?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Cache network interface information
[glibc.git] / nscd / connections.c
blob1b8a9bdba74c3c3572a73860dd6fa28816df7448
1 /* Inner loops of cache daemon.
2 Copyright (C) 1998-2007, 2008, 2009, 2011 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 This program is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published
8 by the Free Software Foundation; version 2 of the License, or
9 (at your option) any later version.
10
11 This program 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
14 GNU General Public License for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with this program; if not, write to the Free Software Foundation,
18 Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. */
19
20 #include <alloca.h>
21 #include <assert.h>
22 #include <atomic.h>
23 #include <error.h>
24 #include <errno.h>
25 #include <fcntl.h>
26 #include <grp.h>
27 #include <ifaddrs.h>
28 #include <libintl.h>
29 #include <pthread.h>
30 #include <pwd.h>
31 #include <resolv.h>
32 #include <stdio.h>
33 #include <stdlib.h>
34 #include <unistd.h>
35 #include <arpa/inet.h>
36 #ifdef HAVE_NETLINK
37 # include <netlink/netlink.h>
38 #endif
39 #ifdef HAVE_EPOLL
40 # include <sys/epoll.h>
41 #endif
42 #ifdef HAVE_INOTIFY
43 # include <sys/inotify.h>
44 #endif
45 #include <sys/mman.h>
46 #include <sys/param.h>
47 #include <sys/poll.h>
48 #ifdef HAVE_SENDFILE
49 # include <sys/sendfile.h>
50 #endif
51 #include <sys/socket.h>
52 #include <sys/stat.h>
53 #include <sys/un.h>
54
55 #include "nscd.h"
56 #include "dbg_log.h"
57 #include "selinux.h"
58 #include <resolv/resolv.h>
59 #ifdef HAVE_SENDFILE
60 # include <kernel-features.h>
61 #endif
62
63
64 /* Support to run nscd as an unprivileged user */
65 const char *server_user;
66 static uid_t server_uid;
67 static gid_t server_gid;
68 const char *stat_user;
69 uid_t stat_uid;
70 static gid_t *server_groups;
71 #ifndef NGROUPS
72 # define NGROUPS 32
73 #endif
74 static int server_ngroups;
75
76 static pthread_attr_t attr;
77
78 static void begin_drop_privileges (void);
79 static void finish_drop_privileges (void);
80
81 /* Map request type to a string. */
82 const char *const serv2str[LASTREQ] =
83 {
84 [GETPWBYNAME] = "GETPWBYNAME",
85 [GETPWBYUID] = "GETPWBYUID",
86 [GETGRBYNAME] = "GETGRBYNAME",
87 [GETGRBYGID] = "GETGRBYGID",
88 [GETHOSTBYNAME] = "GETHOSTBYNAME",
89 [GETHOSTBYNAMEv6] = "GETHOSTBYNAMEv6",
90 [GETHOSTBYADDR] = "GETHOSTBYADDR",
91 [GETHOSTBYADDRv6] = "GETHOSTBYADDRv6",
92 [SHUTDOWN] = "SHUTDOWN",
93 [GETSTAT] = "GETSTAT",
94 [INVALIDATE] = "INVALIDATE",
95 [GETFDPW] = "GETFDPW",
96 [GETFDGR] = "GETFDGR",
97 [GETFDHST] = "GETFDHST",
98 [GETAI] = "GETAI",
99 [INITGROUPS] = "INITGROUPS",
100 [GETSERVBYNAME] = "GETSERVBYNAME",
101 [GETSERVBYPORT] = "GETSERVBYPORT",
102 [GETFDSERV] = "GETFDSERV",
103 [GETNETGRENT] = "GETNETGRENT",
104 [INNETGR] = "INNETGR",
105 [GETFDNETGR] = "GETFDNETGR"
106 };
107
108 /* The control data structures for the services. */
109 struct database_dyn dbs[lastdb] =
110 {
111 [pwddb] = {
112 .lock = PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP,
113 .prune_lock = PTHREAD_MUTEX_INITIALIZER,
114 .prune_run_lock = PTHREAD_MUTEX_INITIALIZER,
115 .enabled = 0,
116 .check_file = 1,
117 .persistent = 0,
118 .propagate = 1,
119 .shared = 0,
120 .max_db_size = DEFAULT_MAX_DB_SIZE,
121 .suggested_module = DEFAULT_SUGGESTED_MODULE,
122 .db_filename = _PATH_NSCD_PASSWD_DB,
123 .disabled_iov = &pwd_iov_disabled,
124 .postimeout = 3600,
125 .negtimeout = 20,
126 .wr_fd = -1,
127 .ro_fd = -1,
128 .mmap_used = false
129 },
130 [grpdb] = {
131 .lock = PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP,
132 .prune_lock = PTHREAD_MUTEX_INITIALIZER,
133 .prune_run_lock = PTHREAD_MUTEX_INITIALIZER,
134 .enabled = 0,
135 .check_file = 1,
136 .persistent = 0,
137 .propagate = 1,
138 .shared = 0,
139 .max_db_size = DEFAULT_MAX_DB_SIZE,
140 .suggested_module = DEFAULT_SUGGESTED_MODULE,
141 .db_filename = _PATH_NSCD_GROUP_DB,
142 .disabled_iov = &grp_iov_disabled,
143 .postimeout = 3600,
144 .negtimeout = 60,
145 .wr_fd = -1,
146 .ro_fd = -1,
147 .mmap_used = false
148 },
149 [hstdb] = {
150 .lock = PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP,
151 .prune_lock = PTHREAD_MUTEX_INITIALIZER,
152 .prune_run_lock = PTHREAD_MUTEX_INITIALIZER,
153 .enabled = 0,
154 .check_file = 1,
155 .persistent = 0,
156 .propagate = 0, /* Not used. */
157 .shared = 0,
158 .max_db_size = DEFAULT_MAX_DB_SIZE,
159 .suggested_module = DEFAULT_SUGGESTED_MODULE,
160 .db_filename = _PATH_NSCD_HOSTS_DB,
161 .disabled_iov = &hst_iov_disabled,
162 .postimeout = 3600,
163 .negtimeout = 20,
164 .wr_fd = -1,
165 .ro_fd = -1,
166 .mmap_used = false
167 },
168 [servdb] = {
169 .lock = PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP,
170 .prune_lock = PTHREAD_MUTEX_INITIALIZER,
171 .prune_run_lock = PTHREAD_MUTEX_INITIALIZER,
172 .enabled = 0,
173 .check_file = 1,
174 .persistent = 0,
175 .propagate = 0, /* Not used. */
176 .shared = 0,
177 .max_db_size = DEFAULT_MAX_DB_SIZE,
178 .suggested_module = DEFAULT_SUGGESTED_MODULE,
179 .db_filename = _PATH_NSCD_SERVICES_DB,
180 .disabled_iov = &serv_iov_disabled,
181 .postimeout = 28800,
182 .negtimeout = 20,
183 .wr_fd = -1,
184 .ro_fd = -1,
185 .mmap_used = false
186 },
187 [netgrdb] = {
188 .lock = PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP,
189 .prune_lock = PTHREAD_MUTEX_INITIALIZER,
190 .prune_run_lock = PTHREAD_MUTEX_INITIALIZER,
191 .enabled = 0,
192 .check_file = 1,
193 .persistent = 0,
194 .propagate = 0, /* Not used. */
195 .shared = 0,
196 .max_db_size = DEFAULT_MAX_DB_SIZE,
197 .suggested_module = DEFAULT_SUGGESTED_MODULE,
198 .db_filename = _PATH_NSCD_NETGROUP_DB,
199 .disabled_iov = &netgroup_iov_disabled,
200 .postimeout = 28800,
201 .negtimeout = 20,
202 .wr_fd = -1,
203 .ro_fd = -1,
204 .mmap_used = false
205 }
206 };
207
208
209 /* Mapping of request type to database. */
210 static struct
211 {
212 bool data_request;
213 struct database_dyn *db;
214 } const reqinfo[LASTREQ] =
215 {
216 [GETPWBYNAME] = { true, &dbs[pwddb] },
217 [GETPWBYUID] = { true, &dbs[pwddb] },
218 [GETGRBYNAME] = { true, &dbs[grpdb] },
219 [GETGRBYGID] = { true, &dbs[grpdb] },
220 [GETHOSTBYNAME] = { true, &dbs[hstdb] },
221 [GETHOSTBYNAMEv6] = { true, &dbs[hstdb] },
222 [GETHOSTBYADDR] = { true, &dbs[hstdb] },
223 [GETHOSTBYADDRv6] = { true, &dbs[hstdb] },
224 [SHUTDOWN] = { false, NULL },
225 [GETSTAT] = { false, NULL },
226 [SHUTDOWN] = { false, NULL },
227 [GETFDPW] = { false, &dbs[pwddb] },
228 [GETFDGR] = { false, &dbs[grpdb] },
229 [GETFDHST] = { false, &dbs[hstdb] },
230 [GETAI] = { true, &dbs[hstdb] },
231 [INITGROUPS] = { true, &dbs[grpdb] },
232 [GETSERVBYNAME] = { true, &dbs[servdb] },
233 [GETSERVBYPORT] = { true, &dbs[servdb] },
234 [GETFDSERV] = { false, &dbs[servdb] },
235 [GETNETGRENT] = { true, &dbs[netgrdb] },
236 [INNETGR] = { true, &dbs[netgrdb] },
237 [GETFDNETGR] = { false, &dbs[netgrdb] }
238 };
239
240
241 /* Initial number of threads to use. */
242 int nthreads = -1;
243 /* Maximum number of threads to use. */
244 int max_nthreads = 32;
245
246 /* Socket for incoming connections. */
247 static int sock;
248
249 #ifdef HAVE_INOTIFY
250 /* Inotify descriptor. */
251 int inotify_fd = -1;
252 #endif
253
254 #ifdef HAVE_NETLINK
255 /* Descriptor for netlink status updates. */
256 static int nl_status_fd = -1;
257 #endif
258
259 #ifndef __ASSUME_SOCK_CLOEXEC
260 /* Negative if SOCK_CLOEXEC is not supported, positive if it is, zero
261 before be know the result. */
262 static int have_sock_cloexec;
263 #endif
264 #ifndef __ASSUME_ACCEPT4
265 static int have_accept4;
266 #endif
267
268 /* Number of times clients had to wait. */
269 unsigned long int client_queued;
270
271
272 ssize_t
273 writeall (int fd, const void *buf, size_t len)
274 {
275 size_t n = len;
276 ssize_t ret;
277 do
278 {
279 ret = TEMP_FAILURE_RETRY (send (fd, buf, n, MSG_NOSIGNAL));
280 if (ret <= 0)
281 break;
282 buf = (const char *) buf + ret;
283 n -= ret;
284 }
285 while (n > 0);
286 return ret < 0 ? ret : len - n;
287 }
288
289
290 #ifdef HAVE_SENDFILE
291 ssize_t
292 sendfileall (int tofd, int fromfd, off_t off, size_t len)
293 {
294 ssize_t n = len;
295 ssize_t ret;
296
297 do
298 {
299 ret = TEMP_FAILURE_RETRY (sendfile (tofd, fromfd, &off, n));
300 if (ret <= 0)
301 break;
302 n -= ret;
303 }
304 while (n > 0);
305 return ret < 0 ? ret : len - n;
306 }
307 #endif
308
309
310 enum usekey
311 {
312 use_not = 0,
313 /* The following three are not really used, they are symbolic constants. */
314 use_first = 16,
315 use_begin = 32,
316 use_end = 64,
317
318 use_he = 1,
319 use_he_begin = use_he | use_begin,
320 use_he_end = use_he | use_end,
321 #if SEPARATE_KEY
322 use_key = 2,
323 use_key_begin = use_key | use_begin,
324 use_key_end = use_key | use_end,
325 use_key_first = use_key_begin | use_first,
326 #endif
327 use_data = 3,
328 use_data_begin = use_data | use_begin,
329 use_data_end = use_data | use_end,
330 use_data_first = use_data_begin | use_first
331 };
332
333
334 static int
335 check_use (const char *data, nscd_ssize_t first_free, uint8_t *usemap,
336 enum usekey use, ref_t start, size_t len)
337 {
338 assert (len >= 2);
339
340 if (start > first_free || start + len > first_free
341 || (start & BLOCK_ALIGN_M1))
342 return 0;
343
344 if (usemap[start] == use_not)
345 {
346 /* Add the start marker. */
347 usemap[start] = use | use_begin;
348 use &= ~use_first;
349
350 while (--len > 0)
351 if (usemap[++start] != use_not)
352 return 0;
353 else
354 usemap[start] = use;
355
356 /* Add the end marker. */
357 usemap[start] = use | use_end;
358 }
359 else if ((usemap[start] & ~use_first) == ((use | use_begin) & ~use_first))
360 {
361 /* Hash entries can't be shared. */
362 if (use == use_he)
363 return 0;
364
365 usemap[start] |= (use & use_first);
366 use &= ~use_first;
367
368 while (--len > 1)
369 if (usemap[++start] != use)
370 return 0;
371
372 if (usemap[++start] != (use | use_end))
373 return 0;
374 }
375 else
376 /* Points to a wrong object or somewhere in the middle. */
377 return 0;
378
379 return 1;
380 }
381
382
383 /* Verify data in persistent database. */
384 static int
385 verify_persistent_db (void *mem, struct database_pers_head *readhead, int dbnr)
386 {
387 assert (dbnr == pwddb || dbnr == grpdb || dbnr == hstdb || dbnr == servdb
388 || dbnr == netgrdb);
389
390 time_t now = time (NULL);
391
392 struct database_pers_head *head = mem;
393 struct database_pers_head head_copy = *head;
394
395 /* Check that the header that was read matches the head in the database. */
396 if (memcmp (head, readhead, sizeof (*head)) != 0)
397 return 0;
398
399 /* First some easy tests: make sure the database header is sane. */
400 if (head->version != DB_VERSION
401 || head->header_size != sizeof (*head)
402 /* We allow a timestamp to be one hour ahead of the current time.
403 This should cover daylight saving time changes. */
404 || head->timestamp > now + 60 * 60 + 60
405 || (head->gc_cycle & 1)
406 || head->module == 0
407 || (size_t) head->module > INT32_MAX / sizeof (ref_t)
408 || (size_t) head->data_size > INT32_MAX - head->module * sizeof (ref_t)
409 || head->first_free < 0
410 || head->first_free > head->data_size
411 || (head->first_free & BLOCK_ALIGN_M1) != 0
412 || head->maxnentries < 0
413 || head->maxnsearched < 0)
414 return 0;
415
416 uint8_t *usemap = calloc (head->first_free, 1);
417 if (usemap == NULL)
418 return 0;
419
420 const char *data = (char *) &head->array[roundup (head->module,
421 ALIGN / sizeof (ref_t))];
422
423 nscd_ssize_t he_cnt = 0;
424 for (nscd_ssize_t cnt = 0; cnt < head->module; ++cnt)
425 {
426 ref_t trail = head->array[cnt];
427 ref_t work = trail;
428 int tick = 0;
429
430 while (work != ENDREF)
431 {
432 if (! check_use (data, head->first_free, usemap, use_he, work,
433 sizeof (struct hashentry)))
434 goto fail;
435
436 /* Now we know we can dereference the record. */
437 struct hashentry *here = (struct hashentry *) (data + work);
438
439 ++he_cnt;
440
441 /* Make sure the record is for this type of service. */
442 if (here->type >= LASTREQ
443 || reqinfo[here->type].db != &dbs[dbnr])
444 goto fail;
445
446 /* Validate boolean field value. */
447 if (here->first != false && here->first != true)
448 goto fail;
449
450 if (here->len < 0)
451 goto fail;
452
453 /* Now the data. */
454 if (here->packet < 0
455 || here->packet > head->first_free
456 || here->packet + sizeof (struct datahead) > head->first_free)
457 goto fail;
458
459 struct datahead *dh = (struct datahead *) (data + here->packet);
460
461 if (! check_use (data, head->first_free, usemap,
462 use_data | (here->first ? use_first : 0),
463 here->packet, dh->allocsize))
464 goto fail;
465
466 if (dh->allocsize < sizeof (struct datahead)
467 || dh->recsize > dh->allocsize
468 || (dh->notfound != false && dh->notfound != true)
469 || (dh->usable != false && dh->usable != true))
470 goto fail;
471
472 if (here->key < here->packet + sizeof (struct datahead)
473 || here->key > here->packet + dh->allocsize
474 || here->key + here->len > here->packet + dh->allocsize)
475 {
476 #if SEPARATE_KEY
477 /* If keys can appear outside of data, this should be done
478 instead. But gc doesn't mark the data in that case. */
479 if (! check_use (data, head->first_free, usemap,
480 use_key | (here->first ? use_first : 0),
481 here->key, here->len))
482 #endif
483 goto fail;
484 }
485
486 work = here->next;
487
488 if (work == trail)
489 /* A circular list, this must not happen. */
490 goto fail;
491 if (tick)
492 trail = ((struct hashentry *) (data + trail))->next;
493 tick = 1 - tick;
494 }
495 }
496
497 if (he_cnt != head->nentries)
498 goto fail;
499
500 /* See if all data and keys had at least one reference from
501 he->first == true hashentry. */
502 for (ref_t idx = 0; idx < head->first_free; ++idx)
503 {
504 #if SEPARATE_KEY
505 if (usemap[idx] == use_key_begin)
506 goto fail;
507 #endif
508 if (usemap[idx] == use_data_begin)
509 goto fail;
510 }
511
512 /* Finally, make sure the database hasn't changed since the first test. */
513 if (memcmp (mem, &head_copy, sizeof (*head)) != 0)
514 goto fail;
515
516 free (usemap);
517 return 1;
518
519 fail:
520 free (usemap);
521 return 0;
522 }
523
524
525 #ifdef O_CLOEXEC
526 # define EXTRA_O_FLAGS O_CLOEXEC
527 #else
528 # define EXTRA_O_FLAGS 0
529 #endif
530
531
532 /* Initialize database information structures. */
533 void
534 nscd_init (void)
535 {
536 /* Look up unprivileged uid/gid/groups before we start listening on the
537 socket */
538 if (server_user != NULL)
539 begin_drop_privileges ();
540
541 if (nthreads == -1)
542 /* No configuration for this value, assume a default. */
543 nthreads = 4;
544
545 for (size_t cnt = 0; cnt < lastdb; ++cnt)
546 if (dbs[cnt].enabled)
547 {
548 pthread_rwlock_init (&dbs[cnt].lock, NULL);
549 pthread_mutex_init (&dbs[cnt].memlock, NULL);
550
551 if (dbs[cnt].persistent)
552 {
553 /* Try to open the appropriate file on disk. */
554 int fd = open (dbs[cnt].db_filename, O_RDWR | EXTRA_O_FLAGS);
555 if (fd != -1)
556 {
557 char *msg = NULL;
558 struct stat64 st;
559 void *mem;
560 size_t total;
561 struct database_pers_head head;
562 ssize_t n = TEMP_FAILURE_RETRY (read (fd, &head,
563 sizeof (head)));
564 if (n != sizeof (head) || fstat64 (fd, &st) != 0)
565 {
566 fail_db_errno:
567 /* The code is single-threaded at this point so
568 using strerror is just fine. */
569 msg = strerror (errno);
570 fail_db:
571 dbg_log (_("invalid persistent database file \"%s\": %s"),
572 dbs[cnt].db_filename, msg);
573 unlink (dbs[cnt].db_filename);
574 }
575 else if (head.module == 0 && head.data_size == 0)
576 {
577 /* The file has been created, but the head has not
578 been initialized yet. */
579 msg = _("uninitialized header");
580 goto fail_db;
581 }
582 else if (head.header_size != (int) sizeof (head))
583 {
584 msg = _("header size does not match");
585 goto fail_db;
586 }
587 else if ((total = (sizeof (head)
588 + roundup (head.module * sizeof (ref_t),
589 ALIGN)
590 + head.data_size))
591 > st.st_size
592 || total < sizeof (head))
593 {
594 msg = _("file size does not match");
595 goto fail_db;
596 }
597 /* Note we map with the maximum size allowed for the
598 database. This is likely much larger than the
599 actual file size. This is OK on most OSes since
600 extensions of the underlying file will
601 automatically translate more pages available for
602 memory access. */
603 else if ((mem = mmap (NULL, dbs[cnt].max_db_size,
604 PROT_READ | PROT_WRITE,
605 MAP_SHARED, fd, 0))
606 == MAP_FAILED)
607 goto fail_db_errno;
608 else if (!verify_persistent_db (mem, &head, cnt))
609 {
610 munmap (mem, total);
611 msg = _("verification failed");
612 goto fail_db;
613 }
614 else
615 {
616 /* Success. We have the database. */
617 dbs[cnt].head = mem;
618 dbs[cnt].memsize = total;
619 dbs[cnt].data = (char *)
620 &dbs[cnt].head->array[roundup (dbs[cnt].head->module,
621 ALIGN / sizeof (ref_t))];
622 dbs[cnt].mmap_used = true;
623
624 if (dbs[cnt].suggested_module > head.module)
625 dbg_log (_("suggested size of table for database %s larger than the persistent database's table"),
626 dbnames[cnt]);
627
628 dbs[cnt].wr_fd = fd;
629 fd = -1;
630 /* We also need a read-only descriptor. */
631 if (dbs[cnt].shared)
632 {
633 dbs[cnt].ro_fd = open (dbs[cnt].db_filename,
634 O_RDONLY | EXTRA_O_FLAGS);
635 if (dbs[cnt].ro_fd == -1)
636 dbg_log (_("\
637 cannot create read-only descriptor for \"%s\"; no mmap"),
638 dbs[cnt].db_filename);
639 }
640
641 // XXX Shall we test whether the descriptors actually
642 // XXX point to the same file?
643 }
644
645 /* Close the file descriptors in case something went
646 wrong in which case the variable have not been
647 assigned -1. */
648 if (fd != -1)
649 close (fd);
650 }
651 else if (errno == EACCES)
652 error (EXIT_FAILURE, 0, _("cannot access '%s'"),
653 dbs[cnt].db_filename);
654 }
655
656 if (dbs[cnt].head == NULL)
657 {
658 /* No database loaded. Allocate the data structure,
659 possibly on disk. */
660 struct database_pers_head head;
661 size_t total = (sizeof (head)
662 + roundup (dbs[cnt].suggested_module
663 * sizeof (ref_t), ALIGN)
664 + (dbs[cnt].suggested_module
665 * DEFAULT_DATASIZE_PER_BUCKET));
666
667 /* Try to create the database. If we do not need a
668 persistent database create a temporary file. */
669 int fd;
670 int ro_fd = -1;
671 if (dbs[cnt].persistent)
672 {
673 fd = open (dbs[cnt].db_filename,
674 O_RDWR | O_CREAT | O_EXCL | O_TRUNC | EXTRA_O_FLAGS,
675 S_IRUSR | S_IWUSR);
676 if (fd != -1 && dbs[cnt].shared)
677 ro_fd = open (dbs[cnt].db_filename,
678 O_RDONLY | EXTRA_O_FLAGS);
679 }
680 else
681 {
682 char fname[] = _PATH_NSCD_XYZ_DB_TMP;
683 fd = mkostemp (fname, EXTRA_O_FLAGS);
684
685 /* We do not need the file name anymore after we
686 opened another file descriptor in read-only mode. */
687 if (fd != -1)
688 {
689 if (dbs[cnt].shared)
690 ro_fd = open (fname, O_RDONLY | EXTRA_O_FLAGS);
691
692 unlink (fname);
693 }
694 }
695
696 if (fd == -1)
697 {
698 if (errno == EEXIST)
699 {
700 dbg_log (_("database for %s corrupted or simultaneously used; remove %s manually if necessary and restart"),
701 dbnames[cnt], dbs[cnt].db_filename);
702 // XXX Correct way to terminate?
703 exit (1);
704 }
705
706 if (dbs[cnt].persistent)
707 dbg_log (_("cannot create %s; no persistent database used"),
708 dbs[cnt].db_filename);
709 else
710 dbg_log (_("cannot create %s; no sharing possible"),
711 dbs[cnt].db_filename);
712
713 dbs[cnt].persistent = 0;
714 // XXX remember: no mmap
715 }
716 else
717 {
718 /* Tell the user if we could not create the read-only
719 descriptor. */
720 if (ro_fd == -1 && dbs[cnt].shared)
721 dbg_log (_("\
722 cannot create read-only descriptor for \"%s\"; no mmap"),
723 dbs[cnt].db_filename);
724
725 /* Before we create the header, initialiye the hash
726 table. So that if we get interrupted if writing
727 the header we can recognize a partially initialized
728 database. */
729 size_t ps = sysconf (_SC_PAGESIZE);
730 char tmpbuf[ps];
731 assert (~ENDREF == 0);
732 memset (tmpbuf, '\xff', ps);
733
734 size_t remaining = dbs[cnt].suggested_module * sizeof (ref_t);
735 off_t offset = sizeof (head);
736
737 size_t towrite;
738 if (offset % ps != 0)
739 {
740 towrite = MIN (remaining, ps - (offset % ps));
741 if (pwrite (fd, tmpbuf, towrite, offset) != towrite)
742 goto write_fail;
743 offset += towrite;
744 remaining -= towrite;
745 }
746
747 while (remaining > ps)
748 {
749 if (pwrite (fd, tmpbuf, ps, offset) == -1)
750 goto write_fail;
751 offset += ps;
752 remaining -= ps;
753 }
754
755 if (remaining > 0
756 && pwrite (fd, tmpbuf, remaining, offset) != remaining)
757 goto write_fail;
758
759 /* Create the header of the file. */
760 struct database_pers_head head =
761 {
762 .version = DB_VERSION,
763 .header_size = sizeof (head),
764 .module = dbs[cnt].suggested_module,
765 .data_size = (dbs[cnt].suggested_module
766 * DEFAULT_DATASIZE_PER_BUCKET),
767 .first_free = 0
768 };
769 void *mem;
770
771 if ((TEMP_FAILURE_RETRY (write (fd, &head, sizeof (head)))
772 != sizeof (head))
773 || (TEMP_FAILURE_RETRY_VAL (posix_fallocate (fd, 0, total))
774 != 0)
775 || (mem = mmap (NULL, dbs[cnt].max_db_size,
776 PROT_READ | PROT_WRITE,
777 MAP_SHARED, fd, 0)) == MAP_FAILED)
778 {
779 write_fail:
780 unlink (dbs[cnt].db_filename);
781 dbg_log (_("cannot write to database file %s: %s"),
782 dbs[cnt].db_filename, strerror (errno));
783 dbs[cnt].persistent = 0;
784 }
785 else
786 {
787 /* Success. */
788 dbs[cnt].head = mem;
789 dbs[cnt].data = (char *)
790 &dbs[cnt].head->array[roundup (dbs[cnt].head->module,
791 ALIGN / sizeof (ref_t))];
792 dbs[cnt].memsize = total;
793 dbs[cnt].mmap_used = true;
794
795 /* Remember the descriptors. */
796 dbs[cnt].wr_fd = fd;
797 dbs[cnt].ro_fd = ro_fd;
798 fd = -1;
799 ro_fd = -1;
800 }
801
802 if (fd != -1)
803 close (fd);
804 if (ro_fd != -1)
805 close (ro_fd);
806 }
807 }
808
809 #if !defined O_CLOEXEC || !defined __ASSUME_O_CLOEXEC
810 /* We do not check here whether the O_CLOEXEC provided to the
811 open call was successful or not. The two fcntl calls are
812 only performed once each per process start-up and therefore
813 is not noticeable at all. */
814 if (paranoia
815 && ((dbs[cnt].wr_fd != -1
816 && fcntl (dbs[cnt].wr_fd, F_SETFD, FD_CLOEXEC) == -1)
817 || (dbs[cnt].ro_fd != -1
818 && fcntl (dbs[cnt].ro_fd, F_SETFD, FD_CLOEXEC) == -1)))
819 {
820 dbg_log (_("\
821 cannot set socket to close on exec: %s; disabling paranoia mode"),
822 strerror (errno));
823 paranoia = 0;
824 }
825 #endif
826
827 if (dbs[cnt].head == NULL)
828 {
829 /* We do not use the persistent database. Just
830 create an in-memory data structure. */
831 assert (! dbs[cnt].persistent);
832
833 dbs[cnt].head = xmalloc (sizeof (struct database_pers_head)
834 + (dbs[cnt].suggested_module
835 * sizeof (ref_t)));
836 memset (dbs[cnt].head, '\0', sizeof (struct database_pers_head));
837 assert (~ENDREF == 0);
838 memset (dbs[cnt].head->array, '\xff',
839 dbs[cnt].suggested_module * sizeof (ref_t));
840 dbs[cnt].head->module = dbs[cnt].suggested_module;
841 dbs[cnt].head->data_size = (DEFAULT_DATASIZE_PER_BUCKET
842 * dbs[cnt].head->module);
843 dbs[cnt].data = xmalloc (dbs[cnt].head->data_size);
844 dbs[cnt].head->first_free = 0;
845
846 dbs[cnt].shared = 0;
847 assert (dbs[cnt].ro_fd == -1);
848 }
849 }
850
851 /* Create the socket. */
852 #ifndef __ASSUME_SOCK_CLOEXEC
853 sock = -1;
854 if (have_sock_cloexec >= 0)
855 #endif
856 {
857 sock = socket (AF_UNIX, SOCK_STREAM | SOCK_CLOEXEC | SOCK_NONBLOCK, 0);
858 #ifndef __ASSUME_SOCK_CLOEXEC
859 if (have_sock_cloexec == 0)
860 have_sock_cloexec = sock != -1 || errno != EINVAL ? 1 : -1;
861 #endif
862 }
863 #ifndef __ASSUME_SOCK_CLOEXEC
864 if (have_sock_cloexec < 0)
865 sock = socket (AF_UNIX, SOCK_STREAM, 0);
866 #endif
867 if (sock < 0)
868 {
869 dbg_log (_("cannot open socket: %s"), strerror (errno));
870 exit (errno == EACCES ? 4 : 1);
871 }
872 /* Bind a name to the socket. */
873 struct sockaddr_un sock_addr;
874 sock_addr.sun_family = AF_UNIX;
875 strcpy (sock_addr.sun_path, _PATH_NSCDSOCKET);
876 if (bind (sock, (struct sockaddr *) &sock_addr, sizeof (sock_addr)) < 0)
877 {
878 dbg_log ("%s: %s", _PATH_NSCDSOCKET, strerror (errno));
879 exit (errno == EACCES ? 4 : 1);
880 }
881
882 #ifndef __ASSUME_SOCK_CLOEXEC
883 if (have_sock_cloexec < 0)
884 {
885 /* We don't want to get stuck on accept. */
886 int fl = fcntl (sock, F_GETFL);
887 if (fl == -1 || fcntl (sock, F_SETFL, fl | O_NONBLOCK) == -1)
888 {
889 dbg_log (_("cannot change socket to nonblocking mode: %s"),
890 strerror (errno));
891 exit (1);
892 }
893
894 /* The descriptor needs to be closed on exec. */
895 if (paranoia && fcntl (sock, F_SETFD, FD_CLOEXEC) == -1)
896 {
897 dbg_log (_("cannot set socket to close on exec: %s"),
898 strerror (errno));
899 exit (1);
900 }
901 }
902 #endif
903
904 /* Set permissions for the socket. */
905 chmod (_PATH_NSCDSOCKET, DEFFILEMODE);
906
907 /* Set the socket up to accept connections. */
908 if (listen (sock, SOMAXCONN) < 0)
909 {
910 dbg_log (_("cannot enable socket to accept connections: %s"),
911 strerror (errno));
912 exit (1);
913 }
914
915 #ifdef HAVE_NETLINK
916 if (dbs[hstdb].enabled)
917 {
918 /* Try to open netlink socket to monitor network setting changes. */
919 nl_status_fd = socket (AF_NETLINK,
920 SOCK_RAW | SOCK_CLOEXEC | SOCK_NONBLOCK,
921 NETLINK_ROUTE);
922 if (nl_status_fd != -1)
923 {
924 struct sockaddr_nl snl;
925 memset (&snl, '\0', sizeof (snl));
926 snl.nl_family = AF_NETLINK;
927 /* XXX Is this the best set to use? */
928 snl.nl_groups = (RTMGRP_IPV4_IFADDR | RTMGRP_TC | RTMGRP_IPV4_MROUTE
929 | RTMGRP_IPV4_ROUTE | RTMGRP_IPV4_RULE
930 | RTMGRP_IPV6_IFADDR | RTMGRP_IPV6_MROUTE
931 | RTMGRP_IPV6_ROUTE | RTMGRP_IPV6_IFINFO
932 | RTMGRP_IPV6_PREFIX);
933
934 if (bind (nl_status_fd, (struct sockaddr *) &snl, sizeof (snl)) != 0)
935 {
936 close (nl_status_fd);
937 nl_status_fd = -1;
938 }
939 else
940 {
941 /* Start the timestamp process. */
942 dbs[hstdb].head->extra_data[NSCD_HST_IDX_CONF_TIMESTAMP]
943 = __bump_nl_timestamp ();
944
945 # ifndef __ASSUME_SOCK_CLOEXEC
946 if (have_sock_cloexec < 0)
947 {
948 /* We don't want to get stuck on accept. */
949 int fl = fcntl (nl_status_fd, F_GETFL);
950 if (fl == -1
951 || fcntl (nl_status_fd, F_SETFL, fl | O_NONBLOCK) == -1)
952 {
953 dbg_log (_("\
954 cannot change socket to nonblocking mode: %s"),
955 strerror (errno));
956 exit (1);
957 }
958
959 /* The descriptor needs to be closed on exec. */
960 if (paranoia
961 && fcntl (nl_status_fd, F_SETFD, FD_CLOEXEC) == -1)
962 {
963 dbg_log (_("cannot set socket to close on exec: %s"),
964 strerror (errno));
965 exit (1);
966 }
967 }
968 # endif
969 }
970 }
971 }
972 #endif
973
974 /* Change to unprivileged uid/gid/groups if specified in config file */
975 if (server_user != NULL)
976 finish_drop_privileges ();
977 }
978
979
980 void
981 register_traced_file (size_t dbidx, struct traced_file *finfo)
982 {
983 if (! dbs[dbidx].enabled || ! dbs[dbidx].check_file)
984 return;
985
986 if (__builtin_expect (debug_level > 0, 0))
987 dbg_log (_("register trace file %s for database %s"),
988 finfo->fname, dbnames[dbidx]);
989
990 #ifdef HAVE_INOTIFY
991 if (inotify_fd < 0
992 || (finfo->inotify_descr = inotify_add_watch (inotify_fd, finfo->fname,
993 IN_DELETE_SELF
994 | IN_MODIFY)) < 0)
995 #endif
996 {
997 /* We need the modification date of the file. */
998 struct stat64 st;
999
1000 if (stat64 (finfo->fname, &st) < 0)
1001 {
1002 /* We cannot stat() the file, disable file checking. */
1003 dbg_log (_("cannot stat() file `%s': %s"),
1004 finfo->fname, strerror (errno));
1005 return;
1006 }
1007
1008 finfo->inotify_descr = -1;
1009 finfo->mtime = st.st_mtime;
1010 }
1011
1012 /* Queue up the file name. */
1013 finfo->next = dbs[dbidx].traced_files;
1014 dbs[dbidx].traced_files = finfo;
1015 }
1016
1017
1018 /* Close the connections. */
1019 void
1020 close_sockets (void)
1021 {
1022 close (sock);
1023 }
1024
1025
1026 static void
1027 invalidate_cache (char *key, int fd)
1028 {
1029 dbtype number;
1030 int32_t resp;
1031
1032 for (number = pwddb; number < lastdb; ++number)
1033 if (strcmp (key, dbnames[number]) == 0)
1034 {
1035 if (number == hstdb)
1036 {
1037 struct traced_file *runp = dbs[hstdb].traced_files;
1038 while (runp != NULL)
1039 if (runp->call_res_init)
1040 {
1041 res_init ();
1042 break;
1043 }
1044 else
1045 runp = runp->next;
1046 }
1047 break;
1048 }
1049
1050 if (number == lastdb)
1051 {
1052 resp = EINVAL;
1053 writeall (fd, &resp, sizeof (resp));
1054 return;
1055 }
1056
1057 if (dbs[number].enabled)
1058 {
1059 pthread_mutex_lock (&dbs[number].prune_run_lock);
1060 prune_cache (&dbs[number], LONG_MAX, fd);
1061 pthread_mutex_unlock (&dbs[number].prune_run_lock);
1062 }
1063 else
1064 {
1065 resp = 0;
1066 writeall (fd, &resp, sizeof (resp));
1067 }
1068 }
1069
1070
1071 #ifdef SCM_RIGHTS
1072 static void
1073 send_ro_fd (struct database_dyn *db, char *key, int fd)
1074 {
1075 /* If we do not have an read-only file descriptor do nothing. */
1076 if (db->ro_fd == -1)
1077 return;
1078
1079 /* We need to send some data along with the descriptor. */
1080 uint64_t mapsize = (db->head->data_size
1081 + roundup (db->head->module * sizeof (ref_t), ALIGN)
1082 + sizeof (struct database_pers_head));
1083 struct iovec iov[2];
1084 iov[0].iov_base = key;
1085 iov[0].iov_len = strlen (key) + 1;
1086 iov[1].iov_base = &mapsize;
1087 iov[1].iov_len = sizeof (mapsize);
1088
1089 /* Prepare the control message to transfer the descriptor. */
1090 union
1091 {
1092 struct cmsghdr hdr;
1093 char bytes[CMSG_SPACE (sizeof (int))];
1094 } buf;
1095 struct msghdr msg = { .msg_iov = iov, .msg_iovlen = 2,
1096 .msg_control = buf.bytes,
1097 .msg_controllen = sizeof (buf) };
1098 struct cmsghdr *cmsg = CMSG_FIRSTHDR (&msg);
1099
1100 cmsg->cmsg_level = SOL_SOCKET;
1101 cmsg->cmsg_type = SCM_RIGHTS;
1102 cmsg->cmsg_len = CMSG_LEN (sizeof (int));
1103
1104 int *ip = (int *) CMSG_DATA (cmsg);
1105 *ip = db->ro_fd;
1106
1107 msg.msg_controllen = cmsg->cmsg_len;
1108
1109 /* Send the control message. We repeat when we are interrupted but
1110 everything else is ignored. */
1111 #ifndef MSG_NOSIGNAL
1112 # define MSG_NOSIGNAL 0
1113 #endif
1114 (void) TEMP_FAILURE_RETRY (sendmsg (fd, &msg, MSG_NOSIGNAL));
1115
1116 if (__builtin_expect (debug_level > 0, 0))
1117 dbg_log (_("provide access to FD %d, for %s"), db->ro_fd, key);
1118 }
1119 #endif /* SCM_RIGHTS */
1120
1121
1122 /* Handle new request. */
1123 static void
1124 handle_request (int fd, request_header *req, void *key, uid_t uid, pid_t pid)
1125 {
1126 if (__builtin_expect (req->version, NSCD_VERSION) != NSCD_VERSION)
1127 {
1128 if (debug_level > 0)
1129 dbg_log (_("\
1130 cannot handle old request version %d; current version is %d"),
1131 req->version, NSCD_VERSION);
1132 return;
1133 }
1134
1135 /* Perform the SELinux check before we go on to the standard checks. */
1136 if (selinux_enabled && nscd_request_avc_has_perm (fd, req->type) != 0)
1137 {
1138 if (debug_level > 0)
1139 {
1140 #ifdef SO_PEERCRED
1141 # ifdef PATH_MAX
1142 char buf[PATH_MAX];
1143 # else
1144 char buf[4096];
1145 # endif
1146
1147 snprintf (buf, sizeof (buf), "/proc/%ld/exe", (long int) pid);
1148 ssize_t n = readlink (buf, buf, sizeof (buf) - 1);
1149
1150 if (n <= 0)
1151 dbg_log (_("\
1152 request from %ld not handled due to missing permission"), (long int) pid);
1153 else
1154 {
1155 buf[n] = '\0';
1156 dbg_log (_("\
1157 request from '%s' [%ld] not handled due to missing permission"),
1158 buf, (long int) pid);
1159 }
1160 #else
1161 dbg_log (_("request not handled due to missing permission"));
1162 #endif
1163 }
1164 return;
1165 }
1166
1167 struct database_dyn *db = reqinfo[req->type].db;
1168
1169 /* See whether we can service the request from the cache. */
1170 if (__builtin_expect (reqinfo[req->type].data_request, true))
1171 {
1172 if (__builtin_expect (debug_level, 0) > 0)
1173 {
1174 if (req->type == GETHOSTBYADDR || req->type == GETHOSTBYADDRv6)
1175 {
1176 char buf[INET6_ADDRSTRLEN];
1177
1178 dbg_log ("\t%s (%s)", serv2str[req->type],
1179 inet_ntop (req->type == GETHOSTBYADDR
1180 ? AF_INET : AF_INET6,
1181 key, buf, sizeof (buf)));
1182 }
1183 else
1184 dbg_log ("\t%s (%s)", serv2str[req->type], (char *) key);
1185 }
1186
1187 /* Is this service enabled? */
1188 if (__builtin_expect (!db->enabled, 0))
1189 {
1190 /* No, sent the prepared record. */
1191 if (TEMP_FAILURE_RETRY (send (fd, db->disabled_iov->iov_base,
1192 db->disabled_iov->iov_len,
1193 MSG_NOSIGNAL))
1194 != (ssize_t) db->disabled_iov->iov_len
1195 && __builtin_expect (debug_level, 0) > 0)
1196 {
1197 /* We have problems sending the result. */
1198 char buf[256];
1199 dbg_log (_("cannot write result: %s"),
1200 strerror_r (errno, buf, sizeof (buf)));
1201 }
1202
1203 return;
1204 }
1205
1206 /* Be sure we can read the data. */
1207 if (__builtin_expect (pthread_rwlock_tryrdlock (&db->lock) != 0, 0))
1208 {
1209 ++db->head->rdlockdelayed;
1210 pthread_rwlock_rdlock (&db->lock);
1211 }
1212
1213 /* See whether we can handle it from the cache. */
1214 struct datahead *cached;
1215 cached = (struct datahead *) cache_search (req->type, key, req->key_len,
1216 db, uid);
1217 if (cached != NULL)
1218 {
1219 /* Hurray it's in the cache. */
1220 ssize_t nwritten;
1221
1222 #ifdef HAVE_SENDFILE
1223 if (__builtin_expect (db->mmap_used, 1))
1224 {
1225 assert (db->wr_fd != -1);
1226 assert ((char *) cached->data > (char *) db->data);
1227 assert ((char *) cached->data - (char *) db->head
1228 + cached->recsize
1229 <= (sizeof (struct database_pers_head)
1230 + db->head->module * sizeof (ref_t)
1231 + db->head->data_size));
1232 nwritten = sendfileall (fd, db->wr_fd,
1233 (char *) cached->data
1234 - (char *) db->head, cached->recsize);
1235 # ifndef __ASSUME_SENDFILE
1236 if (nwritten == -1 && errno == ENOSYS)
1237 goto use_write;
1238 # endif
1239 }
1240 else
1241 # ifndef __ASSUME_SENDFILE
1242 use_write:
1243 # endif
1244 #endif
1245 nwritten = writeall (fd, cached->data, cached->recsize);
1246
1247 if (nwritten != cached->recsize
1248 && __builtin_expect (debug_level, 0) > 0)
1249 {
1250 /* We have problems sending the result. */
1251 char buf[256];
1252 dbg_log (_("cannot write result: %s"),
1253 strerror_r (errno, buf, sizeof (buf)));
1254 }
1255
1256 pthread_rwlock_unlock (&db->lock);
1257
1258 return;
1259 }
1260
1261 pthread_rwlock_unlock (&db->lock);
1262 }
1263 else if (__builtin_expect (debug_level, 0) > 0)
1264 {
1265 if (req->type == INVALIDATE)
1266 dbg_log ("\t%s (%s)", serv2str[req->type], (char *) key);
1267 else
1268 dbg_log ("\t%s", serv2str[req->type]);
1269 }
1270
1271 /* Handle the request. */
1272 switch (req->type)
1273 {
1274 case GETPWBYNAME:
1275 addpwbyname (db, fd, req, key, uid);
1276 break;
1277
1278 case GETPWBYUID:
1279 addpwbyuid (db, fd, req, key, uid);
1280 break;
1281
1282 case GETGRBYNAME:
1283 addgrbyname (db, fd, req, key, uid);
1284 break;
1285
1286 case GETGRBYGID:
1287 addgrbygid (db, fd, req, key, uid);
1288 break;
1289
1290 case GETHOSTBYNAME:
1291 addhstbyname (db, fd, req, key, uid);
1292 break;
1293
1294 case GETHOSTBYNAMEv6:
1295 addhstbynamev6 (db, fd, req, key, uid);
1296 break;
1297
1298 case GETHOSTBYADDR:
1299 addhstbyaddr (db, fd, req, key, uid);
1300 break;
1301
1302 case GETHOSTBYADDRv6:
1303 addhstbyaddrv6 (db, fd, req, key, uid);
1304 break;
1305
1306 case GETAI:
1307 addhstai (db, fd, req, key, uid);
1308 break;
1309
1310 case INITGROUPS:
1311 addinitgroups (db, fd, req, key, uid);
1312 break;
1313
1314 case GETSERVBYNAME:
1315 addservbyname (db, fd, req, key, uid);
1316 break;
1317
1318 case GETSERVBYPORT:
1319 addservbyport (db, fd, req, key, uid);
1320 break;
1321
1322 case GETNETGRENT:
1323 addgetnetgrent (db, fd, req, key, uid);
1324 break;
1325
1326 case INNETGR:
1327 addinnetgr (db, fd, req, key, uid);
1328 break;
1329
1330 case GETSTAT:
1331 case SHUTDOWN:
1332 case INVALIDATE:
1333 {
1334 /* Get the callers credentials. */
1335 #ifdef SO_PEERCRED
1336 struct ucred caller;
1337 socklen_t optlen = sizeof (caller);
1338
1339 if (getsockopt (fd, SOL_SOCKET, SO_PEERCRED, &caller, &optlen) < 0)
1340 {
1341 char buf[256];
1342
1343 dbg_log (_("error getting caller's id: %s"),
1344 strerror_r (errno, buf, sizeof (buf)));
1345 break;
1346 }
1347
1348 uid = caller.uid;
1349 #else
1350 /* Some systems have no SO_PEERCRED implementation. They don't
1351 care about security so we don't as well. */
1352 uid = 0;
1353 #endif
1354 }
1355
1356 /* Accept shutdown, getstat and invalidate only from root. For
1357 the stat call also allow the user specified in the config file. */
1358 if (req->type == GETSTAT)
1359 {
1360 if (uid == 0 || uid == stat_uid)
1361 send_stats (fd, dbs);
1362 }
1363 else if (uid == 0)
1364 {
1365 if (req->type == INVALIDATE)
1366 invalidate_cache (key, fd);
1367 else
1368 termination_handler (0);
1369 }
1370 break;
1371
1372 case GETFDPW:
1373 case GETFDGR:
1374 case GETFDHST:
1375 case GETFDSERV:
1376 case GETFDNETGR:
1377 #ifdef SCM_RIGHTS
1378 send_ro_fd (reqinfo[req->type].db, key, fd);
1379 #endif
1380 break;
1381
1382 default:
1383 /* Ignore the command, it's nothing we know. */
1384 break;
1385 }
1386 }
1387
1388
1389 /* Restart the process. */
1390 static void
1391 restart (void)
1392 {
1393 /* First determine the parameters. We do not use the parameters
1394 passed to main() since in case nscd is started by running the
1395 dynamic linker this will not work. Yes, this is not the usual
1396 case but nscd is part of glibc and we occasionally do this. */
1397 size_t buflen = 1024;
1398 char *buf = alloca (buflen);
1399 size_t readlen = 0;
1400 int fd = open ("/proc/self/cmdline", O_RDONLY);
1401 if (fd == -1)
1402 {
1403 dbg_log (_("\
1404 cannot open /proc/self/cmdline: %s; disabling paranoia mode"),
1405 strerror (errno));
1406
1407 paranoia = 0;
1408 return;
1409 }
1410
1411 while (1)
1412 {
1413 ssize_t n = TEMP_FAILURE_RETRY (read (fd, buf + readlen,
1414 buflen - readlen));
1415 if (n == -1)
1416 {
1417 dbg_log (_("\
1418 cannot read /proc/self/cmdline: %s; disabling paranoia mode"),
1419 strerror (errno));
1420
1421 close (fd);
1422 paranoia = 0;
1423 return;
1424 }
1425
1426 readlen += n;
1427
1428 if (readlen < buflen)
1429 break;
1430
1431 /* We might have to extend the buffer. */
1432 size_t old_buflen = buflen;
1433 char *newp = extend_alloca (buf, buflen, 2 * buflen);
1434 buf = memmove (newp, buf, old_buflen);
1435 }
1436
1437 close (fd);
1438
1439 /* Parse the command line. Worst case scenario: every two
1440 characters form one parameter (one character plus NUL). */
1441 char **argv = alloca ((readlen / 2 + 1) * sizeof (argv[0]));
1442 int argc = 0;
1443
1444 char *cp = buf;
1445 while (cp < buf + readlen)
1446 {
1447 argv[argc++] = cp;
1448 cp = (char *) rawmemchr (cp, '\0') + 1;
1449 }
1450 argv[argc] = NULL;
1451
1452 /* Second, change back to the old user if we changed it. */
1453 if (server_user != NULL)
1454 {
1455 if (setresuid (old_uid, old_uid, old_uid) != 0)
1456 {
1457 dbg_log (_("\
1458 cannot change to old UID: %s; disabling paranoia mode"),
1459 strerror (errno));
1460
1461 paranoia = 0;
1462 return;
1463 }
1464
1465 if (setresgid (old_gid, old_gid, old_gid) != 0)
1466 {
1467 dbg_log (_("\
1468 cannot change to old GID: %s; disabling paranoia mode"),
1469 strerror (errno));
1470
1471 setuid (server_uid);
1472 paranoia = 0;
1473 return;
1474 }
1475 }
1476
1477 /* Next change back to the old working directory. */
1478 if (chdir (oldcwd) == -1)
1479 {
1480 dbg_log (_("\
1481 cannot change to old working directory: %s; disabling paranoia mode"),
1482 strerror (errno));
1483
1484 if (server_user != NULL)
1485 {
1486 setuid (server_uid);
1487 setgid (server_gid);
1488 }
1489 paranoia = 0;
1490 return;
1491 }
1492
1493 /* Synchronize memory. */
1494 int32_t certainly[lastdb];
1495 for (int cnt = 0; cnt < lastdb; ++cnt)
1496 if (dbs[cnt].enabled)
1497 {
1498 /* Make sure nobody keeps using the database. */
1499 dbs[cnt].head->timestamp = 0;
1500 certainly[cnt] = dbs[cnt].head->nscd_certainly_running;
1501 dbs[cnt].head->nscd_certainly_running = 0;
1502
1503 if (dbs[cnt].persistent)
1504 // XXX async OK?
1505 msync (dbs[cnt].head, dbs[cnt].memsize, MS_ASYNC);
1506 }
1507
1508 /* The preparations are done. */
1509 #ifdef PATH_MAX
1510 char pathbuf[PATH_MAX];
1511 #else
1512 char pathbuf[256];
1513 #endif
1514 /* Try to exec the real nscd program so the process name (as reported
1515 in /proc/PID/status) will be 'nscd', but fall back to /proc/self/exe
1516 if readlink or the exec with the result of the readlink call fails. */
1517 ssize_t n = readlink ("/proc/self/exe", pathbuf, sizeof (pathbuf) - 1);
1518 if (n != -1)
1519 {
1520 pathbuf[n] = '\0';
1521 execv (pathbuf, argv);
1522 }
1523 execv ("/proc/self/exe", argv);
1524
1525 /* If we come here, we will never be able to re-exec. */
1526 dbg_log (_("re-exec failed: %s; disabling paranoia mode"),
1527 strerror (errno));
1528
1529 if (server_user != NULL)
1530 {
1531 setuid (server_uid);
1532 setgid (server_gid);
1533 }
1534 if (chdir ("/") != 0)
1535 dbg_log (_("cannot change current working directory to \"/\": %s"),
1536 strerror (errno));
1537 paranoia = 0;
1538
1539 /* Reenable the databases. */
1540 time_t now = time (NULL);
1541 for (int cnt = 0; cnt < lastdb; ++cnt)
1542 if (dbs[cnt].enabled)
1543 {
1544 dbs[cnt].head->timestamp = now;
1545 dbs[cnt].head->nscd_certainly_running = certainly[cnt];
1546 }
1547 }
1548
1549
1550 /* List of file descriptors. */
1551 struct fdlist
1552 {
1553 int fd;
1554 struct fdlist *next;
1555 };
1556 /* Memory allocated for the list. */
1557 static struct fdlist *fdlist;
1558 /* List of currently ready-to-read file descriptors. */
1559 static struct fdlist *readylist;
1560
1561 /* Conditional variable and mutex to signal availability of entries in
1562 READYLIST. The condvar is initialized dynamically since we might
1563 use a different clock depending on availability. */
1564 static pthread_cond_t readylist_cond = PTHREAD_COND_INITIALIZER;
1565 static pthread_mutex_t readylist_lock = PTHREAD_MUTEX_INITIALIZER;
1566
1567 /* The clock to use with the condvar. */
1568 static clockid_t timeout_clock = CLOCK_REALTIME;
1569
1570 /* Number of threads ready to handle the READYLIST. */
1571 static unsigned long int nready;
1572
1573
1574 /* Function for the clean-up threads. */
1575 static void *
1576 __attribute__ ((__noreturn__))
1577 nscd_run_prune (void *p)
1578 {
1579 const long int my_number = (long int) p;
1580 assert (dbs[my_number].enabled);
1581
1582 int dont_need_update = setup_thread (&dbs[my_number]);
1583
1584 time_t now = time (NULL);
1585
1586 /* We are running. */
1587 dbs[my_number].head->timestamp = now;
1588
1589 struct timespec prune_ts;
1590 if (__builtin_expect (clock_gettime (timeout_clock, &prune_ts) == -1, 0))
1591 /* Should never happen. */
1592 abort ();
1593
1594 /* Compute the initial timeout time. Prevent all the timers to go
1595 off at the same time by adding a db-based value. */
1596 prune_ts.tv_sec += CACHE_PRUNE_INTERVAL + my_number;
1597 dbs[my_number].wakeup_time = now + CACHE_PRUNE_INTERVAL + my_number;
1598
1599 pthread_mutex_t *prune_lock = &dbs[my_number].prune_lock;
1600 pthread_mutex_t *prune_run_lock = &dbs[my_number].prune_run_lock;
1601 pthread_cond_t *prune_cond = &dbs[my_number].prune_cond;
1602
1603 pthread_mutex_lock (prune_lock);
1604 while (1)
1605 {
1606 /* Wait, but not forever. */
1607 int e = 0;
1608 if (! dbs[my_number].clear_cache)
1609 e = pthread_cond_timedwait (prune_cond, prune_lock, &prune_ts);
1610 assert (__builtin_expect (e == 0 || e == ETIMEDOUT, 1));
1611
1612 time_t next_wait;
1613 now = time (NULL);
1614 if (e == ETIMEDOUT || now >= dbs[my_number].wakeup_time
1615 || dbs[my_number].clear_cache)
1616 {
1617 /* We will determine the new timout values based on the
1618 cache content. Should there be concurrent additions to
1619 the cache which are not accounted for in the cache
1620 pruning we want to know about it. Therefore set the
1621 timeout to the maximum. It will be descreased when adding
1622 new entries to the cache, if necessary. */
1623 dbs[my_number].wakeup_time = MAX_TIMEOUT_VALUE;
1624
1625 /* Unconditionally reset the flag. */
1626 time_t prune_now = dbs[my_number].clear_cache ? LONG_MAX : now;
1627 dbs[my_number].clear_cache = 0;
1628
1629 pthread_mutex_unlock (prune_lock);
1630
1631 /* We use a separate lock for running the prune function (instead
1632 of keeping prune_lock locked) because this enables concurrent
1633 invocations of cache_add which might modify the timeout value. */
1634 pthread_mutex_lock (prune_run_lock);
1635 next_wait = prune_cache (&dbs[my_number], prune_now, -1);
1636 pthread_mutex_unlock (prune_run_lock);
1637
1638 next_wait = MAX (next_wait, CACHE_PRUNE_INTERVAL);
1639 /* If clients cannot determine for sure whether nscd is running
1640 we need to wake up occasionally to update the timestamp.
1641 Wait 90% of the update period. */
1642 #define UPDATE_MAPPING_TIMEOUT (MAPPING_TIMEOUT * 9 / 10)
1643 if (__builtin_expect (! dont_need_update, 0))
1644 {
1645 next_wait = MIN (UPDATE_MAPPING_TIMEOUT, next_wait);
1646 dbs[my_number].head->timestamp = now;
1647 }
1648
1649 pthread_mutex_lock (prune_lock);
1650
1651 /* Make it known when we will wake up again. */
1652 if (now + next_wait < dbs[my_number].wakeup_time)
1653 dbs[my_number].wakeup_time = now + next_wait;
1654 else
1655 next_wait = dbs[my_number].wakeup_time - now;
1656 }
1657 else
1658 /* The cache was just pruned. Do not do it again now. Just
1659 use the new timeout value. */
1660 next_wait = dbs[my_number].wakeup_time - now;
1661
1662 if (clock_gettime (timeout_clock, &prune_ts) == -1)
1663 /* Should never happen. */
1664 abort ();
1665
1666 /* Compute next timeout time. */
1667 prune_ts.tv_sec += next_wait;
1668 }
1669 }
1670
1671
1672 /* This is the main loop. It is replicated in different threads but
1673 the use of the ready list makes sure only one thread handles an
1674 incoming connection. */
1675 static void *
1676 __attribute__ ((__noreturn__))
1677 nscd_run_worker (void *p)
1678 {
1679 char buf[256];
1680
1681 /* Initial locking. */
1682 pthread_mutex_lock (&readylist_lock);
1683
1684 /* One more thread available. */
1685 ++nready;
1686
1687 while (1)
1688 {
1689 while (readylist == NULL)
1690 pthread_cond_wait (&readylist_cond, &readylist_lock);
1691
1692 struct fdlist *it = readylist->next;
1693 if (readylist->next == readylist)
1694 /* Just one entry on the list. */
1695 readylist = NULL;
1696 else
1697 readylist->next = it->next;
1698
1699 /* Extract the information and mark the record ready to be used
1700 again. */
1701 int fd = it->fd;
1702 it->next = NULL;
1703
1704 /* One more thread available. */
1705 --nready;
1706
1707 /* We are done with the list. */
1708 pthread_mutex_unlock (&readylist_lock);
1709
1710 #ifndef __ASSUME_ACCEPT4
1711 if (have_accept4 < 0)
1712 {
1713 /* We do not want to block on a short read or so. */
1714 int fl = fcntl (fd, F_GETFL);
1715 if (fl == -1 || fcntl (fd, F_SETFL, fl | O_NONBLOCK) == -1)
1716 goto close_and_out;
1717 }
1718 #endif
1719
1720 /* Now read the request. */
1721 request_header req;
1722 if (__builtin_expect (TEMP_FAILURE_RETRY (read (fd, &req, sizeof (req)))
1723 != sizeof (req), 0))
1724 {
1725 /* We failed to read data. Note that this also might mean we
1726 failed because we would have blocked. */
1727 if (debug_level > 0)
1728 dbg_log (_("short read while reading request: %s"),
1729 strerror_r (errno, buf, sizeof (buf)));
1730 goto close_and_out;
1731 }
1732
1733 /* Check whether this is a valid request type. */
1734 if (req.type < GETPWBYNAME || req.type >= LASTREQ)
1735 goto close_and_out;
1736
1737 /* Some systems have no SO_PEERCRED implementation. They don't
1738 care about security so we don't as well. */
1739 uid_t uid = -1;
1740 #ifdef SO_PEERCRED
1741 pid_t pid = 0;
1742
1743 if (__builtin_expect (debug_level > 0, 0))
1744 {
1745 struct ucred caller;
1746 socklen_t optlen = sizeof (caller);
1747
1748 if (getsockopt (fd, SOL_SOCKET, SO_PEERCRED, &caller, &optlen) == 0)
1749 pid = caller.pid;
1750 }
1751 #else
1752 const pid_t pid = 0;
1753 #endif
1754
1755 /* It should not be possible to crash the nscd with a silly
1756 request (i.e., a terribly large key). We limit the size to 1kb. */
1757 if (__builtin_expect (req.key_len, 1) < 0
1758 || __builtin_expect (req.key_len, 1) > MAXKEYLEN)
1759 {
1760 if (debug_level > 0)
1761 dbg_log (_("key length in request too long: %d"), req.key_len);
1762 }
1763 else
1764 {
1765 /* Get the key. */
1766 char keybuf[MAXKEYLEN];
1767
1768 if (__builtin_expect (TEMP_FAILURE_RETRY (read (fd, keybuf,
1769 req.key_len))
1770 != req.key_len, 0))
1771 {
1772 /* Again, this can also mean we would have blocked. */
1773 if (debug_level > 0)
1774 dbg_log (_("short read while reading request key: %s"),
1775 strerror_r (errno, buf, sizeof (buf)));
1776 goto close_and_out;
1777 }
1778
1779 if (__builtin_expect (debug_level, 0) > 0)
1780 {
1781 #ifdef SO_PEERCRED
1782 if (pid != 0)
1783 dbg_log (_("\
1784 handle_request: request received (Version = %d) from PID %ld"),
1785 req.version, (long int) pid);
1786 else
1787 #endif
1788 dbg_log (_("\
1789 handle_request: request received (Version = %d)"), req.version);
1790 }
1791
1792 /* Phew, we got all the data, now process it. */
1793 handle_request (fd, &req, keybuf, uid, pid);
1794 }
1795
1796 close_and_out:
1797 /* We are done. */
1798 close (fd);
1799
1800 /* Re-locking. */
1801 pthread_mutex_lock (&readylist_lock);
1802
1803 /* One more thread available. */
1804 ++nready;
1805 }
1806 /* NOTREACHED */
1807 }
1808
1809
1810 static unsigned int nconns;
1811
1812 static void
1813 fd_ready (int fd)
1814 {
1815 pthread_mutex_lock (&readylist_lock);
1816
1817 /* Find an empty entry in FDLIST. */
1818 size_t inner;
1819 for (inner = 0; inner < nconns; ++inner)
1820 if (fdlist[inner].next == NULL)
1821 break;
1822 assert (inner < nconns);
1823
1824 fdlist[inner].fd = fd;
1825
1826 if (readylist == NULL)
1827 readylist = fdlist[inner].next = &fdlist[inner];
1828 else
1829 {
1830 fdlist[inner].next = readylist->next;
1831 readylist = readylist->next = &fdlist[inner];
1832 }
1833
1834 bool do_signal = true;
1835 if (__builtin_expect (nready == 0, 0))
1836 {
1837 ++client_queued;
1838 do_signal = false;
1839
1840 /* Try to start another thread to help out. */
1841 pthread_t th;
1842 if (nthreads < max_nthreads
1843 && pthread_create (&th, &attr, nscd_run_worker,
1844 (void *) (long int) nthreads) == 0)
1845 {
1846 /* We got another thread. */
1847 ++nthreads;
1848 /* The new thread might need a kick. */
1849 do_signal = true;
1850 }
1851
1852 }
1853
1854 pthread_mutex_unlock (&readylist_lock);
1855
1856 /* Tell one of the worker threads there is work to do. */
1857 if (do_signal)
1858 pthread_cond_signal (&readylist_cond);
1859 }
1860
1861
1862 /* Check whether restarting should happen. */
1863 static inline int
1864 restart_p (time_t now)
1865 {
1866 return (paranoia && readylist == NULL && nready == nthreads
1867 && now >= restart_time);
1868 }
1869
1870
1871 /* Array for times a connection was accepted. */
1872 static time_t *starttime;
1873
1874
1875 static void
1876 __attribute__ ((__noreturn__))
1877 main_loop_poll (void)
1878 {
1879 struct pollfd *conns = (struct pollfd *) xmalloc (nconns
1880 * sizeof (conns[0]));
1881
1882 conns[0].fd = sock;
1883 conns[0].events = POLLRDNORM;
1884 size_t nused = 1;
1885 size_t firstfree = 1;
1886
1887 #ifdef HAVE_INOTIFY
1888 if (inotify_fd != -1)
1889 {
1890 conns[1].fd = inotify_fd;
1891 conns[1].events = POLLRDNORM;
1892 nused = 2;
1893 firstfree = 2;
1894 }
1895 #endif
1896
1897 #ifdef HAVE_NETLINK
1898 size_t idx_nl_status_fd = 0;
1899 if (nl_status_fd != -1)
1900 {
1901 idx_nl_status_fd = nused;
1902 conns[nused].fd = nl_status_fd;
1903 conns[nused].events = POLLRDNORM;
1904 ++nused;
1905 firstfree = nused;
1906 }
1907 #endif
1908
1909 while (1)
1910 {
1911 /* Wait for any event. We wait at most a couple of seconds so
1912 that we can check whether we should close any of the accepted
1913 connections since we have not received a request. */
1914 #define MAX_ACCEPT_TIMEOUT 30
1915 #define MIN_ACCEPT_TIMEOUT 5
1916 #define MAIN_THREAD_TIMEOUT \
1917 (MAX_ACCEPT_TIMEOUT * 1000 \
1918 - ((MAX_ACCEPT_TIMEOUT - MIN_ACCEPT_TIMEOUT) * 1000 * nused) / (2 * nconns))
1919
1920 int n = poll (conns, nused, MAIN_THREAD_TIMEOUT);
1921
1922 time_t now = time (NULL);
1923
1924 /* If there is a descriptor ready for reading or there is a new
1925 connection, process this now. */
1926 if (n > 0)
1927 {
1928 if (conns[0].revents != 0)
1929 {
1930 /* We have a new incoming connection. Accept the connection. */
1931 int fd;
1932
1933 #ifndef __ASSUME_ACCEPT4
1934 fd = -1;
1935 if (have_accept4 >= 0)
1936 #endif
1937 {
1938 fd = TEMP_FAILURE_RETRY (accept4 (sock, NULL, NULL,
1939 SOCK_NONBLOCK));
1940 #ifndef __ASSUME_ACCEPT4
1941 if (have_accept4 == 0)
1942 have_accept4 = fd != -1 || errno != ENOSYS ? 1 : -1;
1943 #endif
1944 }
1945 #ifndef __ASSUME_ACCEPT4
1946 if (have_accept4 < 0)
1947 fd = TEMP_FAILURE_RETRY (accept (sock, NULL, NULL));
1948 #endif
1949
1950 /* Use the descriptor if we have not reached the limit. */
1951 if (fd >= 0)
1952 {
1953 if (firstfree < nconns)
1954 {
1955 conns[firstfree].fd = fd;
1956 conns[firstfree].events = POLLRDNORM;
1957 starttime[firstfree] = now;
1958 if (firstfree >= nused)
1959 nused = firstfree + 1;
1960
1961 do
1962 ++firstfree;
1963 while (firstfree < nused && conns[firstfree].fd != -1);
1964 }
1965 else
1966 /* We cannot use the connection so close it. */
1967 close (fd);
1968 }
1969
1970 --n;
1971 }
1972
1973 size_t first = 1;
1974 #ifdef HAVE_INOTIFY
1975 if (inotify_fd != -1 && conns[1].fd == inotify_fd)
1976 {
1977 if (conns[1].revents != 0)
1978 {
1979 bool to_clear[lastdb] = { false, };
1980 union
1981 {
1982 # ifndef PATH_MAX
1983 # define PATH_MAX 1024
1984 # endif
1985 struct inotify_event i;
1986 char buf[sizeof (struct inotify_event) + PATH_MAX];
1987 } inev;
1988
1989 while (1)
1990 {
1991 ssize_t nb = TEMP_FAILURE_RETRY (read (inotify_fd, &inev,
1992 sizeof (inev)));
1993 if (nb < (ssize_t) sizeof (struct inotify_event))
1994 {
1995 if (__builtin_expect (nb == -1 && errno != EAGAIN,
1996 0))
1997 {
1998 /* Something went wrong when reading the inotify
1999 data. Better disable inotify. */
2000 dbg_log (_("\
2001 disabled inotify after read error %d"),
2002 errno);
2003 conns[1].fd = -1;
2004 firstfree = 1;
2005 if (nused == 2)
2006 nused = 1;
2007 close (inotify_fd);
2008 inotify_fd = -1;
2009 }
2010 break;
2011 }
2012
2013 /* Check which of the files changed. */
2014 for (size_t dbcnt = 0; dbcnt < lastdb; ++dbcnt)
2015 {
2016 struct traced_file *finfo = dbs[dbcnt].traced_files;
2017
2018 while (finfo != NULL)
2019 {
2020 if (finfo->inotify_descr == inev.i.wd)
2021 {
2022 to_clear[dbcnt] = true;
2023 if (finfo->call_res_init)
2024 res_init ();
2025 goto next;
2026 }
2027
2028 finfo = finfo->next;
2029 }
2030 }
2031 next:;
2032 }
2033
2034 /* Actually perform the cache clearing. */
2035 for (size_t dbcnt = 0; dbcnt < lastdb; ++dbcnt)
2036 if (to_clear[dbcnt])
2037 {
2038 pthread_mutex_lock (&dbs[dbcnt].prune_lock);
2039 dbs[dbcnt].clear_cache = 1;
2040 pthread_mutex_unlock (&dbs[dbcnt].prune_lock);
2041 pthread_cond_signal (&dbs[dbcnt].prune_cond);
2042 }
2043
2044 --n;
2045 }
2046
2047 first = 2;
2048 }
2049 #endif
2050
2051 #ifdef HAVE_NETLINK
2052 if (idx_nl_status_fd != 0 && conns[idx_nl_status_fd].revents != 0)
2053 {
2054 char buf[4096];
2055 /* Read all the data. We do not interpret it here. */
2056 while (TEMP_FAILURE_RETRY (read (nl_status_fd, buf,
2057 sizeof (buf))) != -1)
2058 ;
2059
2060 dbs[hstdb].head->extra_data[NSCD_HST_IDX_CONF_TIMESTAMP]
2061 = __bump_nl_timestamp ();
2062 }
2063 #endif
2064
2065 for (size_t cnt = first; cnt < nused && n > 0; ++cnt)
2066 if (conns[cnt].revents != 0)
2067 {
2068 fd_ready (conns[cnt].fd);
2069
2070 /* Clean up the CONNS array. */
2071 conns[cnt].fd = -1;
2072 if (cnt < firstfree)
2073 firstfree = cnt;
2074 if (cnt == nused - 1)
2075 do
2076 --nused;
2077 while (conns[nused - 1].fd == -1);
2078
2079 --n;
2080 }
2081 }
2082
2083 /* Now find entries which have timed out. */
2084 assert (nused > 0);
2085
2086 /* We make the timeout length depend on the number of file
2087 descriptors currently used. */
2088 #define ACCEPT_TIMEOUT \
2089 (MAX_ACCEPT_TIMEOUT \
2090 - ((MAX_ACCEPT_TIMEOUT - MIN_ACCEPT_TIMEOUT) * nused) / nconns)
2091 time_t laststart = now - ACCEPT_TIMEOUT;
2092
2093 for (size_t cnt = nused - 1; cnt > 0; --cnt)
2094 {
2095 if (conns[cnt].fd != -1 && starttime[cnt] < laststart)
2096 {
2097 /* Remove the entry, it timed out. */
2098 (void) close (conns[cnt].fd);
2099 conns[cnt].fd = -1;
2100
2101 if (cnt < firstfree)
2102 firstfree = cnt;
2103 if (cnt == nused - 1)
2104 do
2105 --nused;
2106 while (conns[nused - 1].fd == -1);
2107 }
2108 }
2109
2110 if (restart_p (now))
2111 restart ();
2112 }
2113 }
2114
2115
2116 #ifdef HAVE_EPOLL
2117 static void
2118 main_loop_epoll (int efd)
2119 {
2120 struct epoll_event ev = { 0, };
2121 int nused = 1;
2122 size_t highest = 0;
2123
2124 /* Add the socket. */
2125 ev.events = EPOLLRDNORM;
2126 ev.data.fd = sock;
2127 if (epoll_ctl (efd, EPOLL_CTL_ADD, sock, &ev) == -1)
2128 /* We cannot use epoll. */
2129 return;
2130
2131 # ifdef HAVE_INOTIFY
2132 if (inotify_fd != -1)
2133 {
2134 ev.events = EPOLLRDNORM;
2135 ev.data.fd = inotify_fd;
2136 if (epoll_ctl (efd, EPOLL_CTL_ADD, inotify_fd, &ev) == -1)
2137 /* We cannot use epoll. */
2138 return;
2139 nused = 2;
2140 }
2141 # endif
2142
2143 # ifdef HAVE_NETLINK
2144 if (nl_status_fd != -1)
2145 {
2146 ev.events = EPOLLRDNORM;
2147 ev.data.fd = nl_status_fd;
2148 if (epoll_ctl (efd, EPOLL_CTL_ADD, nl_status_fd, &ev) == -1)
2149 /* We cannot use epoll. */
2150 return;
2151 }
2152 # endif
2153
2154 while (1)
2155 {
2156 struct epoll_event revs[100];
2157 # define nrevs (sizeof (revs) / sizeof (revs[0]))
2158
2159 int n = epoll_wait (efd, revs, nrevs, MAIN_THREAD_TIMEOUT);
2160
2161 time_t now = time (NULL);
2162
2163 for (int cnt = 0; cnt < n; ++cnt)
2164 if (revs[cnt].data.fd == sock)
2165 {
2166 /* A new connection. */
2167 int fd;
2168
2169 # ifndef __ASSUME_ACCEPT4
2170 fd = -1;
2171 if (have_accept4 >= 0)
2172 # endif
2173 {
2174 fd = TEMP_FAILURE_RETRY (accept4 (sock, NULL, NULL,
2175 SOCK_NONBLOCK));
2176 # ifndef __ASSUME_ACCEPT4
2177 if (have_accept4 == 0)
2178 have_accept4 = fd != -1 || errno != ENOSYS ? 1 : -1;
2179 # endif
2180 }
2181 # ifndef __ASSUME_ACCEPT4
2182 if (have_accept4 < 0)
2183 fd = TEMP_FAILURE_RETRY (accept (sock, NULL, NULL));
2184 # endif
2185
2186 /* Use the descriptor if we have not reached the limit. */
2187 if (fd >= 0)
2188 {
2189 /* Try to add the new descriptor. */
2190 ev.data.fd = fd;
2191 if (fd >= nconns
2192 || epoll_ctl (efd, EPOLL_CTL_ADD, fd, &ev) == -1)
2193 /* The descriptor is too large or something went
2194 wrong. Close the descriptor. */
2195 close (fd);
2196 else
2197 {
2198 /* Remember when we accepted the connection. */
2199 starttime[fd] = now;
2200
2201 if (fd > highest)
2202 highest = fd;
2203
2204 ++nused;
2205 }
2206 }
2207 }
2208 # ifdef HAVE_INOTIFY
2209 else if (revs[cnt].data.fd == inotify_fd)
2210 {
2211 bool to_clear[lastdb] = { false, };
2212 union
2213 {
2214 struct inotify_event i;
2215 char buf[sizeof (struct inotify_event) + PATH_MAX];
2216 } inev;
2217
2218 while (1)
2219 {
2220 ssize_t nb = TEMP_FAILURE_RETRY (read (inotify_fd, &inev,
2221 sizeof (inev)));
2222 if (nb < (ssize_t) sizeof (struct inotify_event))
2223 {
2224 if (__builtin_expect (nb == -1 && errno != EAGAIN, 0))
2225 {
2226 /* Something went wrong when reading the inotify
2227 data. Better disable inotify. */
2228 dbg_log (_("disabled inotify after read error %d"),
2229 errno);
2230 (void) epoll_ctl (efd, EPOLL_CTL_DEL, inotify_fd,
2231 NULL);
2232 close (inotify_fd);
2233 inotify_fd = -1;
2234 }
2235 break;
2236 }
2237
2238 /* Check which of the files changed. */
2239 for (size_t dbcnt = 0; dbcnt < lastdb; ++dbcnt)
2240 {
2241 struct traced_file *finfo = dbs[dbcnt].traced_files;
2242
2243 while (finfo != NULL)
2244 {
2245 if (finfo->inotify_descr == inev.i.wd)
2246 {
2247 to_clear[dbcnt] = true;
2248 if (finfo->call_res_init)
2249 res_init ();
2250 goto next;
2251 }
2252
2253 finfo = finfo->next;
2254 }
2255 }
2256 next:;
2257 }
2258
2259 /* Actually perform the cache clearing. */
2260 for (size_t dbcnt = 0; dbcnt < lastdb; ++dbcnt)
2261 if (to_clear[dbcnt])
2262 {
2263 pthread_mutex_lock (&dbs[dbcnt].prune_lock);
2264 dbs[dbcnt].clear_cache = 1;
2265 pthread_mutex_unlock (&dbs[dbcnt].prune_lock);
2266 pthread_cond_signal (&dbs[dbcnt].prune_cond);
2267 }
2268 }
2269 # endif
2270 # ifdef HAVE_NETLINK
2271 else if (revs[cnt].data.fd == nl_status_fd)
2272 {
2273 char buf[4096];
2274 /* Read all the data. We do not interpret it here. */
2275 while (TEMP_FAILURE_RETRY (read (nl_status_fd, buf,
2276 sizeof (buf))) != -1)
2277 ;
2278
2279 __bump_nl_timestamp ();
2280 }
2281 # endif
2282 else
2283 {
2284 /* Remove the descriptor from the epoll descriptor. */
2285 (void) epoll_ctl (efd, EPOLL_CTL_DEL, revs[cnt].data.fd, NULL);
2286
2287 /* Get a worker to handle the request. */
2288 fd_ready (revs[cnt].data.fd);
2289
2290 /* Reset the time. */
2291 starttime[revs[cnt].data.fd] = 0;
2292 if (revs[cnt].data.fd == highest)
2293 do
2294 --highest;
2295 while (highest > 0 && starttime[highest] == 0);
2296
2297 --nused;
2298 }
2299
2300 /* Now look for descriptors for accepted connections which have
2301 no reply in too long of a time. */
2302 time_t laststart = now - ACCEPT_TIMEOUT;
2303 assert (starttime[sock] == 0);
2304 assert (inotify_fd == -1 || starttime[inotify_fd] == 0);
2305 assert (nl_status_fd == -1 || starttime[nl_status_fd] == 0);
2306 for (int cnt = highest; cnt > STDERR_FILENO; --cnt)
2307 if (starttime[cnt] != 0 && starttime[cnt] < laststart)
2308 {
2309 /* We are waiting for this one for too long. Close it. */
2310 (void) epoll_ctl (efd, EPOLL_CTL_DEL, cnt, NULL);
2311
2312 (void) close (cnt);
2313
2314 starttime[cnt] = 0;
2315 if (cnt == highest)
2316 --highest;
2317 }
2318 else if (cnt != sock && starttime[cnt] == 0 && cnt == highest)
2319 --highest;
2320
2321 if (restart_p (now))
2322 restart ();
2323 }
2324 }
2325 #endif
2326
2327
2328 /* Start all the threads we want. The initial process is thread no. 1. */
2329 void
2330 start_threads (void)
2331 {
2332 /* Initialize the conditional variable we will use. The only
2333 non-standard attribute we might use is the clock selection. */
2334 pthread_condattr_t condattr;
2335 pthread_condattr_init (&condattr);
2336
2337 #if defined _POSIX_CLOCK_SELECTION && _POSIX_CLOCK_SELECTION >= 0 \
2338 && defined _POSIX_MONOTONIC_CLOCK && _POSIX_MONOTONIC_CLOCK >= 0
2339 /* Determine whether the monotonous clock is available. */
2340 struct timespec dummy;
2341 # if _POSIX_MONOTONIC_CLOCK == 0
2342 if (sysconf (_SC_MONOTONIC_CLOCK) > 0)
2343 # endif
2344 # if _POSIX_CLOCK_SELECTION == 0
2345 if (sysconf (_SC_CLOCK_SELECTION) > 0)
2346 # endif
2347 if (clock_getres (CLOCK_MONOTONIC, &dummy) == 0
2348 && pthread_condattr_setclock (&condattr, CLOCK_MONOTONIC) == 0)
2349 timeout_clock = CLOCK_MONOTONIC;
2350 #endif
2351
2352 /* Create the attribute for the threads. They are all created
2353 detached. */
2354 pthread_attr_init (&attr);
2355 pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
2356 /* Use 1MB stacks, twice as much for 64-bit architectures. */
2357 pthread_attr_setstacksize (&attr, NSCD_THREAD_STACKSIZE);
2358
2359 /* We allow less than LASTDB threads only for debugging. */
2360 if (debug_level == 0)
2361 nthreads = MAX (nthreads, lastdb);
2362
2363 /* Create the threads which prune the databases. */
2364 // XXX Ideally this work would be done by some of the worker threads.
2365 // XXX But this is problematic since we would need to be able to wake
2366 // XXX them up explicitly as well as part of the group handling the
2367 // XXX ready-list. This requires an operation where we can wait on
2368 // XXX two conditional variables at the same time. This operation
2369 // XXX does not exist (yet).
2370 for (long int i = 0; i < lastdb; ++i)
2371 {
2372 /* Initialize the conditional variable. */
2373 if (pthread_cond_init (&dbs[i].prune_cond, &condattr) != 0)
2374 {
2375 dbg_log (_("could not initialize conditional variable"));
2376 exit (1);
2377 }
2378
2379 pthread_t th;
2380 if (dbs[i].enabled
2381 && pthread_create (&th, &attr, nscd_run_prune, (void *) i) != 0)
2382 {
2383 dbg_log (_("could not start clean-up thread; terminating"));
2384 exit (1);
2385 }
2386 }
2387
2388 pthread_condattr_destroy (&condattr);
2389
2390 for (long int i = 0; i < nthreads; ++i)
2391 {
2392 pthread_t th;
2393 if (pthread_create (&th, &attr, nscd_run_worker, NULL) != 0)
2394 {
2395 if (i == 0)
2396 {
2397 dbg_log (_("could not start any worker thread; terminating"));
2398 exit (1);
2399 }
2400
2401 break;
2402 }
2403 }
2404
2405 /* Determine how much room for descriptors we should initially
2406 allocate. This might need to change later if we cap the number
2407 with MAXCONN. */
2408 const long int nfds = sysconf (_SC_OPEN_MAX);
2409 #define MINCONN 32
2410 #define MAXCONN 16384
2411 if (nfds == -1 || nfds > MAXCONN)
2412 nconns = MAXCONN;
2413 else if (nfds < MINCONN)
2414 nconns = MINCONN;
2415 else
2416 nconns = nfds;
2417
2418 /* We need memory to pass descriptors on to the worker threads. */
2419 fdlist = (struct fdlist *) xcalloc (nconns, sizeof (fdlist[0]));
2420 /* Array to keep track when connection was accepted. */
2421 starttime = (time_t *) xcalloc (nconns, sizeof (starttime[0]));
2422
2423 /* In the main thread we execute the loop which handles incoming
2424 connections. */
2425 #ifdef HAVE_EPOLL
2426 int efd = epoll_create (100);
2427 if (efd != -1)
2428 {
2429 main_loop_epoll (efd);
2430 close (efd);
2431 }
2432 #endif
2433
2434 main_loop_poll ();
2435 }
2436
2437
2438 /* Look up the uid, gid, and supplementary groups to run nscd as. When
2439 this function is called, we are not listening on the nscd socket yet so
2440 we can just use the ordinary lookup functions without causing a lockup */
2441 static void
2442 begin_drop_privileges (void)
2443 {
2444 struct passwd *pwd = getpwnam (server_user);
2445
2446 if (pwd == NULL)
2447 {
2448 dbg_log (_("Failed to run nscd as user '%s'"), server_user);
2449 error (EXIT_FAILURE, 0, _("Failed to run nscd as user '%s'"),
2450 server_user);
2451 }
2452
2453 server_uid = pwd->pw_uid;
2454 server_gid = pwd->pw_gid;
2455
2456 /* Save the old UID/GID if we have to change back. */
2457 if (paranoia)
2458 {
2459 old_uid = getuid ();
2460 old_gid = getgid ();
2461 }
2462
2463 if (getgrouplist (server_user, server_gid, NULL, &server_ngroups) == 0)
2464 {
2465 /* This really must never happen. */
2466 dbg_log (_("Failed to run nscd as user '%s'"), server_user);
2467 error (EXIT_FAILURE, errno, _("initial getgrouplist failed"));
2468 }
2469
2470 server_groups = (gid_t *) xmalloc (server_ngroups * sizeof (gid_t));
2471
2472 if (getgrouplist (server_user, server_gid, server_groups, &server_ngroups)
2473 == -1)
2474 {
2475 dbg_log (_("Failed to run nscd as user '%s'"), server_user);
2476 error (EXIT_FAILURE, errno, _("getgrouplist failed"));
2477 }
2478 }
2479
2480
2481 /* Call setgroups(), setgid(), and setuid() to drop root privileges and
2482 run nscd as the user specified in the configuration file. */
2483 static void
2484 finish_drop_privileges (void)
2485 {
2486 #if defined HAVE_LIBAUDIT && defined HAVE_LIBCAP
2487 /* We need to preserve the capabilities to connect to the audit daemon. */
2488 cap_t new_caps = preserve_capabilities ();
2489 #endif
2490
2491 if (setgroups (server_ngroups, server_groups) == -1)
2492 {
2493 dbg_log (_("Failed to run nscd as user '%s'"), server_user);
2494 error (EXIT_FAILURE, errno, _("setgroups failed"));
2495 }
2496
2497 int res;
2498 if (paranoia)
2499 res = setresgid (server_gid, server_gid, old_gid);
2500 else
2501 res = setgid (server_gid);
2502 if (res == -1)
2503 {
2504 dbg_log (_("Failed to run nscd as user '%s'"), server_user);
2505 perror ("setgid");
2506 exit (4);
2507 }
2508
2509 if (paranoia)
2510 res = setresuid (server_uid, server_uid, old_uid);
2511 else
2512 res = setuid (server_uid);
2513 if (res == -1)
2514 {
2515 dbg_log (_("Failed to run nscd as user '%s'"), server_user);
2516 perror ("setuid");
2517 exit (4);
2518 }
2519
2520 #if defined HAVE_LIBAUDIT && defined HAVE_LIBCAP
2521 /* Remove the temporary capabilities. */
2522 install_real_capabilities (new_caps);
2523 #endif
2524 }
sources.redhat.com git mirror of glibc CVS