nscd/nscd_helper.c

   1 /* Copyright (C) 1998-2002,2003,2004,2005,2006,2007
   2    Free Software Foundation, Inc.
   3    This file is part of the GNU C Library.
   4    Contributed by Ulrich Drepper <drepper@cygnus.com>, 1998.
   5
   6    The GNU C Library is free software; you can redistribute it and/or
   7    modify it under the terms of the GNU Lesser General Public
   8    License as published by the Free Software Foundation; either
   9    version 2.1 of the License, or (at your option) any later version.
  10
  11    The GNU C Library is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14    Lesser General Public License for more details.
  15
  16    You should have received a copy of the GNU Lesser General Public
  17    License along with the GNU C Library; if not, write to the Free
  18    Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
  19    02111-1307 USA.  */
  20
  21 #include <assert.h>
  22 #include <errno.h>
  23 #include <fcntl.h>
  24 #include <stdbool.h>
  25 #include <string.h>
  26 #include <time.h>
  27 #include <unistd.h>
  28 #include <sys/mman.h>
  29 #include <sys/poll.h>
  30 #include <sys/socket.h>
  31 #include <sys/stat.h>
  32 #include <sys/time.h>
  33 #include <sys/uio.h>
  34 #include <sys/un.h>
  35 #include <not-cancel.h>
  36 #include <nis/rpcsvc/nis.h>
  37
  38 #include "nscd-client.h"
  39
  40
  41 ssize_t
  42 __readall (int fd, void *buf, size_t len)
  43 {
  44   size_t n = len;
  45   ssize_t ret;
  46   do
  47     {
  48       ret = TEMP_FAILURE_RETRY (__read (fd, buf, n));
  49       if (ret <= 0)
  50         break;
  51       buf = (char *) buf + ret;
  52       n -= ret;
  53     }
  54   while (n > 0);
  55   return ret < 0 ? ret : len - n;
  56 }
  57
  58
  59 ssize_t
  60 __readvall (int fd, const struct iovec *iov, int iovcnt)
  61 {
  62   ssize_t ret = TEMP_FAILURE_RETRY (__readv (fd, iov, iovcnt));
  63   if (ret <= 0)
  64     return ret;
  65
  66   size_t total = 0;
  67   for (int i = 0; i < iovcnt; ++i)
  68     total += iov[i].iov_len;
  69
  70   if (ret < total)
  71     {
  72       struct iovec iov_buf[iovcnt];
  73       ssize_t r = ret;
  74
  75       struct iovec *iovp = memcpy (iov_buf, iov, iovcnt * sizeof (*iov));
  76       do
  77         {
  78           while (iovp->iov_len <= r)
  79             {
  80               r -= iovp->iov_len;
  81               --iovcnt;
  82               ++iovp;
  83             }
  84           iovp->iov_base = (char *) iovp->iov_base + r;
  85           iovp->iov_len -= r;
  86           r = TEMP_FAILURE_RETRY (__readv (fd, iovp, iovcnt));
  87           if (r <= 0)
  88             break;
  89           ret += r;
  90         }
  91       while (ret < total);
  92       if (r < 0)
  93         ret = r;
  94     }
  95   return ret;
  96 }
  97
  98
  99 static int
 100 open_socket (request_type type, const char *key, size_t keylen)
 101 {
 102   int sock = __socket (PF_UNIX, SOCK_STREAM, 0);
 103   if (sock < 0)
 104     return -1;
 105
 106   struct
 107   {
 108     request_header req;
 109     char key[keylen];
 110   } reqdata;
 111   size_t real_sizeof_reqdata = sizeof (request_header) + keylen;
 112
 113   /* Make socket non-blocking.  */
 114   __fcntl (sock, F_SETFL, O_RDWR | O_NONBLOCK);
 115
 116   struct sockaddr_un sun;
 117   sun.sun_family = AF_UNIX;
 118   strcpy (sun.sun_path, _PATH_NSCDSOCKET);
 119   if (__connect (sock, (struct sockaddr *) &sun, sizeof (sun)) < 0
 120       && errno != EINPROGRESS)
 121     goto out;
 122
 123   reqdata.req.version = NSCD_VERSION;
 124   reqdata.req.type = type;
 125   reqdata.req.key_len = keylen;
 126
 127   memcpy (reqdata.key, key, keylen);
 128
 129   bool first_try = true;
 130   struct timeval tvend;
 131   /* Fake initializing tvend.  */
 132   asm ("" : "=m" (tvend));
 133   while (1)
 134     {
 135 #ifndef MSG_NOSIGNAL
 136 # define MSG_NOSIGNAL 0
 137 #endif
 138       ssize_t wres = TEMP_FAILURE_RETRY (__send (sock, &reqdata,
 139                                                  real_sizeof_reqdata,
 140                                                  MSG_NOSIGNAL));
 141       if (__builtin_expect (wres == (ssize_t) real_sizeof_reqdata, 1))
 142         /* We managed to send the request.  */
 143         return sock;
 144
 145       if (wres != -1 || errno != EAGAIN)
 146         /* Something is really wrong, no chance to continue.  */
 147         break;
 148
 149       /* The daemon is busy wait for it.  */
 150       int to;
 151       struct timeval now;
 152       (void) __gettimeofday (&now, NULL);
 153       if (first_try)
 154         {
 155           tvend.tv_usec = now.tv_usec;
 156           tvend.tv_sec = now.tv_sec + 5;
 157           to = 5 * 1000;
 158           first_try = false;
 159         }
 160       else
 161         to = ((tvend.tv_sec - now.tv_sec) * 1000
 162               + (tvend.tv_usec - now.tv_usec) / 1000);
 163
 164       struct pollfd fds[1];
 165       fds[0].fd = sock;
 166       fds[0].events = POLLOUT | POLLERR | POLLHUP;
 167       if (__poll (fds, 1, to) <= 0)
 168         /* The connection timed out or broke down.  */
 169         break;
 170
 171       /* We try to write again.  */
 172     }
 173
 174  out:
 175   close_not_cancel_no_status (sock);
 176
 177   return -1;
 178 }
 179
 180
 181 void
 182 __nscd_unmap (struct mapped_database *mapped)
 183 {
 184   assert (mapped->counter == 0);
 185   __munmap ((void *) mapped->head, mapped->mapsize);
 186   free (mapped);
 187 }
 188
 189
 190 static int
 191 wait_on_socket (int sock)
 192 {
 193   struct pollfd fds[1];
 194   fds[0].fd = sock;
 195   fds[0].events = POLLIN | POLLERR | POLLHUP;
 196   int n = __poll (fds, 1, 5 * 1000);
 197   if (n == -1 && __builtin_expect (errno == EINTR, 0))
 198     {
 199       /* Handle the case where the poll() call is interrupted by a
 200          signal.  We cannot just use TEMP_FAILURE_RETRY since it might
 201          lead to infinite loops.  */
 202       struct timeval now;
 203       (void) __gettimeofday (&now, NULL);
 204       long int end = (now.tv_sec + 5) * 1000 + (now.tv_usec + 500) / 1000;
 205       while (1)
 206         {
 207           long int timeout = end - (now.tv_sec * 1000
 208                                     + (now.tv_usec + 500) / 1000);
 209           n = __poll (fds, 1, timeout);
 210           if (n != -1 || errno != EINTR)
 211             break;
 212           (void) __gettimeofday (&now, NULL);
 213         }
 214     }
 215
 216   return n;
 217 }
 218
 219
 220 /* Try to get a file descriptor for the shared meory segment
 221    containing the database.  */
 222 static struct mapped_database *
 223 get_mapping (request_type type, const char *key,
 224              struct mapped_database **mappedp)
 225 {
 226   struct mapped_database *result = NO_MAPPING;
 227 #ifdef SCM_RIGHTS
 228   const size_t keylen = strlen (key) + 1;
 229   int saved_errno = errno;
 230
 231   int mapfd = -1;
 232   char resdata[keylen];
 233
 234   /* Open a socket and send the request.  */
 235   int sock = open_socket (type, key, keylen);
 236   if (sock < 0)
 237     goto out;
 238
 239   /* Room for the data sent along with the file descriptor.  We expect
 240      the key name back.  */
 241   uint64_t mapsize;
 242   struct iovec iov[2];
 243   iov[0].iov_base = resdata;
 244   iov[0].iov_len = keylen;
 245   iov[1].iov_base = &mapsize;
 246   iov[1].iov_len = sizeof (mapsize);
 247
 248   union
 249   {
 250     struct cmsghdr hdr;
 251     char bytes[CMSG_SPACE (sizeof (int))];
 252   } buf;
 253   struct msghdr msg = { .msg_iov = iov, .msg_iovlen = 2,
 254                         .msg_control = buf.bytes,
 255                         .msg_controllen = sizeof (buf) };
 256   struct cmsghdr *cmsg = CMSG_FIRSTHDR (&msg);
 257
 258   cmsg->cmsg_level = SOL_SOCKET;
 259   cmsg->cmsg_type = SCM_RIGHTS;
 260   cmsg->cmsg_len = CMSG_LEN (sizeof (int));
 261
 262   /* This access is well-aligned since BUF is correctly aligned for an
 263      int and CMSG_DATA preserves this alignment.  */
 264   *(int *) CMSG_DATA (cmsg) = -1;
 265
 266   msg.msg_controllen = cmsg->cmsg_len;
 267
 268   if (wait_on_socket (sock) <= 0)
 269     goto out_close2;
 270
 271 # ifndef MSG_CMSG_CLOEXEC
 272 #  define MSG_CMSG_CLOEXEC 0
 273 # endif
 274   ssize_t n = TEMP_FAILURE_RETRY (__recvmsg (sock, &msg, MSG_CMSG_CLOEXEC));
 275
 276   if (__builtin_expect (CMSG_FIRSTHDR (&msg) == NULL
 277                         || (CMSG_FIRSTHDR (&msg)->cmsg_len
 278                             != CMSG_LEN (sizeof (int))), 0))
 279     goto out_close2;
 280
 281   mapfd = *(int *) CMSG_DATA (cmsg);
 282
 283   if (__builtin_expect (n != keylen && n != keylen + sizeof (mapsize), 0))
 284     goto out_close;
 285
 286   if (__builtin_expect (strcmp (resdata, key) != 0, 0))
 287     goto out_close;
 288
 289   if (__builtin_expect (n == keylen, 0))
 290     {
 291       struct stat64 st;
 292       if (__builtin_expect (fstat64 (mapfd, &st) != 0, 0)
 293           || __builtin_expect (st.st_size < sizeof (struct database_pers_head),
 294                                0))
 295         goto out_close;
 296
 297       mapsize = st.st_size;
 298     }
 299
 300   /* The file is large enough, map it now.  */
 301   void *mapping = __mmap (NULL, mapsize, PROT_READ, MAP_SHARED, mapfd, 0);
 302   if (__builtin_expect (mapping != MAP_FAILED, 1))
 303     {
 304       /* Check whether the database is correct and up-to-date.  */
 305       struct database_pers_head *head = mapping;
 306
 307       if (__builtin_expect (head->version != DB_VERSION, 0)
 308           || __builtin_expect (head->header_size != sizeof (*head), 0)
 309           /* This really should not happen but who knows, maybe the update
 310              thread got stuck.  */
 311           || __builtin_expect (! head->nscd_certainly_running
 312                                && (head->timestamp + MAPPING_TIMEOUT
 313                                    < time (NULL)), 0))
 314         {
 315         out_unmap:
 316           __munmap (mapping, mapsize);
 317           goto out_close;
 318         }
 319
 320       size_t size = (sizeof (*head) + roundup (head->module * sizeof (ref_t),
 321                                                ALIGN)
 322                      + head->data_size);
 323
 324       if (__builtin_expect (mapsize < size, 0))
 325         goto out_unmap;
 326
 327       /* Allocate a record for the mapping.  */
 328       struct mapped_database *newp = malloc (sizeof (*newp));
 329       if (newp == NULL)
 330         /* Ugh, after all we went through the memory allocation failed.  */
 331         goto out_unmap;
 332
 333       newp->head = mapping;
 334       newp->data = ((char *) mapping + head->header_size
 335                     + roundup (head->module * sizeof (ref_t), ALIGN));
 336       newp->mapsize = size;
 337       newp->datasize = head->data_size;
 338       /* Set counter to 1 to show it is usable.  */
 339       newp->counter = 1;
 340
 341       result = newp;
 342     }
 343
 344  out_close:
 345   __close (mapfd);
 346  out_close2:
 347   __close (sock);
 348  out:
 349   __set_errno (saved_errno);
 350 #endif  /* SCM_RIGHTS */
 351
 352   struct mapped_database *oldval = *mappedp;
 353   *mappedp = result;
 354
 355   if (oldval != NULL && atomic_decrement_val (&oldval->counter) == 0)
 356     __nscd_unmap (oldval);
 357
 358   return result;
 359 }
 360
 361
 362 struct mapped_database *
 363 __nscd_get_map_ref (request_type type, const char *name,
 364                     volatile struct locked_map_ptr *mapptr, int *gc_cyclep)
 365 {
 366   struct mapped_database *cur = mapptr->mapped;
 367   if (cur == NO_MAPPING)
 368     return cur;
 369
 370   int cnt = 0;
 371   while (__builtin_expect (atomic_compare_and_exchange_val_acq (&mapptr->lock,
 372                                                                 1, 0) != 0, 0))
 373     {
 374       // XXX Best number of rounds?
 375       if (__builtin_expect (++cnt > 5, 0))
 376         return NO_MAPPING;
 377
 378       atomic_delay ();
 379     }
 380
 381   cur = mapptr->mapped;
 382
 383   if (__builtin_expect (cur != NO_MAPPING, 1))
 384     {
 385       /* If not mapped or timestamp not updated, request new map.  */
 386       if (cur == NULL
 387           || (cur->head->nscd_certainly_running == 0
 388               && cur->head->timestamp + MAPPING_TIMEOUT < time (NULL))
 389           || cur->head->data_size > cur->datasize)
 390         cur = get_mapping (type, name,
 391                            (struct mapped_database **) &mapptr->mapped);
 392
 393       if (__builtin_expect (cur != NO_MAPPING, 1))
 394         {
 395           if (__builtin_expect (((*gc_cyclep = cur->head->gc_cycle) & 1) != 0,
 396                                 0))
 397             cur = NO_MAPPING;
 398           else
 399             atomic_increment (&cur->counter);
 400         }
 401     }
 402
 403   mapptr->lock = 0;
 404
 405   return cur;
 406 }
 407
 408
 409 /* Don't return const struct datahead *, as eventhough the record
 410    is normally constant, it can change arbitrarily during nscd
 411    garbage collection.  */
 412 struct datahead *
 413 __nscd_cache_search (request_type type, const char *key, size_t keylen,
 414                      const struct mapped_database *mapped)
 415 {
 416   unsigned long int hash = __nis_hash (key, keylen) % mapped->head->module;
 417   size_t datasize = mapped->datasize;
 418
 419   ref_t work = mapped->head->array[hash];
 420   while (work != ENDREF && work + sizeof (struct hashentry) <= datasize)
 421     {
 422       struct hashentry *here = (struct hashentry *) (mapped->data + work);
 423
 424 #ifndef _STRING_ARCH_unaligned
 425       /* Although during garbage collection when moving struct hashentry
 426          records around we first copy from old to new location and then
 427          adjust pointer from previous hashentry to it, there is no barrier
 428          between those memory writes.  It is very unlikely to hit it,
 429          so check alignment only if a misaligned load can crash the
 430          application.  */
 431       if ((uintptr_t) here & (__alignof__ (*here) - 1))
 432         return NULL;
 433 #endif
 434
 435       if (type == here->type
 436           && keylen == here->len
 437           && here->key + keylen <= datasize
 438           && memcmp (key, mapped->data + here->key, keylen) == 0
 439           && here->packet + sizeof (struct datahead) <= datasize)
 440         {
 441           /* We found the entry.  Increment the appropriate counter.  */
 442           struct datahead *dh
 443             = (struct datahead *) (mapped->data + here->packet);
 444
 445 #ifndef _STRING_ARCH_unaligned
 446           if ((uintptr_t) dh & (__alignof__ (*dh) - 1))
 447             return NULL;
 448 #endif
 449
 450           /* See whether we must ignore the entry or whether something
 451              is wrong because garbage collection is in progress.  */
 452           if (dh->usable && here->packet + dh->allocsize <= datasize)
 453             return dh;
 454         }
 455
 456       work = here->next;
 457     }
 458
 459   return NULL;
 460 }
 461
 462
 463 /* Create a socket connected to a name. */
 464 int
 465 __nscd_open_socket (const char *key, size_t keylen, request_type type,
 466                     void *response, size_t responselen)
 467 {
 468   /* This should never happen and it is something the nscd daemon
 469      enforces, too.  He it helps to limit the amount of stack
 470      used.  */
 471   if (keylen > MAXKEYLEN)
 472     return -1;
 473
 474   int saved_errno = errno;
 475
 476   int sock = open_socket (type, key, keylen);
 477   if (sock >= 0)
 478     {
 479       /* Wait for data.  */
 480       if (wait_on_socket (sock) > 0)
 481         {
 482           ssize_t nbytes = TEMP_FAILURE_RETRY (__read (sock, response,
 483                                                        responselen));
 484           if (nbytes == (ssize_t) responselen)
 485             return sock;
 486         }
 487
 488       close_not_cancel_no_status (sock);
 489     }
 490
 491   __set_errno (saved_errno);
 492
 493   return -1;
 494 }