (ISC_R_NOMEMORY);
1390         }
1391
1392         isc_random_get(&r);
1393         newentry->locknum = r % DEFAULT_ACACHE_ENTRY_LOCK_COUNT;
1394
1395         result = isc_refcount_init(&newentry->references, 1);
1396         if (result != ISC_R_SUCCESS) {
1397                 isc_mem_put(acache->mctx, newentry, sizeof(*newentry));
1398                 return (result);
1399         };
1400
1401         ISC_LINK_INIT(newentry, link);
1402         ISC_LINK_INIT(newentry, olink);
1403         ISC_LINK_INIT(newentry, rlink);
1404
1405         newentry->acache = NULL;
1406         dns_acache_attach(acache, &newentry->acache);
1407
1408         newentry->zone = NULL;
1409         newentry->db = NULL;
1410         newentry->version = NULL;
1411         newentry->node = NULL;
1412         newentry->foundname = NULL;
1413
1414         newentry->callback = callback;
1415         newentry->cbarg = cbarg;
1416         newentry->origdb = NULL;
1417         dns_db_attach(origdb, &newentry->origdb);
1418
1419         isc_stdtime_get(&newentry->lastused);
1420
1421         newentry->magic = ACACHEENTRY_MAGIC;
1422
1423         *entryp = newentry;
1424
1425         return (ISC_R_SUCCESS);
1426 }
1427
1428 isc_result_t
1429 dns_acache_getentry(dns_acacheentry_t *entry, dns_zone_t **zonep,
1430                     dns_db_t **dbp, dns_dbversion_t **versionp,
1431                     dns_dbnode_t **nodep, dns_name_t *fname,
1432                     dns_message_t *msg, isc_stdtime_t now)
1433 {
1434         isc_result_t result = ISC_R_SUCCESS;
1435         dns_rdataset_t *erdataset;
1436         isc_stdtime32_t now32;
1437         dns_acache_t *acache;
1438         int locknum;
1439
1440         REQUIRE(DNS_ACACHEENTRY_VALID(entry));
1441         REQUIRE(zonep == NULL || *zonep == NULL);
1442         REQUIRE(dbp != NULL && *dbp == NULL);
1443         REQUIRE(versionp != NULL && *versionp == NULL);
1444         REQUIRE(nodep != NULL && *nodep == NULL);
1445         REQUIRE(fname != NULL);
1446         REQUIRE(msg != NULL);
1447         acache = entry->acache;
1448         REQUIRE(DNS_ACACHE_VALID(acache));
1449
1450         locknum = entry->locknum;
1451         ACACHE_LOCK(&acache->entrylocks[locknum], isc_rwlocktype_read);
1452
1453         isc_stdtime_convert32(now, &now32);
1454         acache_storetime(entry, now32);
1455
1456         if (entry->zone != NULL && zonep != NULL)
1457                 dns_zone_attach(entry->zone, zonep);
1458
1459         if (entry->db == NULL) {
1460                 *dbp = NULL;
1461                 *versionp = NULL;
1462         } else {
1463                 dns_db_attach(entry->db, dbp);
1464                 dns_db_attachversion(entry->db, entry->version, versionp);
1465         }
1466         if (entry->node == NULL)
1467                 *nodep = NULL;
1468         else {
1469                 dns_db_attachnode(entry->db, entry->node, nodep);
1470
1471                 INSIST(entry->foundname != NULL);
1472                 dns_name_copy(entry->foundname, fname, NULL);
1473                 for (erdataset = ISC_LIST_HEAD(entry->foundname->list);
1474                      erdataset != NULL;
1475                      erdataset = ISC_LIST_NEXT(erdataset, link)) {
1476                         dns_rdataset_t *ardataset;
1477
1478                         ardataset = NULL;
1479                         result = dns_message_gettemprdataset(msg, &ardataset);
1480                         if (result != ISC_R_SUCCESS) {
1481                                 ACACHE_UNLOCK(&acache->entrylocks[locknum],
1482                                               isc_rwlocktype_read);
1483                                 goto fail;
1484                         }
1485
1486                         /*
1487                          * XXXJT: if we simply clone the rdataset, we'll get
1488                          * lost wrt cyclic ordering.  We'll need an additional
1489                          * trick to get the latest counter from the original
1490                          * header.
1491                          */
1492                         dns_rdataset_init(ardataset);
1493                         dns_rdataset_clone(erdataset, ardataset);
1494                         ISC_LIST_APPEND(fname->list, ardataset, link);
1495                 }
1496         }
1497
1498         entry->acache->stats.hits++; /* XXXMLG danger: unlocked! */
1499         entry->acache->stats.queries++;
1500
1501         ACACHE_UNLOCK(&acache->entrylocks[locknum], isc_rwlocktype_read);
1502
1503         return (result);
1504
1505   fail:
1506         while ((erdataset = ISC_LIST_HEAD(fname->list)) != NULL) {
1507                 ISC_LIST_UNLINK(fname->list, erdataset, link);
1508                 dns_rdataset_disassociate(erdataset);
1509                 dns_message_puttemprdataset(msg, &erdataset);
1510         }
1511         if (*nodep != NULL)
1512                 dns_db_detachnode(*dbp, nodep);
1513         if (*versionp != NULL)
1514                 dns_db_closeversion(*dbp, versionp, ISC_FALSE);
1515         if (*dbp != NULL)
1516                 dns_db_detach(dbp);
1517         if (zonep != NULL && *zonep != NULL)
1518                 dns_zone_detach(zonep);
1519
1520         return (result);
1521 }
1522
1523 isc_result_t
1524 dns_acache_setentry(dns_acache_t *acache, dns_acacheentry_t *entry,
1525                     dns_zone_t *zone, dns_db_t *db, dns_dbversion_t *version,
1526                     dns_dbnode_t *node, dns_name_t *fname)
1527 {
1528         isc_result_t result;
1529         dbentry_t *odbent;
1530         dbentry_t *rdbent = NULL;
1531         isc_boolean_t close_version = ISC_FALSE;
1532         dns_acacheentry_t *dummy_entry = NULL;
1533
1534         REQUIRE(DNS_ACACHE_VALID(acache));
1535         REQUIRE(DNS_ACACHEENTRY_VALID(entry));
1536
1537         LOCK(&acache->lock);    /* XXX: need to lock it here for ordering */
1538         ACACHE_LOCK(&acache->entrylocks[entry->locknum], isc_rwlocktype_write);
1539
1540         /* Set zone */
1541         if (zone != NULL)
1542                 dns_zone_attach(zone, &entry->zone);
1543         /* Set DB */
1544         if (db != NULL)
1545                 dns_db_attach(db, &entry->db);
1546         /*
1547          * Set DB version.  If the version is not given by the caller,
1548          * which is the case for glue or cache DBs, use the current version.
1549          */
1550         if (version == NULL) {
1551                 if (db != NULL) {
1552                         dns_db_currentversion(db, &version);
1553                         close_version = ISC_TRUE;
1554                 }
1555         }
1556         if (version != NULL) {
1557                 INSIST(db != NULL);
1558                 dns_db_attachversion(db, version, &entry->version);
1559         }
1560         if (close_version)
1561                 dns_db_closeversion(db, &version, ISC_FALSE);
1562         /* Set DB node. */
1563         if (node != NULL) {
1564                 INSIST(db != NULL);
1565                 dns_db_attachnode(db, node, &entry->node);
1566         }
1567
1568         /*
1569          * Set list of the corresponding rdatasets, if given.
1570          * To minimize the overhead and memory consumption, we'll do this for
1571          * positive cache only, in which case the DB node is non NULL.
1572          * We do not want to cache incomplete information, so give up the
1573          * entire entry when a memory shortage happen during the process.
1574          */
1575         if (node != NULL) {
1576                 dns_rdataset_t *ardataset, *crdataset;
1577
1578                 entry->foundname = isc_mem_get(acache->mctx,
1579                                                sizeof(*entry->foundname));
1580
1581                 if (entry->foundname == NULL) {
1582                         result = ISC_R_NOMEMORY;
1583                         goto fail;
1584                 }
1585                 dns_name_init(entry->foundname, NULL);
1586                 result = dns_name_dup(fname, acache->mctx,
1587                                       entry->foundname);
1588                 if (result != ISC_R_SUCCESS)
1589                         goto fail;
1590
1591                 for (ardataset = ISC_LIST_HEAD(fname->list);
1592                      ardataset != NULL;
1593                      ardataset = ISC_LIST_NEXT(ardataset, link)) {
1594                         crdataset = isc_mem_get(acache->mctx,
1595                                                 sizeof(*crdataset));
1596                         if (crdataset == NULL) {
1597                                 result = ISC_R_NOMEMORY;
1598                                 goto fail;
1599                         }
1600
1601                         dns_rdataset_init(crdataset);
1602                         dns_rdataset_clone(ardataset, crdataset);
1603                         ISC_LIST_APPEND(entry->foundname->list, crdataset,
1604                                         link);
1605                 }
1606         }
1607
1608         odbent = NULL;
1609         result = finddbent(acache, entry->origdb, &odbent);
1610         if (result != ISC_R_SUCCESS)
1611                 goto fail;
1612         if (db != NULL) {
1613                 rdbent = NULL;
1614                 result = finddbent(acache, db, &rdbent);
1615                 if (result != ISC_R_SUCCESS)
1616                         goto fail;
1617         }
1618
1619         ISC_LIST_APPEND(acache->entries, entry, link);
1620         ISC_LIST_APPEND(odbent->originlist, entry, olink);
1621         if (rdbent != NULL)
1622                 ISC_LIST_APPEND(rdbent->referlist, entry, rlink);
1623
1624         /*
1625          * The additional cache needs an implicit reference to entries in its
1626          * link.
1627          */
1628         dns_acache_attachentry(entry, &dummy_entry);
1629
1630         ACACHE_UNLOCK(&acache->entrylocks[entry->locknum],
1631                       isc_rwlocktype_write);
1632
1633         acache->stats.adds++;
1634         UNLOCK(&acache->lock);
1635
1636         return (ISC_R_SUCCESS);
1637
1638  fail:
1639         clear_entry(acache, entry);
1640
1641         ACACHE_UNLOCK(&acache->entrylocks[entry->locknum],
1642                       isc_rwlocktype_write);
1643         UNLOCK(&acache->lock);
1644
1645         return (result);
1646 }
1647
1648 void
1649 dns_acache_cancelentry(dns_acacheentry_t *entry) {
1650         dns_acache_t *acache = entry->acache;
1651
1652         REQUIRE(DNS_ACACHEENTRY_VALID(entry));
1653         INSIST(DNS_ACACHE_VALID(acache));
1654
1655         LOCK(&acache->lock);
1656         ACACHE_LOCK(&acache->entrylocks[entry->locknum], isc_rwlocktype_write);
1657
1658         /*
1659          * Release dependencies stored in this entry as much as possible.
1660          * The main link cannot be released, since the acache object has
1661          * a reference to this entry; the empty entry will be released in
1662          * the next cleaning action.
1663          */
1664         unlink_dbentries(acache, entry);
1665         clear_entry(entry->acache, entry);
1666
1667         entry->callback = NULL;
1668         entry->cbarg = NULL;
1669
1670         ACACHE_UNLOCK(&acache->entrylocks[entry->locknum],
1671                       isc_rwlocktype_write);
1672         UNLOCK(&acache->lock);
1673 }
1674
1675 void
1676 dns_acache_attachentry(dns_acacheentry_t *source,
1677                        dns_acacheentry_t **targetp)
1678 {
1679         REQUIRE(DNS_ACACHEENTRY_VALID(source));
1680         REQUIRE(targetp != NULL && *targetp == NULL);
1681
1682         isc_refcount_increment(&source->references, NULL);
1683
1684         *targetp = source;
1685 }
1686
1687 void
1688 dns_acache_detachentry(dns_acacheentry_t **entryp) {
1689         dns_acacheentry_t *entry;
1690         unsigned int refs;
1691
1692         REQUIRE(entryp != NULL && DNS_ACACHEENTRY_VALID(*entryp));
1693         entry = *entryp;
1694
1695         isc_refcount_decrement(&entry->references, &refs);
1696
1697         /*
1698          * If there are no references to the entry, the entry must have been
1699          * unlinked and can be destroyed safely.
1700          */
1701         if (refs == 0) {
1702                 INSIST(!ISC_LINK_LINKED(entry, link));
1703                 (*entryp)->acache->stats.deleted++;
1704                 destroy_entry(entry);
1705         }
1706
1707         *entryp = NULL;
1708 }
1709
1710 void
1711 dns_acache_setcleaninginterval(dns_acache_t *acache, unsigned int t) {
1712         isc_interval_t interval;
1713         isc_result_t result;
1714
1715         REQUIRE(DNS_ACACHE_VALID(acache));
1716
1717         ATRACE("dns_acache_setcleaninginterval");
1718
1719         LOCK(&acache->lock);
1720
1721         /*
1722          * It may be the case that the acache has already shut down.
1723          * If so, it has no timer.  (Not sure if this can really happen.)
1724          */
1725         if (acache->cleaner.cleaning_timer == NULL)
1726                 goto unlock;
1727
1728         acache->cleaner.cleaning_interval = t;
1729
1730         if (t == 0) {
1731                 result = isc_timer_reset(acache->cleaner.cleaning_timer,
1732                                          isc_timertype_inactive,
1733                                          NULL, NULL, ISC_TRUE);
1734         } else {
1735                 isc_interval_set(&interval, acache->cleaner.cleaning_interval,
1736                                  0);
1737                 result = isc_timer_reset(acache->cleaner.cleaning_timer,
1738                                          isc_timertype_ticker,
1739                                          NULL, &interval, ISC_FALSE);
1740         }
1741         if (result != ISC_R_SUCCESS)
1742                 isc_log_write(dns_lctx, DNS_LOGCATEGORY_DATABASE,
1743                               DNS_LOGMODULE_ACACHE, ISC_LOG_WARNING,
1744                               "could not set acache cleaning interval: %s",
1745                               isc_result_totext(result));
1746         else
1747                 isc_log_write(dns_lctx, DNS_LOGCATEGORY_DATABASE,
1748                               DNS_LOGMODULE_ACACHE, ISC_LOG_NOTICE,
1749                               "acache %p cleaning interval set to %d.",
1750                               acache, t);
1751
1752  unlock:
1753         UNLOCK(&acache->lock);
1754 }
1755
1756 /*
1757  * This function was derived from cache.c:dns_cache_setcachesize().  See the
1758  * function for more details about the logic.
1759  */
1760 void
1761 dns_acache_setcachesize(dns_acache_t *acache, isc_uint32_t size) {
1762         isc_uint32_t lowater;
1763         isc_uint32_t hiwater;
1764
1765         REQUIRE(DNS_ACACHE_VALID(acache));
1766
1767         if (size != 0 && size < DNS_ACACHE_MINSIZE)
1768                 size = DNS_ACACHE_MINSIZE;
1769
1770         hiwater = size - (size >> 3);
1771         lowater = size - (size >> 2);
1772
1773         if (size == 0 || hiwater == 0 || lowater == 0)
1774                 isc_mem_setwater(acache->mctx, water, acache, 0, 0);
1775         else
1776                 isc_mem_setwater(acache->mctx, water, acache,
1777                                  hiwater, lowater);
1778 }