nscd/cache.c

   1 /* Copyright (c) 1998, 1999, 2003-2008, 2009 Free Software Foundation, Inc.
   2    This file is part of the GNU C Library.
   3    Contributed by Ulrich Drepper <drepper@cygnus.com>, 1998.
   4
   5    This program is free software; you can redistribute it and/or modify
   6    it under the terms of the GNU General Public License as published
   7    by the Free Software Foundation; version 2 of the License, or
   8    (at your option) any later version.
   9
  10    This program is distributed in the hope that it will be useful,
  11    but WITHOUT ANY WARRANTY; without even the implied warranty of
  12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13    GNU General Public License for more details.
  14
  15    You should have received a copy of the GNU General Public License
  16    along with this program; if not, write to the Free Software Foundation,
  17    Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.  */
  18
  19 #include <assert.h>
  20 #include <atomic.h>
  21 #include <errno.h>
  22 #include <error.h>
  23 #include <inttypes.h>
  24 #include <limits.h>
  25 #include <stdlib.h>
  26 #include <string.h>
  27 #include <libintl.h>
  28 #include <arpa/inet.h>
  29 #include <rpcsvc/nis.h>
  30 #include <sys/mman.h>
  31 #include <sys/param.h>
  32 #include <sys/stat.h>
  33 #include <sys/uio.h>
  34
  35 #include "nscd.h"
  36 #include "dbg_log.h"
  37
  38
  39 /* Wrapper functions with error checking for standard functions.  */
  40 extern void *xcalloc (size_t n, size_t s);
  41
  42
  43 /* Number of times a value is reloaded without being used.  UINT_MAX
  44    means unlimited.  */
  45 unsigned int reload_count = DEFAULT_RELOAD_LIMIT;
  46
  47
  48 static void (*const readdfcts[LASTREQ]) (struct database_dyn *,
  49                                          struct hashentry *,
  50                                          struct datahead *) =
  51 {
  52   [GETPWBYNAME] = readdpwbyname,
  53   [GETPWBYUID] = readdpwbyuid,
  54   [GETGRBYNAME] = readdgrbyname,
  55   [GETGRBYGID] = readdgrbygid,
  56   [GETHOSTBYNAME] = readdhstbyname,
  57   [GETHOSTBYNAMEv6] = readdhstbynamev6,
  58   [GETHOSTBYADDR] = readdhstbyaddr,
  59   [GETHOSTBYADDRv6] = readdhstbyaddrv6,
  60   [GETAI] = readdhstai,
  61   [INITGROUPS] = readdinitgroups,
  62   [GETSERVBYNAME] = readdservbyname,
  63   [GETSERVBYPORT] = readdservbyport
  64 };
  65
  66
  67 /* Search the cache for a matching entry and return it when found.  If
  68    this fails search the negative cache and return (void *) -1 if this
  69    search was successful.  Otherwise return NULL.
  70
  71    This function must be called with the read-lock held.  */
  72 struct datahead *
  73 cache_search (request_type type, void *key, size_t len,
  74               struct database_dyn *table, uid_t owner)
  75 {
  76   unsigned long int hash = __nis_hash (key, len) % table->head->module;
  77
  78   unsigned long int nsearched = 0;
  79   struct datahead *result = NULL;
  80
  81   ref_t work = table->head->array[hash];
  82   while (work != ENDREF)
  83     {
  84       ++nsearched;
  85
  86       struct hashentry *here = (struct hashentry *) (table->data + work);
  87
  88       if (type == here->type && len == here->len
  89           && memcmp (key, table->data + here->key, len) == 0
  90           && here->owner == owner)
  91         {
  92           /* We found the entry.  Increment the appropriate counter.  */
  93           struct datahead *dh
  94             = (struct datahead *) (table->data + here->packet);
  95
  96           /* See whether we must ignore the entry.  */
  97           if (dh->usable)
  98             {
  99               /* We do not synchronize the memory here.  The statistics
 100                  data is not crucial, we synchronize only once in a while
 101                  in the cleanup threads.  */
 102               if (dh->notfound)
 103                 ++table->head->neghit;
 104               else
 105                 {
 106                   ++table->head->poshit;
 107
 108                   if (dh->nreloads != 0)
 109                     dh->nreloads = 0;
 110                 }
 111
 112               result = dh;
 113               break;
 114             }
 115         }
 116
 117       work = here->next;
 118     }
 119
 120   if (nsearched > table->head->maxnsearched)
 121     table->head->maxnsearched = nsearched;
 122
 123   return result;
 124 }
 125
 126 /* Add a new entry to the cache.  The return value is zero if the function
 127    call was successful.
 128
 129    This function must be called with the read-lock held.
 130
 131    We modify the table but we nevertheless only acquire a read-lock.
 132    This is ok since we use operations which would be safe even without
 133    locking, given that the `prune_cache' function never runs.  Using
 134    the readlock reduces the chance of conflicts.  */
 135 int
 136 cache_add (int type, const void *key, size_t len, struct datahead *packet,
 137            bool first, struct database_dyn *table,
 138            uid_t owner, bool prune_wakeup)
 139 {
 140   if (__builtin_expect (debug_level >= 2, 0))
 141     {
 142       const char *str;
 143       char buf[INET6_ADDRSTRLEN + 1];
 144       if (type == GETHOSTBYADDR || type == GETHOSTBYADDRv6)
 145         str = inet_ntop (type == GETHOSTBYADDR ? AF_INET : AF_INET6,
 146                          key, buf, sizeof (buf));
 147       else
 148         str = key;
 149
 150       dbg_log (_("add new entry \"%s\" of type %s for %s to cache%s"),
 151                str, serv2str[type], dbnames[table - dbs],
 152                first ? _(" (first)") : "");
 153     }
 154
 155   unsigned long int hash = __nis_hash (key, len) % table->head->module;
 156   struct hashentry *newp;
 157
 158   newp = mempool_alloc (table, sizeof (struct hashentry), 0);
 159   /* If we cannot allocate memory, just do not do anything.  */
 160   if (newp == NULL)
 161     {
 162       /* If necessary mark the entry as unusable so that lookups will
 163          not use it.  */
 164       if (first)
 165         packet->usable = false;
 166
 167       return -1;
 168     }
 169
 170   newp->type = type;
 171   newp->first = first;
 172   newp->len = len;
 173   newp->key = (char *) key - table->data;
 174   assert (newp->key + newp->len <= table->head->first_free);
 175   newp->owner = owner;
 176   newp->packet = (char *) packet - table->data;
 177   assert ((newp->packet & BLOCK_ALIGN_M1) == 0);
 178
 179   /* Put the new entry in the first position.  */
 180   do
 181     newp->next = table->head->array[hash];
 182   while (atomic_compare_and_exchange_bool_acq (&table->head->array[hash],
 183                                                (ref_t) ((char *) newp
 184                                                         - table->data),
 185                                                (ref_t) newp->next));
 186
 187   /* Update the statistics.  */
 188   if (packet->notfound)
 189     ++table->head->negmiss;
 190   else if (first)
 191     ++table->head->posmiss;
 192
 193   /* We depend on this value being correct and at least as high as the
 194      real number of entries.  */
 195   atomic_increment (&table->head->nentries);
 196
 197   /* It does not matter that we are not loading the just increment
 198      value, this is just for statistics.  */
 199   unsigned long int nentries = table->head->nentries;
 200   if (nentries > table->head->maxnentries)
 201     table->head->maxnentries = nentries;
 202
 203   if (table->persistent)
 204     // XXX async OK?
 205     msync ((void *) table->head,
 206            (char *) &table->head->array[hash] - (char *) table->head
 207            + sizeof (ref_t), MS_ASYNC);
 208
 209   /* We do not have to worry about the pruning thread if we are
 210      re-adding the data since this is done by the pruning thread.  We
 211      also do not have to do anything in case this is not the first
 212      time the data is entered since different data heads all have the
 213      same timeout.  */
 214   if (first && prune_wakeup)
 215     {
 216       /* Perhaps the prune thread for the table is not running in a long
 217          time.  Wake it if necessary.  */
 218       pthread_mutex_lock (&table->prune_lock);
 219       time_t next_wakeup = table->wakeup_time;
 220       bool do_wakeup = false;
 221       if (next_wakeup > packet->timeout + CACHE_PRUNE_INTERVAL)
 222         {
 223           table->wakeup_time = packet->timeout;
 224           do_wakeup = true;
 225         }
 226       pthread_mutex_unlock (&table->prune_lock);
 227       if (do_wakeup)
 228         pthread_cond_signal (&table->prune_cond);
 229     }
 230
 231   return 0;
 232 }
 233
 234 /* Walk through the table and remove all entries which lifetime ended.
 235
 236    We have a problem here.  To actually remove the entries we must get
 237    the write-lock.  But since we want to keep the time we have the
 238    lock as short as possible we cannot simply acquire the lock when we
 239    start looking for timedout entries.
 240
 241    Therefore we do it in two stages: first we look for entries which
 242    must be invalidated and remember them.  Then we get the lock and
 243    actually remove them.  This is complicated by the way we have to
 244    free the data structures since some hash table entries share the same
 245    data.  */
 246 time_t
 247 prune_cache (struct database_dyn *table, time_t now, int fd)
 248 {
 249   size_t cnt = table->head->module;
 250
 251   /* If this table is not actually used don't do anything.  */
 252   if (cnt == 0)
 253     {
 254       if (fd != -1)
 255         {
 256           /* Reply to the INVALIDATE initiator.  */
 257           int32_t resp = 0;
 258           writeall (fd, &resp, sizeof (resp));
 259         }
 260
 261       /* No need to do this again anytime soon.  */
 262       return 24 * 60 * 60;
 263     }
 264
 265   /* If we check for the modification of the underlying file we invalidate
 266      the entries also in this case.  */
 267   if (table->inotify_descr < 0 && table->check_file && now != LONG_MAX)
 268     {
 269       struct stat64 st;
 270
 271       if (stat64 (table->filename, &st) < 0)
 272         {
 273           char buf[128];
 274           /* We cannot stat() the file, disable file checking if the
 275              file does not exist.  */
 276           dbg_log (_("cannot stat() file `%s': %s"),
 277                    table->filename, strerror_r (errno, buf, sizeof (buf)));
 278           if (errno == ENOENT)
 279             table->check_file = 0;
 280         }
 281       else
 282         {
 283           if (st.st_mtime != table->file_mtime)
 284             {
 285               /* The file changed.  Invalidate all entries.  */
 286               now = LONG_MAX;
 287               table->file_mtime = st.st_mtime;
 288             }
 289         }
 290     }
 291
 292   /* We run through the table and find values which are not valid anymore.
 293
 294      Note that for the initial step, finding the entries to be removed,
 295      we don't need to get any lock.  It is at all timed assured that the
 296      linked lists are set up correctly and that no second thread prunes
 297      the cache.  */
 298   bool *mark;
 299   size_t memory_needed = cnt * sizeof (bool);
 300   bool mark_use_alloca;
 301   if (__builtin_expect (memory_needed <= MAX_STACK_USE, 1))
 302     {
 303       mark = alloca (cnt * sizeof (bool));
 304       memset (mark, '\0', memory_needed);
 305       mark_use_alloca = true;
 306     }
 307   else
 308     {
 309       mark = xcalloc (1, memory_needed);
 310       mark_use_alloca = false;
 311     }
 312   size_t first = cnt + 1;
 313   size_t last = 0;
 314   char *const data = table->data;
 315   bool any = false;
 316
 317   if (__builtin_expect (debug_level > 2, 0))
 318     dbg_log (_("pruning %s cache; time %ld"),
 319              dbnames[table - dbs], (long int) now);
 320
 321 #define NO_TIMEOUT LONG_MAX
 322   time_t next_timeout = NO_TIMEOUT;
 323   do
 324     {
 325       ref_t run = table->head->array[--cnt];
 326
 327       while (run != ENDREF)
 328         {
 329           struct hashentry *runp = (struct hashentry *) (data + run);
 330           struct datahead *dh = (struct datahead *) (data + runp->packet);
 331
 332           /* Some debug support.  */
 333           if (__builtin_expect (debug_level > 2, 0))
 334             {
 335               char buf[INET6_ADDRSTRLEN];
 336               const char *str;
 337
 338               if (runp->type == GETHOSTBYADDR || runp->type == GETHOSTBYADDRv6)
 339                 {
 340                   inet_ntop (runp->type == GETHOSTBYADDR ? AF_INET : AF_INET6,
 341                              data + runp->key, buf, sizeof (buf));
 342                   str = buf;
 343                 }
 344               else
 345                 str = data + runp->key;
 346
 347               dbg_log (_("considering %s entry \"%s\", timeout %" PRIu64),
 348                        serv2str[runp->type], str, dh->timeout);
 349             }
 350
 351           /* Check whether the entry timed out.  */
 352           if (dh->timeout < now)
 353             {
 354               /* This hash bucket could contain entries which need to
 355                  be looked at.  */
 356               mark[cnt] = true;
 357
 358               first = MIN (first, cnt);
 359               last = MAX (last, cnt);
 360
 361               /* We only have to look at the data of the first entries
 362                  since the count information is kept in the data part
 363                  which is shared.  */
 364               if (runp->first)
 365                 {
 366
 367                   /* At this point there are two choices: we reload the
 368                      value or we discard it.  Do not change NRELOADS if
 369                      we never not reload the record.  */
 370                   if ((reload_count != UINT_MAX
 371                        && __builtin_expect (dh->nreloads >= reload_count, 0))
 372                       /* We always remove negative entries.  */
 373                       || dh->notfound
 374                       /* Discard everything if the user explicitly
 375                          requests it.  */
 376                       || now == LONG_MAX)
 377                     {
 378                       /* Remove the value.  */
 379                       dh->usable = false;
 380
 381                       /* We definitely have some garbage entries now.  */
 382                       any = true;
 383                     }
 384                   else
 385                     {
 386                       /* Reload the value.  We do this only for the
 387                          initially used key, not the additionally
 388                          added derived value.  */
 389                       assert (runp->type < LASTREQ
 390                               && readdfcts[runp->type] != NULL);
 391
 392                       readdfcts[runp->type] (table, runp, dh);
 393
 394                       /* If the entry has been replaced, we might need
 395                          cleanup.  */
 396                       any |= !dh->usable;
 397                     }
 398                 }
 399             }
 400           else
 401             {
 402               assert (dh->usable);
 403               next_timeout = MIN (next_timeout, dh->timeout);
 404             }
 405
 406           run = runp->next;
 407         }
 408     }
 409   while (cnt > 0);
 410
 411   if (__builtin_expect (fd != -1, 0))
 412     {
 413       /* Reply to the INVALIDATE initiator that the cache has been
 414          invalidated.  */
 415       int32_t resp = 0;
 416       writeall (fd, &resp, sizeof (resp));
 417     }
 418
 419   if (first <= last)
 420     {
 421       struct hashentry *head = NULL;
 422
 423       /* Now we have to get the write lock since we are about to modify
 424          the table.  */
 425       if (__builtin_expect (pthread_rwlock_trywrlock (&table->lock) != 0, 0))
 426         {
 427           ++table->head->wrlockdelayed;
 428           pthread_rwlock_wrlock (&table->lock);
 429         }
 430
 431       while (first <= last)
 432         {
 433           if (mark[first])
 434             {
 435               ref_t *old = &table->head->array[first];
 436               ref_t run = table->head->array[first];
 437
 438               assert (run != ENDREF);
 439               do
 440                 {
 441                   struct hashentry *runp = (struct hashentry *) (data + run);
 442                   struct datahead *dh
 443                     = (struct datahead *) (data + runp->packet);
 444
 445                   if (! dh->usable)
 446                     {
 447                       /* We need the list only for debugging but it is
 448                          more costly to avoid creating the list than
 449                          doing it.  */
 450                       runp->dellist = head;
 451                       head = runp;
 452
 453                       /* No need for an atomic operation, we have the
 454                          write lock.  */
 455                       --table->head->nentries;
 456
 457                       run = *old = runp->next;
 458                     }
 459                   else
 460                     {
 461                       old = &runp->next;
 462                       run = runp->next;
 463                     }
 464                 }
 465               while (run != ENDREF);
 466             }
 467
 468           ++first;
 469         }
 470
 471       /* It's all done.  */
 472       pthread_rwlock_unlock (&table->lock);
 473
 474       /* Make sure the data is saved to disk.  */
 475       if (table->persistent)
 476         msync (table->head,
 477                data + table->head->first_free - (char *) table->head,
 478                MS_ASYNC);
 479
 480       /* One extra pass if we do debugging.  */
 481       if (__builtin_expect (debug_level > 0, 0))
 482         {
 483           struct hashentry *runp = head;
 484
 485           while (runp != NULL)
 486             {
 487               char buf[INET6_ADDRSTRLEN];
 488               const char *str;
 489
 490               if (runp->type == GETHOSTBYADDR || runp->type == GETHOSTBYADDRv6)
 491                 {
 492                   inet_ntop (runp->type == GETHOSTBYADDR ? AF_INET : AF_INET6,
 493                              data + runp->key, buf, sizeof (buf));
 494                   str = buf;
 495                 }
 496               else
 497                 str = data + runp->key;
 498
 499               dbg_log ("remove %s entry \"%s\"", serv2str[runp->type], str);
 500
 501               runp = runp->dellist;
 502             }
 503         }
 504     }
 505
 506   if (__builtin_expect (! mark_use_alloca, 0))
 507     free (mark);
 508
 509   /* Run garbage collection if any entry has been removed or replaced.  */
 510   if (any)
 511     gc (table);
 512
 513   /* If there is no entry in the database and we therefore have no new
 514      timeout value, tell the caller to wake up in 24 hours.  */
 515   return next_timeout == NO_TIMEOUT ? 24 * 60 * 60 : next_timeout - now;
 516 }