/* Inner loops of cache daemon.
- Copyright (C) 1998-2003, 2004, 2005 Free Software Foundation, Inc.
+ Copyright (C) 1998-2016 Free Software Foundation, Inc.
This file is part of the GNU C Library.
Contributed by Ulrich Drepper <drepper@cygnus.com>, 1998.
- The GNU C Library is free software; you can redistribute it and/or
- modify it under the terms of the GNU Lesser General Public
- License as published by the Free Software Foundation; either
- version 2.1 of the License, or (at your option) any later version.
+ This program is free software; you can redistribute it and/or modify
+ it under the terms of the GNU General Public License as published
+ by the Free Software Foundation; version 2 of the License, or
+ (at your option) any later version.
- The GNU C Library is distributed in the hope that it will be useful,
+ This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- Lesser General Public License for more details.
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ GNU General Public License for more details.
- You should have received a copy of the GNU Lesser General Public
- License along with the GNU C Library; if not, write to the Free
- Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
- 02111-1307 USA. */
+ You should have received a copy of the GNU General Public License
+ along with this program; if not, see <http://www.gnu.org/licenses/>. */
#include <alloca.h>
#include <assert.h>
#include <errno.h>
#include <fcntl.h>
#include <grp.h>
+#include <ifaddrs.h>
#include <libintl.h>
#include <pthread.h>
#include <pwd.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
+#include <stdint.h>
#include <arpa/inet.h>
+#ifdef HAVE_NETLINK
+# include <linux/netlink.h>
+# include <linux/rtnetlink.h>
+#endif
#ifdef HAVE_EPOLL
# include <sys/epoll.h>
#endif
+#ifdef HAVE_INOTIFY
+# include <sys/inotify.h>
+#endif
#include <sys/mman.h>
#include <sys/param.h>
#include <sys/poll.h>
+#ifdef HAVE_SENDFILE
+# include <sys/sendfile.h>
+#endif
#include <sys/socket.h>
#include <sys/stat.h>
#include <sys/un.h>
#include "nscd.h"
#include "dbg_log.h"
#include "selinux.h"
+#include <resolv/resolv.h>
-
-/* Number of bytes of data we initially reserve for each hash table bucket. */
-#define DEFAULT_DATASIZE_PER_BUCKET 1024
+#include <kernel-features.h>
+#include <libc-internal.h>
-/* Wrapper functions with error checking for standard functions. */
-extern void *xmalloc (size_t n);
-extern void *xcalloc (size_t n, size_t s);
-extern void *xrealloc (void *o, size_t n);
-
/* Support to run nscd as an unprivileged user */
const char *server_user;
static uid_t server_uid;
static void finish_drop_privileges (void);
/* Map request type to a string. */
-const char *serv2str[LASTREQ] =
+const char *const serv2str[LASTREQ] =
{
[GETPWBYNAME] = "GETPWBYNAME",
[GETPWBYUID] = "GETPWBYUID",
[GETFDGR] = "GETFDGR",
[GETFDHST] = "GETFDHST",
[GETAI] = "GETAI",
- [INITGROUPS] = "INITGROUPS"
+ [INITGROUPS] = "INITGROUPS",
+ [GETSERVBYNAME] = "GETSERVBYNAME",
+ [GETSERVBYPORT] = "GETSERVBYPORT",
+ [GETFDSERV] = "GETFDSERV",
+ [GETNETGRENT] = "GETNETGRENT",
+ [INNETGR] = "INNETGR",
+ [GETFDNETGR] = "GETFDNETGR"
};
/* The control data structures for the services. */
{
[pwddb] = {
.lock = PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP,
+ .prune_lock = PTHREAD_MUTEX_INITIALIZER,
+ .prune_run_lock = PTHREAD_MUTEX_INITIALIZER,
.enabled = 0,
.check_file = 1,
.persistent = 0,
+ .propagate = 1,
.shared = 0,
- .filename = "/etc/passwd",
+ .max_db_size = DEFAULT_MAX_DB_SIZE,
+ .suggested_module = DEFAULT_SUGGESTED_MODULE,
.db_filename = _PATH_NSCD_PASSWD_DB,
.disabled_iov = &pwd_iov_disabled,
.postimeout = 3600,
},
[grpdb] = {
.lock = PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP,
+ .prune_lock = PTHREAD_MUTEX_INITIALIZER,
+ .prune_run_lock = PTHREAD_MUTEX_INITIALIZER,
.enabled = 0,
.check_file = 1,
.persistent = 0,
+ .propagate = 1,
.shared = 0,
- .filename = "/etc/group",
+ .max_db_size = DEFAULT_MAX_DB_SIZE,
+ .suggested_module = DEFAULT_SUGGESTED_MODULE,
.db_filename = _PATH_NSCD_GROUP_DB,
.disabled_iov = &grp_iov_disabled,
.postimeout = 3600,
},
[hstdb] = {
.lock = PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP,
+ .prune_lock = PTHREAD_MUTEX_INITIALIZER,
+ .prune_run_lock = PTHREAD_MUTEX_INITIALIZER,
.enabled = 0,
.check_file = 1,
.persistent = 0,
+ .propagate = 0, /* Not used. */
.shared = 0,
- .filename = "/etc/hosts",
+ .max_db_size = DEFAULT_MAX_DB_SIZE,
+ .suggested_module = DEFAULT_SUGGESTED_MODULE,
.db_filename = _PATH_NSCD_HOSTS_DB,
.disabled_iov = &hst_iov_disabled,
.postimeout = 3600,
.wr_fd = -1,
.ro_fd = -1,
.mmap_used = false
+ },
+ [servdb] = {
+ .lock = PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP,
+ .prune_lock = PTHREAD_MUTEX_INITIALIZER,
+ .prune_run_lock = PTHREAD_MUTEX_INITIALIZER,
+ .enabled = 0,
+ .check_file = 1,
+ .persistent = 0,
+ .propagate = 0, /* Not used. */
+ .shared = 0,
+ .max_db_size = DEFAULT_MAX_DB_SIZE,
+ .suggested_module = DEFAULT_SUGGESTED_MODULE,
+ .db_filename = _PATH_NSCD_SERVICES_DB,
+ .disabled_iov = &serv_iov_disabled,
+ .postimeout = 28800,
+ .negtimeout = 20,
+ .wr_fd = -1,
+ .ro_fd = -1,
+ .mmap_used = false
+ },
+ [netgrdb] = {
+ .lock = PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP,
+ .prune_lock = PTHREAD_MUTEX_INITIALIZER,
+ .prune_run_lock = PTHREAD_MUTEX_INITIALIZER,
+ .enabled = 0,
+ .check_file = 1,
+ .persistent = 0,
+ .propagate = 0, /* Not used. */
+ .shared = 0,
+ .max_db_size = DEFAULT_MAX_DB_SIZE,
+ .suggested_module = DEFAULT_SUGGESTED_MODULE,
+ .db_filename = _PATH_NSCD_NETGROUP_DB,
+ .disabled_iov = &netgroup_iov_disabled,
+ .postimeout = 28800,
+ .negtimeout = 20,
+ .wr_fd = -1,
+ .ro_fd = -1,
+ .mmap_used = false
}
};
/* Mapping of request type to database. */
-static struct database_dyn *const serv2db[LASTREQ] =
+static struct
+{
+ bool data_request;
+ struct database_dyn *db;
+} const reqinfo[LASTREQ] =
{
- [GETPWBYNAME] = &dbs[pwddb],
- [GETPWBYUID] = &dbs[pwddb],
- [GETGRBYNAME] = &dbs[grpdb],
- [GETGRBYGID] = &dbs[grpdb],
- [GETHOSTBYNAME] = &dbs[hstdb],
- [GETHOSTBYNAMEv6] = &dbs[hstdb],
- [GETHOSTBYADDR] = &dbs[hstdb],
- [GETHOSTBYADDRv6] = &dbs[hstdb],
- [GETFDPW] = &dbs[pwddb],
- [GETFDGR] = &dbs[grpdb],
- [GETFDHST] = &dbs[hstdb],
- [GETAI] = &dbs[hstdb],
- [INITGROUPS] = &dbs[grpdb]
+ [GETPWBYNAME] = { true, &dbs[pwddb] },
+ [GETPWBYUID] = { true, &dbs[pwddb] },
+ [GETGRBYNAME] = { true, &dbs[grpdb] },
+ [GETGRBYGID] = { true, &dbs[grpdb] },
+ [GETHOSTBYNAME] = { true, &dbs[hstdb] },
+ [GETHOSTBYNAMEv6] = { true, &dbs[hstdb] },
+ [GETHOSTBYADDR] = { true, &dbs[hstdb] },
+ [GETHOSTBYADDRv6] = { true, &dbs[hstdb] },
+ [SHUTDOWN] = { false, NULL },
+ [GETSTAT] = { false, NULL },
+ [SHUTDOWN] = { false, NULL },
+ [GETFDPW] = { false, &dbs[pwddb] },
+ [GETFDGR] = { false, &dbs[grpdb] },
+ [GETFDHST] = { false, &dbs[hstdb] },
+ [GETAI] = { true, &dbs[hstdb] },
+ [INITGROUPS] = { true, &dbs[grpdb] },
+ [GETSERVBYNAME] = { true, &dbs[servdb] },
+ [GETSERVBYPORT] = { true, &dbs[servdb] },
+ [GETFDSERV] = { false, &dbs[servdb] },
+ [GETNETGRENT] = { true, &dbs[netgrdb] },
+ [INNETGR] = { true, &dbs[netgrdb] },
+ [GETFDNETGR] = { false, &dbs[netgrdb] }
};
-/* Number of seconds between two cache pruning runs. */
-#define CACHE_PRUNE_INTERVAL 15
-
-
/* Initial number of threads to use. */
int nthreads = -1;
/* Maximum number of threads to use. */
/* Socket for incoming connections. */
static int sock;
+#ifdef HAVE_INOTIFY
+/* Inotify descriptor. */
+int inotify_fd = -1;
+#endif
+
+#ifdef HAVE_NETLINK
+/* Descriptor for netlink status updates. */
+static int nl_status_fd = -1;
+#endif
+
+#ifndef __ASSUME_ACCEPT4
+static int have_accept4;
+#endif
+
/* Number of times clients had to wait. */
unsigned long int client_queued;
-/* Initialize database information structures. */
-void
-nscd_init (void)
+ssize_t
+writeall (int fd, const void *buf, size_t len)
{
- struct sockaddr_un sock_addr;
- size_t cnt;
+ size_t n = len;
+ ssize_t ret;
+ do
+ {
+ ret = TEMP_FAILURE_RETRY (send (fd, buf, n, MSG_NOSIGNAL));
+ if (ret <= 0)
+ break;
+ buf = (const char *) buf + ret;
+ n -= ret;
+ }
+ while (n > 0);
+ return ret < 0 ? ret : len - n;
+}
+
- /* Secure mode and unprivileged mode are incompatible */
- if (server_user != NULL && secure_in_use)
+#ifdef HAVE_SENDFILE
+ssize_t
+sendfileall (int tofd, int fromfd, off_t off, size_t len)
+{
+ ssize_t n = len;
+ ssize_t ret;
+
+ do
{
- dbg_log (_("Cannot run nscd in secure mode as unprivileged user"));
- exit (1);
+ ret = TEMP_FAILURE_RETRY (sendfile (tofd, fromfd, &off, n));
+ if (ret <= 0)
+ break;
+ n -= ret;
}
+ while (n > 0);
+ return ret < 0 ? ret : len - n;
+}
+#endif
+
+
+enum usekey
+ {
+ use_not = 0,
+ /* The following three are not really used, they are symbolic constants. */
+ use_first = 16,
+ use_begin = 32,
+ use_end = 64,
+
+ use_he = 1,
+ use_he_begin = use_he | use_begin,
+ use_he_end = use_he | use_end,
+ use_data = 3,
+ use_data_begin = use_data | use_begin,
+ use_data_end = use_data | use_end,
+ use_data_first = use_data_begin | use_first
+ };
+
+static int
+check_use (const char *data, nscd_ssize_t first_free, uint8_t *usemap,
+ enum usekey use, ref_t start, size_t len)
+{
+ assert (len >= 2);
+
+ if (start > first_free || start + len > first_free
+ || (start & BLOCK_ALIGN_M1))
+ return 0;
+
+ if (usemap[start] == use_not)
+ {
+ /* Add the start marker. */
+ usemap[start] = use | use_begin;
+ use &= ~use_first;
+
+ while (--len > 0)
+ if (usemap[++start] != use_not)
+ return 0;
+ else
+ usemap[start] = use;
+
+ /* Add the end marker. */
+ usemap[start] = use | use_end;
+ }
+ else if ((usemap[start] & ~use_first) == ((use | use_begin) & ~use_first))
+ {
+ /* Hash entries can't be shared. */
+ if (use == use_he)
+ return 0;
+
+ usemap[start] |= (use & use_first);
+ use &= ~use_first;
+
+ while (--len > 1)
+ if (usemap[++start] != use)
+ return 0;
+
+ if (usemap[++start] != (use | use_end))
+ return 0;
+ }
+ else
+ /* Points to a wrong object or somewhere in the middle. */
+ return 0;
+
+ return 1;
+}
+
+
+/* Verify data in persistent database. */
+static int
+verify_persistent_db (void *mem, struct database_pers_head *readhead, int dbnr)
+{
+ assert (dbnr == pwddb || dbnr == grpdb || dbnr == hstdb || dbnr == servdb
+ || dbnr == netgrdb);
+
+ time_t now = time (NULL);
+
+ struct database_pers_head *head = mem;
+ struct database_pers_head head_copy = *head;
+
+ /* Check that the header that was read matches the head in the database. */
+ if (memcmp (head, readhead, sizeof (*head)) != 0)
+ return 0;
+
+ /* First some easy tests: make sure the database header is sane. */
+ if (head->version != DB_VERSION
+ || head->header_size != sizeof (*head)
+ /* We allow a timestamp to be one hour ahead of the current time.
+ This should cover daylight saving time changes. */
+ || head->timestamp > now + 60 * 60 + 60
+ || (head->gc_cycle & 1)
+ || head->module == 0
+ || (size_t) head->module > INT32_MAX / sizeof (ref_t)
+ || (size_t) head->data_size > INT32_MAX - head->module * sizeof (ref_t)
+ || head->first_free < 0
+ || head->first_free > head->data_size
+ || (head->first_free & BLOCK_ALIGN_M1) != 0
+ || head->maxnentries < 0
+ || head->maxnsearched < 0)
+ return 0;
+
+ uint8_t *usemap = calloc (head->first_free, 1);
+ if (usemap == NULL)
+ return 0;
+
+ const char *data = (char *) &head->array[roundup (head->module,
+ ALIGN / sizeof (ref_t))];
+
+ nscd_ssize_t he_cnt = 0;
+ for (nscd_ssize_t cnt = 0; cnt < head->module; ++cnt)
+ {
+ ref_t trail = head->array[cnt];
+ ref_t work = trail;
+ int tick = 0;
+
+ while (work != ENDREF)
+ {
+ if (! check_use (data, head->first_free, usemap, use_he, work,
+ sizeof (struct hashentry)))
+ goto fail;
+
+ /* Now we know we can dereference the record. */
+ struct hashentry *here = (struct hashentry *) (data + work);
+
+ ++he_cnt;
+
+ /* Make sure the record is for this type of service. */
+ if (here->type >= LASTREQ
+ || reqinfo[here->type].db != &dbs[dbnr])
+ goto fail;
+
+ /* Validate boolean field value. */
+ if (here->first != false && here->first != true)
+ goto fail;
+
+ if (here->len < 0)
+ goto fail;
+
+ /* Now the data. */
+ if (here->packet < 0
+ || here->packet > head->first_free
+ || here->packet + sizeof (struct datahead) > head->first_free)
+ goto fail;
+
+ struct datahead *dh = (struct datahead *) (data + here->packet);
+
+ if (! check_use (data, head->first_free, usemap,
+ use_data | (here->first ? use_first : 0),
+ here->packet, dh->allocsize))
+ goto fail;
+
+ if (dh->allocsize < sizeof (struct datahead)
+ || dh->recsize > dh->allocsize
+ || (dh->notfound != false && dh->notfound != true)
+ || (dh->usable != false && dh->usable != true))
+ goto fail;
+
+ if (here->key < here->packet + sizeof (struct datahead)
+ || here->key > here->packet + dh->allocsize
+ || here->key + here->len > here->packet + dh->allocsize)
+ goto fail;
+
+ work = here->next;
+
+ if (work == trail)
+ /* A circular list, this must not happen. */
+ goto fail;
+ if (tick)
+ trail = ((struct hashentry *) (data + trail))->next;
+ tick = 1 - tick;
+ }
+ }
+
+ if (he_cnt != head->nentries)
+ goto fail;
+
+ /* See if all data and keys had at least one reference from
+ he->first == true hashentry. */
+ for (ref_t idx = 0; idx < head->first_free; ++idx)
+ {
+ if (usemap[idx] == use_data_begin)
+ goto fail;
+ }
+
+ /* Finally, make sure the database hasn't changed since the first test. */
+ if (memcmp (mem, &head_copy, sizeof (*head)) != 0)
+ goto fail;
+
+ free (usemap);
+ return 1;
+
+fail:
+ free (usemap);
+ return 0;
+}
+
+
+#ifdef O_CLOEXEC
+# define EXTRA_O_FLAGS O_CLOEXEC
+#else
+# define EXTRA_O_FLAGS 0
+#endif
+
+
+/* Initialize database information structures. */
+void
+nscd_init (void)
+{
/* Look up unprivileged uid/gid/groups before we start listening on the
socket */
if (server_user != NULL)
if (nthreads == -1)
/* No configuration for this value, assume a default. */
- nthreads = 2 * lastdb;
+ nthreads = 4;
- for (cnt = 0; cnt < lastdb; ++cnt)
+ for (size_t cnt = 0; cnt < lastdb; ++cnt)
if (dbs[cnt].enabled)
{
pthread_rwlock_init (&dbs[cnt].lock, NULL);
if (dbs[cnt].persistent)
{
/* Try to open the appropriate file on disk. */
- int fd = open (dbs[cnt].db_filename, O_RDWR);
+ int fd = open (dbs[cnt].db_filename, O_RDWR | EXTRA_O_FLAGS);
if (fd != -1)
{
+ char *msg = NULL;
struct stat64 st;
void *mem;
size_t total;
sizeof (head)));
if (n != sizeof (head) || fstat64 (fd, &st) != 0)
{
+ fail_db_errno:
+ /* The code is single-threaded at this point so
+ using strerror is just fine. */
+ msg = strerror (errno);
fail_db:
dbg_log (_("invalid persistent database file \"%s\": %s"),
- dbs[cnt].db_filename, strerror (errno));
- dbs[cnt].persistent = 0;
+ dbs[cnt].db_filename, msg);
+ unlink (dbs[cnt].db_filename);
}
else if (head.module == 0 && head.data_size == 0)
{
- /* The file has been created, but the head has not been
- initialized yet. Remove the old file. */
- unlink (dbs[cnt].db_filename);
+ /* The file has been created, but the head has not
+ been initialized yet. */
+ msg = _("uninitialized header");
+ goto fail_db;
}
else if (head.header_size != (int) sizeof (head))
{
- dbg_log (_("invalid persistent database file \"%s\": %s"),
- dbs[cnt].db_filename,
- _("header size does not match"));
- dbs[cnt].persistent = 0;
+ msg = _("header size does not match");
+ goto fail_db;
}
else if ((total = (sizeof (head)
+ roundup (head.module * sizeof (ref_t),
ALIGN)
+ head.data_size))
- > st.st_size)
+ > st.st_size
+ || total < sizeof (head))
{
- dbg_log (_("invalid persistent database file \"%s\": %s"),
- dbs[cnt].db_filename,
- _("file size does not match"));
- dbs[cnt].persistent = 0;
+ msg = _("file size does not match");
+ goto fail_db;
+ }
+ /* Note we map with the maximum size allowed for the
+ database. This is likely much larger than the
+ actual file size. This is OK on most OSes since
+ extensions of the underlying file will
+ automatically translate more pages available for
+ memory access. */
+ else if ((mem = mmap (NULL, dbs[cnt].max_db_size,
+ PROT_READ | PROT_WRITE,
+ MAP_SHARED, fd, 0))
+ == MAP_FAILED)
+ goto fail_db_errno;
+ else if (!verify_persistent_db (mem, &head, cnt))
+ {
+ munmap (mem, total);
+ msg = _("verification failed");
+ goto fail_db;
}
- else if ((mem = mmap (NULL, total, PROT_READ | PROT_WRITE,
- MAP_SHARED, fd, 0)) == MAP_FAILED)
- goto fail_db;
else
{
/* Success. We have the database. */
/* We also need a read-only descriptor. */
if (dbs[cnt].shared)
{
- dbs[cnt].ro_fd = open (dbs[cnt].db_filename, O_RDONLY);
+ dbs[cnt].ro_fd = open (dbs[cnt].db_filename,
+ O_RDONLY | EXTRA_O_FLAGS);
if (dbs[cnt].ro_fd == -1)
dbg_log (_("\
cannot create read-only descriptor for \"%s\"; no mmap"),
if (fd != -1)
close (fd);
}
+ else if (errno == EACCES)
+ do_exit (EXIT_FAILURE, 0, _("cannot access '%s'"),
+ dbs[cnt].db_filename);
}
if (dbs[cnt].head == NULL)
if (dbs[cnt].persistent)
{
fd = open (dbs[cnt].db_filename,
- O_RDWR | O_CREAT | O_EXCL | O_TRUNC,
+ O_RDWR | O_CREAT | O_EXCL | O_TRUNC | EXTRA_O_FLAGS,
S_IRUSR | S_IWUSR);
if (fd != -1 && dbs[cnt].shared)
- ro_fd = open (dbs[cnt].db_filename, O_RDONLY);
+ ro_fd = open (dbs[cnt].db_filename,
+ O_RDONLY | EXTRA_O_FLAGS);
}
else
{
char fname[] = _PATH_NSCD_XYZ_DB_TMP;
- fd = mkstemp (fname);
+ fd = mkostemp (fname, EXTRA_O_FLAGS);
/* We do not need the file name anymore after we
opened another file descriptor in read-only mode. */
if (fd != -1)
{
if (dbs[cnt].shared)
- ro_fd = open (fname, O_RDONLY);
+ ro_fd = open (fname, O_RDONLY | EXTRA_O_FLAGS);
unlink (fname);
}
{
dbg_log (_("database for %s corrupted or simultaneously used; remove %s manually if necessary and restart"),
dbnames[cnt], dbs[cnt].db_filename);
- // XXX Correct way to terminate?
- exit (1);
+ do_exit (1, 0, NULL);
}
if (dbs[cnt].persistent)
cannot create read-only descriptor for \"%s\"; no mmap"),
dbs[cnt].db_filename);
- /* Before we create the header, initialiye the hash
- table. So that if we get interrupted if writing
+ /* Before we create the header, initialize the hash
+ table. That way if we get interrupted while writing
the header we can recognize a partially initialized
database. */
size_t ps = sysconf (_SC_PAGESIZE);
if (offset % ps != 0)
{
towrite = MIN (remaining, ps - (offset % ps));
- pwrite (fd, tmpbuf, towrite, offset);
+ if (pwrite (fd, tmpbuf, towrite, offset) != towrite)
+ goto write_fail;
offset += towrite;
remaining -= towrite;
}
while (remaining > ps)
{
- pwrite (fd, tmpbuf, ps, offset);
+ if (pwrite (fd, tmpbuf, ps, offset) == -1)
+ goto write_fail;
offset += ps;
remaining -= ps;
}
- if (remaining > 0)
- pwrite (fd, tmpbuf, remaining, offset);
+ if (remaining > 0
+ && pwrite (fd, tmpbuf, remaining, offset) != remaining)
+ goto write_fail;
/* Create the header of the file. */
struct database_pers_head head =
if ((TEMP_FAILURE_RETRY (write (fd, &head, sizeof (head)))
!= sizeof (head))
- || ftruncate (fd, total) != 0
- || (mem = mmap (NULL, total, PROT_READ | PROT_WRITE,
+ || (TEMP_FAILURE_RETRY_VAL (posix_fallocate (fd, 0, total))
+ != 0)
+ || (mem = mmap (NULL, dbs[cnt].max_db_size,
+ PROT_READ | PROT_WRITE,
MAP_SHARED, fd, 0)) == MAP_FAILED)
{
+ write_fail:
unlink (dbs[cnt].db_filename);
dbg_log (_("cannot write to database file %s: %s"),
dbs[cnt].db_filename, strerror (errno));
}
}
+#if !defined O_CLOEXEC || !defined __ASSUME_O_CLOEXEC
+ /* We do not check here whether the O_CLOEXEC provided to the
+ open call was successful or not. The two fcntl calls are
+ only performed once each per process start-up and therefore
+ is not noticeable at all. */
if (paranoia
&& ((dbs[cnt].wr_fd != -1
&& fcntl (dbs[cnt].wr_fd, F_SETFD, FD_CLOEXEC) == -1)
strerror (errno));
paranoia = 0;
}
+#endif
if (dbs[cnt].head == NULL)
{
dbs[cnt].head = xmalloc (sizeof (struct database_pers_head)
+ (dbs[cnt].suggested_module
* sizeof (ref_t)));
- memset (dbs[cnt].head, '\0', sizeof (dbs[cnt].head));
+ memset (dbs[cnt].head, '\0', sizeof (struct database_pers_head));
assert (~ENDREF == 0);
memset (dbs[cnt].head->array, '\xff',
dbs[cnt].suggested_module * sizeof (ref_t));
dbs[cnt].shared = 0;
assert (dbs[cnt].ro_fd == -1);
}
-
- if (dbs[cnt].check_file)
- {
- /* We need the modification date of the file. */
- struct stat st;
-
- if (stat (dbs[cnt].filename, &st) < 0)
- {
- /* We cannot stat() the file, disable file checking. */
- dbg_log (_("cannot stat() file `%s': %s"),
- dbs[cnt].filename, strerror (errno));
- dbs[cnt].check_file = 0;
- }
- else
- dbs[cnt].file_mtime = st.st_mtime;
- }
}
/* Create the socket. */
- sock = socket (AF_UNIX, SOCK_STREAM, 0);
+ sock = socket (AF_UNIX, SOCK_STREAM | SOCK_CLOEXEC | SOCK_NONBLOCK, 0);
if (sock < 0)
{
dbg_log (_("cannot open socket: %s"), strerror (errno));
- exit (1);
+ do_exit (errno == EACCES ? 4 : 1, 0, NULL);
}
/* Bind a name to the socket. */
+ struct sockaddr_un sock_addr;
sock_addr.sun_family = AF_UNIX;
strcpy (sock_addr.sun_path, _PATH_NSCDSOCKET);
if (bind (sock, (struct sockaddr *) &sock_addr, sizeof (sock_addr)) < 0)
{
dbg_log ("%s: %s", _PATH_NSCDSOCKET, strerror (errno));
- exit (1);
- }
-
- /* We don't want to get stuck on accept. */
- int fl = fcntl (sock, F_GETFL);
- if (fl == -1 || fcntl (sock, F_SETFL, fl | O_NONBLOCK) == -1)
- {
- dbg_log (_("cannot change socket to nonblocking mode: %s"),
- strerror (errno));
- exit (1);
- }
-
- /* The descriptor needs to be closed on exec. */
- if (paranoia && fcntl (sock, F_SETFD, FD_CLOEXEC) == -1)
- {
- dbg_log (_("cannot set socket to close on exec: %s"),
- strerror (errno));
- exit (1);
+ do_exit (errno == EACCES ? 4 : 1, 0, NULL);
}
/* Set permissions for the socket. */
{
dbg_log (_("cannot enable socket to accept connections: %s"),
strerror (errno));
- exit (1);
+ do_exit (1, 0, NULL);
+ }
+
+#ifdef HAVE_NETLINK
+ if (dbs[hstdb].enabled)
+ {
+ /* Try to open netlink socket to monitor network setting changes. */
+ nl_status_fd = socket (AF_NETLINK,
+ SOCK_RAW | SOCK_CLOEXEC | SOCK_NONBLOCK,
+ NETLINK_ROUTE);
+ if (nl_status_fd != -1)
+ {
+ struct sockaddr_nl snl;
+ memset (&snl, '\0', sizeof (snl));
+ snl.nl_family = AF_NETLINK;
+ /* XXX Is this the best set to use? */
+ snl.nl_groups = (RTMGRP_IPV4_IFADDR | RTMGRP_TC | RTMGRP_IPV4_MROUTE
+ | RTMGRP_IPV4_ROUTE | RTMGRP_IPV4_RULE
+ | RTMGRP_IPV6_IFADDR | RTMGRP_IPV6_MROUTE
+ | RTMGRP_IPV6_ROUTE | RTMGRP_IPV6_IFINFO
+ | RTMGRP_IPV6_PREFIX);
+
+ if (bind (nl_status_fd, (struct sockaddr *) &snl, sizeof (snl)) != 0)
+ {
+ close (nl_status_fd);
+ nl_status_fd = -1;
+ }
+ else
+ {
+ /* Start the timestamp process. */
+ dbs[hstdb].head->extra_data[NSCD_HST_IDX_CONF_TIMESTAMP]
+ = __bump_nl_timestamp ();
+ }
+ }
}
+#endif
- /* Change to unprivileged uid/gid/groups if specifed in config file */
+ /* Change to unprivileged uid/gid/groups if specified in config file */
if (server_user != NULL)
finish_drop_privileges ();
}
+#ifdef HAVE_INOTIFY
+#define TRACED_FILE_MASK (IN_DELETE_SELF | IN_CLOSE_WRITE | IN_MOVE_SELF)
+#define TRACED_DIR_MASK (IN_DELETE_SELF | IN_CREATE | IN_MOVED_TO | IN_MOVE_SELF)
+void
+install_watches (struct traced_file *finfo)
+{
+ /* Use inotify support if we have it. */
+ if (finfo->inotify_descr[TRACED_FILE] < 0)
+ finfo->inotify_descr[TRACED_FILE] = inotify_add_watch (inotify_fd,
+ finfo->fname,
+ TRACED_FILE_MASK);
+ if (finfo->inotify_descr[TRACED_FILE] < 0)
+ {
+ dbg_log (_("disabled inotify-based monitoring for file `%s': %s"),
+ finfo->fname, strerror (errno));
+ return;
+ }
+ dbg_log (_("monitoring file `%s` (%d)"),
+ finfo->fname, finfo->inotify_descr[TRACED_FILE]);
+ /* Additionally listen for events in the file's parent directory.
+ We do this because the file to be watched might be
+ deleted and then added back again. When it is added back again
+ we must re-add the watch. We must also cover IN_MOVED_TO to
+ detect a file being moved into the directory. */
+ if (finfo->inotify_descr[TRACED_DIR] < 0)
+ finfo->inotify_descr[TRACED_DIR] = inotify_add_watch (inotify_fd,
+ finfo->dname,
+ TRACED_DIR_MASK);
+ if (finfo->inotify_descr[TRACED_DIR] < 0)
+ {
+ dbg_log (_("disabled inotify-based monitoring for directory `%s': %s"),
+ finfo->fname, strerror (errno));
+ return;
+ }
+ dbg_log (_("monitoring directory `%s` (%d)"),
+ finfo->dname, finfo->inotify_descr[TRACED_DIR]);
+}
+#endif
+
+/* Register the file in FINFO as a traced file for the database DBS[DBIX].
+
+ We support registering multiple files per database. Each call to
+ register_traced_file adds to the list of registered files.
+
+ When we prune the database, either through timeout or a request to
+ invalidate, we will check to see if any of the registered files has changed.
+ When we accept new connections to handle a cache request we will also
+ check to see if any of the registered files has changed.
+
+ If we have inotify support then we install an inotify fd to notify us of
+ file deletion or modification, both of which will require we invalidate
+ the cache for the database. Without inotify support we stat the file and
+ store st_mtime to determine if the file has been modified. */
+void
+register_traced_file (size_t dbidx, struct traced_file *finfo)
+{
+ /* If the database is disabled or file checking is disabled
+ then ignore the registration. */
+ if (! dbs[dbidx].enabled || ! dbs[dbidx].check_file)
+ return;
+
+ if (__glibc_unlikely (debug_level > 0))
+ dbg_log (_("monitoring file %s for database %s"),
+ finfo->fname, dbnames[dbidx]);
+
+#ifdef HAVE_INOTIFY
+ install_watches (finfo);
+#endif
+ struct stat64 st;
+ if (stat64 (finfo->fname, &st) < 0)
+ {
+ /* We cannot stat() the file. Set mtime to zero and try again later. */
+ dbg_log (_("stat failed for file `%s'; will try again later: %s"),
+ finfo->fname, strerror (errno));
+ finfo->mtime = 0;
+ }
+ else
+ finfo->mtime = st.st_mtime;
+
+ /* Queue up the file name. */
+ finfo->next = dbs[dbidx].traced_files;
+ dbs[dbidx].traced_files = finfo;
+}
+
/* Close the connections. */
void
static void
-invalidate_cache (char *key)
+invalidate_cache (char *key, int fd)
{
dbtype number;
+ int32_t resp;
- if (strcmp (key, "passwd") == 0)
- number = pwddb;
- else if (strcmp (key, "group") == 0)
- number = grpdb;
- else if (__builtin_expect (strcmp (key, "hosts"), 0) == 0)
- {
- number = hstdb;
+ for (number = pwddb; number < lastdb; ++number)
+ if (strcmp (key, dbnames[number]) == 0)
+ {
+ struct traced_file *runp = dbs[number].traced_files;
+ while (runp != NULL)
+ {
+ /* Make sure we reload from file when checking mtime. */
+ runp->mtime = 0;
+#ifdef HAVE_INOTIFY
+ /* During an invalidation we try to reload the traced
+ file watches. This allows the user to re-sync if
+ inotify events were lost. Similar to what we do during
+ pruning. */
+ install_watches (runp);
+#endif
+ if (runp->call_res_init)
+ {
+ res_init ();
+ break;
+ }
+ runp = runp->next;
+ }
+ break;
+ }
- /* Re-initialize the resolver. resolv.conf might have changed. */
- res_init ();
+ if (number == lastdb)
+ {
+ resp = EINVAL;
+ writeall (fd, &resp, sizeof (resp));
+ return;
}
- else
- return;
if (dbs[number].enabled)
- prune_cache (&dbs[number], LONG_MAX);
+ {
+ pthread_mutex_lock (&dbs[number].prune_run_lock);
+ prune_cache (&dbs[number], LONG_MAX, fd);
+ pthread_mutex_unlock (&dbs[number].prune_run_lock);
+ }
+ else
+ {
+ resp = 0;
+ writeall (fd, &resp, sizeof (resp));
+ }
}
return;
/* We need to send some data along with the descriptor. */
- struct iovec iov[1];
+ uint64_t mapsize = (db->head->data_size
+ + roundup (db->head->module * sizeof (ref_t), ALIGN)
+ + sizeof (struct database_pers_head));
+ struct iovec iov[2];
iov[0].iov_base = key;
iov[0].iov_len = strlen (key) + 1;
+ iov[1].iov_base = &mapsize;
+ iov[1].iov_len = sizeof (mapsize);
/* Prepare the control message to transfer the descriptor. */
- char buf[CMSG_SPACE (sizeof (int))];
- struct msghdr msg = { .msg_iov = iov, .msg_iovlen = 1,
- .msg_control = buf, .msg_controllen = sizeof (buf) };
+ union
+ {
+ struct cmsghdr hdr;
+ char bytes[CMSG_SPACE (sizeof (int))];
+ } buf;
+ struct msghdr msg = { .msg_iov = iov, .msg_iovlen = 2,
+ .msg_control = buf.bytes,
+ .msg_controllen = sizeof (buf) };
struct cmsghdr *cmsg = CMSG_FIRSTHDR (&msg);
cmsg->cmsg_level = SOL_SOCKET;
cmsg->cmsg_type = SCM_RIGHTS;
cmsg->cmsg_len = CMSG_LEN (sizeof (int));
- *(int *) CMSG_DATA (cmsg) = db->ro_fd;
+ int *ip = (int *) CMSG_DATA (cmsg);
+ *ip = db->ro_fd;
msg.msg_controllen = cmsg->cmsg_len;
#endif
(void) TEMP_FAILURE_RETRY (sendmsg (fd, &msg, MSG_NOSIGNAL));
- if (__builtin_expect (debug_level > 0, 0))
+ if (__glibc_unlikely (debug_level > 0))
dbg_log (_("provide access to FD %d, for %s"), db->ro_fd, key);
}
#endif /* SCM_RIGHTS */
/* Handle new request. */
static void
-handle_request (int fd, request_header *req, void *key, uid_t uid)
+handle_request (int fd, request_header *req, void *key, uid_t uid, pid_t pid)
{
if (__builtin_expect (req->version, NSCD_VERSION) != NSCD_VERSION)
{
return;
}
- /* Make the SELinux check before we go on to the standard checks. We
- need to verify that the request type is valid, since it has not
- yet been checked at this point. */
- if (selinux_enabled
- && __builtin_expect (req->type, GETPWBYNAME) >= GETPWBYNAME
- && __builtin_expect (req->type, LASTREQ) < LASTREQ
- && nscd_request_avc_has_perm (fd, req->type) != 0)
- return;
+ /* Perform the SELinux check before we go on to the standard checks. */
+ if (selinux_enabled && nscd_request_avc_has_perm (fd, req->type) != 0)
+ {
+ if (debug_level > 0)
+ {
+#ifdef SO_PEERCRED
+# ifdef PATH_MAX
+ char buf[PATH_MAX];
+# else
+ char buf[4096];
+# endif
- struct database_dyn *db = serv2db[req->type];
+ snprintf (buf, sizeof (buf), "/proc/%ld/exe", (long int) pid);
+ ssize_t n = readlink (buf, buf, sizeof (buf) - 1);
- // XXX Clean up so that each new command need not introduce a
- // XXX new conditional.
- if ((__builtin_expect (req->type, GETPWBYNAME) >= GETPWBYNAME
- && __builtin_expect (req->type, LASTDBREQ) <= LASTDBREQ)
- || req->type == GETAI || req->type == INITGROUPS)
+ if (n <= 0)
+ dbg_log (_("\
+request from %ld not handled due to missing permission"), (long int) pid);
+ else
+ {
+ buf[n] = '\0';
+ dbg_log (_("\
+request from '%s' [%ld] not handled due to missing permission"),
+ buf, (long int) pid);
+ }
+#else
+ dbg_log (_("request not handled due to missing permission"));
+#endif
+ }
+ return;
+ }
+
+ struct database_dyn *db = reqinfo[req->type].db;
+
+ /* See whether we can service the request from the cache. */
+ if (__builtin_expect (reqinfo[req->type].data_request, true))
{
if (__builtin_expect (debug_level, 0) > 0)
{
}
/* Is this service enabled? */
- if (!db->enabled)
+ if (__glibc_unlikely (!db->enabled))
{
/* No, sent the prepared record. */
- if (TEMP_FAILURE_RETRY (write (fd, db->disabled_iov->iov_base,
- db->disabled_iov->iov_len))
+ if (TEMP_FAILURE_RETRY (send (fd, db->disabled_iov->iov_base,
+ db->disabled_iov->iov_len,
+ MSG_NOSIGNAL))
!= (ssize_t) db->disabled_iov->iov_len
&& __builtin_expect (debug_level, 0) > 0)
{
}
/* Be sure we can read the data. */
- if (__builtin_expect (pthread_rwlock_tryrdlock (&db->lock) != 0, 0))
+ if (__glibc_unlikely (pthread_rwlock_tryrdlock (&db->lock) != 0))
{
++db->head->rdlockdelayed;
pthread_rwlock_rdlock (&db->lock);
if (cached != NULL)
{
/* Hurray it's in the cache. */
- if (TEMP_FAILURE_RETRY (write (fd, cached->data, cached->recsize))
- != cached->recsize
+ ssize_t nwritten;
+
+#ifdef HAVE_SENDFILE
+ if (__glibc_likely (db->mmap_used))
+ {
+ assert (db->wr_fd != -1);
+ assert ((char *) cached->data > (char *) db->data);
+ assert ((char *) cached->data - (char *) db->head
+ + cached->recsize
+ <= (sizeof (struct database_pers_head)
+ + db->head->module * sizeof (ref_t)
+ + db->head->data_size));
+ nwritten = sendfileall (fd, db->wr_fd,
+ (char *) cached->data
+ - (char *) db->head, cached->recsize);
+# ifndef __ASSUME_SENDFILE
+ if (nwritten == -1 && errno == ENOSYS)
+ goto use_write;
+# endif
+ }
+ else
+# ifndef __ASSUME_SENDFILE
+ use_write:
+# endif
+#endif
+ nwritten = writeall (fd, cached->data, cached->recsize);
+
+ if (nwritten != cached->recsize
&& __builtin_expect (debug_level, 0) > 0)
{
/* We have problems sending the result. */
addinitgroups (db, fd, req, key, uid);
break;
+ case GETSERVBYNAME:
+ addservbyname (db, fd, req, key, uid);
+ break;
+
+ case GETSERVBYPORT:
+ addservbyport (db, fd, req, key, uid);
+ break;
+
+ case GETNETGRENT:
+ addgetnetgrent (db, fd, req, key, uid);
+ break;
+
+ case INNETGR:
+ addinnetgr (db, fd, req, key, uid);
+ break;
+
case GETSTAT:
case SHUTDOWN:
case INVALIDATE:
- if (! secure_in_use)
- {
- /* Get the callers credentials. */
+ {
+ /* Get the callers credentials. */
#ifdef SO_PEERCRED
- struct ucred caller;
- socklen_t optlen = sizeof (caller);
+ struct ucred caller;
+ socklen_t optlen = sizeof (caller);
- if (getsockopt (fd, SOL_SOCKET, SO_PEERCRED, &caller, &optlen) < 0)
- {
- char buf[256];
+ if (getsockopt (fd, SOL_SOCKET, SO_PEERCRED, &caller, &optlen) < 0)
+ {
+ char buf[256];
- dbg_log (_("error getting callers id: %s"),
- strerror_r (errno, buf, sizeof (buf)));
- break;
- }
+ dbg_log (_("error getting caller's id: %s"),
+ strerror_r (errno, buf, sizeof (buf)));
+ break;
+ }
- uid = caller.uid;
+ uid = caller.uid;
#else
- /* Some systems have no SO_PEERCRED implementation. They don't
- care about security so we don't as well. */
- uid = 0;
+ /* Some systems have no SO_PEERCRED implementation. They don't
+ care about security so we don't as well. */
+ uid = 0;
#endif
- }
+ }
/* Accept shutdown, getstat and invalidate only from root. For
the stat call also allow the user specified in the config file. */
else if (uid == 0)
{
if (req->type == INVALIDATE)
- invalidate_cache (key);
+ invalidate_cache (key, fd);
else
termination_handler (0);
}
case GETFDPW:
case GETFDGR:
case GETFDHST:
+ case GETFDSERV:
+ case GETFDNETGR:
#ifdef SCM_RIGHTS
- send_ro_fd (serv2db[req->type], key, fd);
+ send_ro_fd (reqinfo[req->type].db, key, fd);
#endif
break;
if (n == -1)
{
dbg_log (_("\
-cannot open /proc/self/cmdline: %s; disabling paranoia mode"),
+cannot read /proc/self/cmdline: %s; disabling paranoia mode"),
strerror (errno));
close (fd);
/* Second, change back to the old user if we changed it. */
if (server_user != NULL)
{
- if (setuid (old_uid) != 0)
+ if (setresuid (old_uid, old_uid, old_uid) != 0)
{
dbg_log (_("\
cannot change to old UID: %s; disabling paranoia mode"),
return;
}
- if (setgid (old_gid) != 0)
+ if (setresgid (old_gid, old_gid, old_gid) != 0)
{
dbg_log (_("\
cannot change to old GID: %s; disabling paranoia mode"),
strerror (errno));
- setuid (server_uid);
+ ignore_value (setuid (server_uid));
paranoia = 0;
return;
}
if (server_user != NULL)
{
- setuid (server_uid);
- setgid (server_gid);
+ ignore_value (setuid (server_uid));
+ ignore_value (setgid (server_gid));
}
paranoia = 0;
return;
}
/* Synchronize memory. */
+ int32_t certainly[lastdb];
for (int cnt = 0; cnt < lastdb; ++cnt)
- {
- /* Make sure nobody keeps using the database. */
- dbs[cnt].head->timestamp = 0;
+ if (dbs[cnt].enabled)
+ {
+ /* Make sure nobody keeps using the database. */
+ dbs[cnt].head->timestamp = 0;
+ certainly[cnt] = dbs[cnt].head->nscd_certainly_running;
+ dbs[cnt].head->nscd_certainly_running = 0;
- if (dbs[cnt].persistent)
- // XXX async OK?
- msync (dbs[cnt].head, dbs[cnt].memsize, MS_ASYNC);
- }
+ if (dbs[cnt].persistent)
+ // XXX async OK?
+ msync (dbs[cnt].head, dbs[cnt].memsize, MS_ASYNC);
+ }
/* The preparations are done. */
+#ifdef PATH_MAX
+ char pathbuf[PATH_MAX];
+#else
+ char pathbuf[256];
+#endif
+ /* Try to exec the real nscd program so the process name (as reported
+ in /proc/PID/status) will be 'nscd', but fall back to /proc/self/exe
+ if readlink or the exec with the result of the readlink call fails. */
+ ssize_t n = readlink ("/proc/self/exe", pathbuf, sizeof (pathbuf) - 1);
+ if (n != -1)
+ {
+ pathbuf[n] = '\0';
+ execv (pathbuf, argv);
+ }
execv ("/proc/self/exe", argv);
/* If we come here, we will never be able to re-exec. */
if (server_user != NULL)
{
- setuid (server_uid);
- setgid (server_gid);
+ ignore_value (setuid (server_uid));
+ ignore_value (setgid (server_gid));
}
- chdir ("/");
+ if (chdir ("/") != 0)
+ dbg_log (_("cannot change current working directory to \"/\": %s"),
+ strerror (errno));
paranoia = 0;
+
+ /* Reenable the databases. */
+ time_t now = time (NULL);
+ for (int cnt = 0; cnt < lastdb; ++cnt)
+ if (dbs[cnt].enabled)
+ {
+ dbs[cnt].head->timestamp = now;
+ dbs[cnt].head->nscd_certainly_running = certainly[cnt];
+ }
}
/* Conditional variable and mutex to signal availability of entries in
READYLIST. The condvar is initialized dynamically since we might
use a different clock depending on availability. */
-static pthread_cond_t readylist_cond;
+static pthread_cond_t readylist_cond = PTHREAD_COND_INITIALIZER;
static pthread_mutex_t readylist_lock = PTHREAD_MUTEX_INITIALIZER;
/* The clock to use with the condvar. */
static unsigned long int nready;
-/* This is the main loop. It is replicated in different threads but the
- `poll' call makes sure only one thread handles an incoming connection. */
+/* Function for the clean-up threads. */
static void *
__attribute__ ((__noreturn__))
-nscd_run (void *p)
+nscd_run_prune (void *p)
{
const long int my_number = (long int) p;
- const int run_prune = my_number < lastdb && dbs[my_number].enabled;
+ assert (dbs[my_number].enabled);
+
+ int dont_need_update = setup_thread (&dbs[my_number]);
+
+ time_t now = time (NULL);
+
+ /* We are running. */
+ dbs[my_number].head->timestamp = now;
+
struct timespec prune_ts;
- int to = 0;
- char buf[256];
+ if (__glibc_unlikely (clock_gettime (timeout_clock, &prune_ts) == -1))
+ /* Should never happen. */
+ abort ();
+
+ /* Compute the initial timeout time. Prevent all the timers to go
+ off at the same time by adding a db-based value. */
+ prune_ts.tv_sec += CACHE_PRUNE_INTERVAL + my_number;
+ dbs[my_number].wakeup_time = now + CACHE_PRUNE_INTERVAL + my_number;
+
+ pthread_mutex_t *prune_lock = &dbs[my_number].prune_lock;
+ pthread_mutex_t *prune_run_lock = &dbs[my_number].prune_run_lock;
+ pthread_cond_t *prune_cond = &dbs[my_number].prune_cond;
- if (run_prune)
+ pthread_mutex_lock (prune_lock);
+ while (1)
{
- setup_thread (&dbs[my_number]);
+ /* Wait, but not forever. */
+ int e = 0;
+ if (! dbs[my_number].clear_cache)
+ e = pthread_cond_timedwait (prune_cond, prune_lock, &prune_ts);
+ assert (__builtin_expect (e == 0 || e == ETIMEDOUT, 1));
+
+ time_t next_wait;
+ now = time (NULL);
+ if (e == ETIMEDOUT || now >= dbs[my_number].wakeup_time
+ || dbs[my_number].clear_cache)
+ {
+ /* We will determine the new timout values based on the
+ cache content. Should there be concurrent additions to
+ the cache which are not accounted for in the cache
+ pruning we want to know about it. Therefore set the
+ timeout to the maximum. It will be descreased when adding
+ new entries to the cache, if necessary. */
+ dbs[my_number].wakeup_time = MAX_TIMEOUT_VALUE;
+
+ /* Unconditionally reset the flag. */
+ time_t prune_now = dbs[my_number].clear_cache ? LONG_MAX : now;
+ dbs[my_number].clear_cache = 0;
+
+ pthread_mutex_unlock (prune_lock);
+
+ /* We use a separate lock for running the prune function (instead
+ of keeping prune_lock locked) because this enables concurrent
+ invocations of cache_add which might modify the timeout value. */
+ pthread_mutex_lock (prune_run_lock);
+ next_wait = prune_cache (&dbs[my_number], prune_now, -1);
+ pthread_mutex_unlock (prune_run_lock);
+
+ next_wait = MAX (next_wait, CACHE_PRUNE_INTERVAL);
+ /* If clients cannot determine for sure whether nscd is running
+ we need to wake up occasionally to update the timestamp.
+ Wait 90% of the update period. */
+#define UPDATE_MAPPING_TIMEOUT (MAPPING_TIMEOUT * 9 / 10)
+ if (__glibc_unlikely (! dont_need_update))
+ {
+ next_wait = MIN (UPDATE_MAPPING_TIMEOUT, next_wait);
+ dbs[my_number].head->timestamp = now;
+ }
- /* We are running. */
- dbs[my_number].head->timestamp = time (NULL);
+ pthread_mutex_lock (prune_lock);
+
+ /* Make it known when we will wake up again. */
+ if (now + next_wait < dbs[my_number].wakeup_time)
+ dbs[my_number].wakeup_time = now + next_wait;
+ else
+ next_wait = dbs[my_number].wakeup_time - now;
+ }
+ else
+ /* The cache was just pruned. Do not do it again now. Just
+ use the new timeout value. */
+ next_wait = dbs[my_number].wakeup_time - now;
if (clock_gettime (timeout_clock, &prune_ts) == -1)
/* Should never happen. */
abort ();
- /* Compute timeout time. */
- prune_ts.tv_sec += CACHE_PRUNE_INTERVAL;
+ /* Compute next timeout time. */
+ prune_ts.tv_sec += next_wait;
}
+}
+
+
+/* This is the main loop. It is replicated in different threads but
+ the use of the ready list makes sure only one thread handles an
+ incoming connection. */
+static void *
+__attribute__ ((__noreturn__))
+nscd_run_worker (void *p)
+{
+ char buf[256];
/* Initial locking. */
pthread_mutex_lock (&readylist_lock);
while (1)
{
while (readylist == NULL)
- {
- if (run_prune)
- {
- /* Wait, but not forever. */
- to = pthread_cond_timedwait (&readylist_cond, &readylist_lock,
- &prune_ts);
-
- /* If we were woken and there is no work to be done,
- just start pruning. */
- if (readylist == NULL && to == ETIMEDOUT)
- {
- --nready;
- pthread_mutex_unlock (&readylist_lock);
- goto only_prune;
- }
- }
- else
- /* No need to timeout. */
- pthread_cond_wait (&readylist_cond, &readylist_lock);
- }
+ pthread_cond_wait (&readylist_cond, &readylist_lock);
struct fdlist *it = readylist->next;
if (readylist->next == readylist)
/* We are done with the list. */
pthread_mutex_unlock (&readylist_lock);
- /* We do not want to block on a short read or so. */
- int fl = fcntl (fd, F_GETFL);
- if (fl == -1 || fcntl (fd, F_SETFL, fl | O_NONBLOCK) == -1)
- goto close_and_out;
+#ifndef __ASSUME_ACCEPT4
+ if (have_accept4 < 0)
+ {
+ /* We do not want to block on a short read or so. */
+ int fl = fcntl (fd, F_GETFL);
+ if (fl == -1 || fcntl (fd, F_SETFL, fl | O_NONBLOCK) == -1)
+ goto close_and_out;
+ }
+#endif
/* Now read the request. */
request_header req;
#ifdef SO_PEERCRED
pid_t pid = 0;
- if (secure_in_use)
- {
- struct ucred caller;
- socklen_t optlen = sizeof (caller);
-
- if (getsockopt (fd, SOL_SOCKET, SO_PEERCRED, &caller, &optlen) < 0)
- {
- dbg_log (_("error getting callers id: %s"),
- strerror_r (errno, buf, sizeof (buf)));
- goto close_and_out;
- }
-
- if (req.type < GETPWBYNAME || req.type > LASTDBREQ
- || serv2db[req.type]->secure)
- uid = caller.uid;
-
- pid = caller.pid;
- }
- else if (__builtin_expect (debug_level > 0, 0))
+ if (__glibc_unlikely (debug_level > 0))
{
struct ucred caller;
socklen_t optlen = sizeof (caller);
if (getsockopt (fd, SOL_SOCKET, SO_PEERCRED, &caller, &optlen) == 0)
pid = caller.pid;
}
+#else
+ const pid_t pid = 0;
#endif
/* It should not be possible to crash the nscd with a silly
request (i.e., a terribly large key). We limit the size to 1kb. */
-#define MAXKEYLEN 1024
if (__builtin_expect (req.key_len, 1) < 0
|| __builtin_expect (req.key_len, 1) > MAXKEYLEN)
{
else
{
/* Get the key. */
- char keybuf[MAXKEYLEN];
+ char keybuf[MAXKEYLEN + 1];
if (__builtin_expect (TEMP_FAILURE_RETRY (read (fd, keybuf,
req.key_len))
strerror_r (errno, buf, sizeof (buf)));
goto close_and_out;
}
+ keybuf[req.key_len] = '\0';
if (__builtin_expect (debug_level, 0) > 0)
{
}
/* Phew, we got all the data, now process it. */
- handle_request (fd, &req, keybuf, uid);
+ handle_request (fd, &req, keybuf, uid, pid);
}
close_and_out:
/* We are done. */
close (fd);
- /* Check whether we should be pruning the cache. */
- assert (run_prune || to == 0);
- if (to == ETIMEDOUT)
- {
- only_prune:
- /* The pthread_cond_timedwait() call timed out. It is time
- to clean up the cache. */
- assert (my_number < lastdb);
- prune_cache (&dbs[my_number],
- prune_ts.tv_sec + (prune_ts.tv_nsec >= 500000000));
-
- if (clock_gettime (timeout_clock, &prune_ts) == -1)
- /* Should never happen. */
- abort ();
-
- /* Compute next timeout time. */
- prune_ts.tv_sec += CACHE_PRUNE_INTERVAL;
-
- /* In case the list is emtpy we do not want to run the prune
- code right away again. */
- to = 0;
- }
-
/* Re-locking. */
pthread_mutex_lock (&readylist_lock);
/* One more thread available. */
++nready;
}
+ /* NOTREACHED */
}
}
bool do_signal = true;
- if (__builtin_expect (nready == 0, 0))
+ if (__glibc_unlikely (nready == 0))
{
++client_queued;
do_signal = false;
/* Try to start another thread to help out. */
pthread_t th;
if (nthreads < max_nthreads
- && pthread_create (&th, &attr, nscd_run,
+ && pthread_create (&th, &attr, nscd_run_worker,
(void *) (long int) nthreads) == 0)
{
/* We got another thread. */
++nthreads;
- /* The new thread might new a kick. */
+ /* The new thread might need a kick. */
do_signal = true;
}
/* Check whether restarting should happen. */
-static inline int
+static bool
restart_p (time_t now)
{
return (paranoia && readylist == NULL && nready == nthreads
/* Array for times a connection was accepted. */
static time_t *starttime;
+#ifdef HAVE_INOTIFY
+/* Inotify event for changed file. */
+union __inev
+{
+ struct inotify_event i;
+# ifndef PATH_MAX
+# define PATH_MAX 1024
+# endif
+ char buf[sizeof (struct inotify_event) + PATH_MAX];
+};
+
+/* Returns 0 if the file is there otherwise -1. */
+int
+check_file (struct traced_file *finfo)
+{
+ struct stat64 st;
+ /* We could check mtime and if different re-add
+ the watches, and invalidate the database, but we
+ don't because we are called from inotify_check_files
+ which should be doing that work. If sufficient inotify
+ events were lost then the next pruning or invalidation
+ will do the stat and mtime check. We don't do it here to
+ keep the logic simple. */
+ if (stat64 (finfo->fname, &st) < 0)
+ return -1;
+ return 0;
+}
+
+/* Process the inotify event in INEV. If the event matches any of the files
+ registered with a database then mark that database as requiring its cache
+ to be cleared. We indicate the cache needs clearing by setting
+ TO_CLEAR[DBCNT] to true for the matching database. */
+static void
+inotify_check_files (bool *to_clear, union __inev *inev)
+{
+ /* Check which of the files changed. */
+ for (size_t dbcnt = 0; dbcnt < lastdb; ++dbcnt)
+ {
+ struct traced_file *finfo = dbs[dbcnt].traced_files;
+
+ while (finfo != NULL)
+ {
+ /* The configuration file was moved or deleted.
+ We stop watching it at that point, and reinitialize. */
+ if (finfo->inotify_descr[TRACED_FILE] == inev->i.wd
+ && ((inev->i.mask & IN_MOVE_SELF)
+ || (inev->i.mask & IN_DELETE_SELF)
+ || (inev->i.mask & IN_IGNORED)))
+ {
+ int ret;
+ bool moved = (inev->i.mask & IN_MOVE_SELF) != 0;
+
+ if (check_file (finfo) == 0)
+ {
+ dbg_log (_("ignored inotify event for `%s` (file exists)"),
+ finfo->fname);
+ return;
+ }
+
+ dbg_log (_("monitored file `%s` was %s, removing watch"),
+ finfo->fname, moved ? "moved" : "deleted");
+ /* File was moved out, remove the watch. Watches are
+ automatically removed when the file is deleted. */
+ if (moved)
+ {
+ ret = inotify_rm_watch (inotify_fd, inev->i.wd);
+ if (ret < 0)
+ dbg_log (_("failed to remove file watch `%s`: %s"),
+ finfo->fname, strerror (errno));
+ }
+ finfo->inotify_descr[TRACED_FILE] = -1;
+ to_clear[dbcnt] = true;
+ if (finfo->call_res_init)
+ res_init ();
+ return;
+ }
+ /* The configuration file was open for writing and has just closed.
+ We reset the cache and reinitialize. */
+ if (finfo->inotify_descr[TRACED_FILE] == inev->i.wd
+ && inev->i.mask & IN_CLOSE_WRITE)
+ {
+ /* Mark cache as needing to be cleared and reinitialize. */
+ dbg_log (_("monitored file `%s` was written to"), finfo->fname);
+ to_clear[dbcnt] = true;
+ if (finfo->call_res_init)
+ res_init ();
+ return;
+ }
+ /* The parent directory was moved or deleted. We trigger one last
+ invalidation. At the next pruning or invalidation we may add
+ this watch back if the file is present again. */
+ if (finfo->inotify_descr[TRACED_DIR] == inev->i.wd
+ && ((inev->i.mask & IN_DELETE_SELF)
+ || (inev->i.mask & IN_MOVE_SELF)
+ || (inev->i.mask & IN_IGNORED)))
+ {
+ bool moved = (inev->i.mask & IN_MOVE_SELF) != 0;
+ /* The directory watch may have already been removed
+ but we don't know so we just remove it again and
+ ignore the error. Then we remove the file watch.
+ Note: watches are automatically removed for deleted
+ files. */
+ if (moved)
+ inotify_rm_watch (inotify_fd, inev->i.wd);
+ if (finfo->inotify_descr[TRACED_FILE] != -1)
+ {
+ dbg_log (_("monitored parent directory `%s` was %s, removing watch on `%s`"),
+ finfo->dname, moved ? "moved" : "deleted", finfo->fname);
+ if (inotify_rm_watch (inotify_fd, finfo->inotify_descr[TRACED_FILE]) < 0)
+ dbg_log (_("failed to remove file watch `%s`: %s"),
+ finfo->dname, strerror (errno));
+ }
+ finfo->inotify_descr[TRACED_FILE] = -1;
+ finfo->inotify_descr[TRACED_DIR] = -1;
+ to_clear[dbcnt] = true;
+ if (finfo->call_res_init)
+ res_init ();
+ /* Continue to the next entry since this might be the
+ parent directory for multiple registered files and
+ we want to remove watches for all registered files. */
+ continue;
+ }
+ /* The parent directory had a create or moved to event. */
+ if (finfo->inotify_descr[TRACED_DIR] == inev->i.wd
+ && ((inev->i.mask & IN_MOVED_TO)
+ || (inev->i.mask & IN_CREATE))
+ && strcmp (inev->i.name, finfo->sfname) == 0)
+ {
+ /* We detected a directory change. We look for the creation
+ of the file we are tracking or the move of the same file
+ into the directory. */
+ int ret;
+ dbg_log (_("monitored file `%s` was %s, adding watch"),
+ finfo->fname,
+ inev->i.mask & IN_CREATE ? "created" : "moved into place");
+ /* File was moved in or created. Regenerate the watch. */
+ if (finfo->inotify_descr[TRACED_FILE] != -1)
+ inotify_rm_watch (inotify_fd,
+ finfo->inotify_descr[TRACED_FILE]);
+
+ ret = inotify_add_watch (inotify_fd,
+ finfo->fname,
+ TRACED_FILE_MASK);
+ if (ret < 0)
+ dbg_log (_("failed to add file watch `%s`: %s"),
+ finfo->fname, strerror (errno));
+
+ finfo->inotify_descr[TRACED_FILE] = ret;
+
+ /* The file is new or moved so mark cache as needing to
+ be cleared and reinitialize. */
+ to_clear[dbcnt] = true;
+ if (finfo->call_res_init)
+ res_init ();
+
+ /* Done re-adding the watch. Don't return, we may still
+ have other files in this same directory, same watch
+ descriptor, and need to process them. */
+ }
+ /* Other events are ignored, and we move on to the next file. */
+ finfo = finfo->next;
+ }
+ }
+}
+
+/* If an entry in the array of booleans TO_CLEAR is TRUE then clear the cache
+ for the associated database, otherwise do nothing. The TO_CLEAR array must
+ have LASTDB entries. */
+static inline void
+clear_db_cache (bool *to_clear)
+{
+ for (size_t dbcnt = 0; dbcnt < lastdb; ++dbcnt)
+ if (to_clear[dbcnt])
+ {
+ pthread_mutex_lock (&dbs[dbcnt].prune_lock);
+ dbs[dbcnt].clear_cache = 1;
+ pthread_mutex_unlock (&dbs[dbcnt].prune_lock);
+ pthread_cond_signal (&dbs[dbcnt].prune_cond);
+ }
+}
+
+int
+handle_inotify_events (void)
+{
+ bool to_clear[lastdb] = { false, };
+ union __inev inev;
+
+ /* Read all inotify events for files registered via
+ register_traced_file(). */
+ while (1)
+ {
+ /* Potentially read multiple events into buf. */
+ ssize_t nb = TEMP_FAILURE_RETRY (read (inotify_fd,
+ &inev.buf,
+ sizeof (inev)));
+ if (nb < (ssize_t) sizeof (struct inotify_event))
+ {
+ /* Not even 1 event. */
+ if (__glibc_unlikely (nb == -1 && errno != EAGAIN))
+ return -1;
+ /* Done reading events that are ready. */
+ break;
+ }
+ /* Process all events. The normal inotify interface delivers
+ complete events on a read and never a partial event. */
+ char *eptr = &inev.buf[0];
+ ssize_t count;
+ while (1)
+ {
+ /* Check which of the files changed. */
+ inotify_check_files (to_clear, &inev);
+ count = sizeof (struct inotify_event) + inev.i.len;
+ eptr += count;
+ nb -= count;
+ if (nb >= (ssize_t) sizeof (struct inotify_event))
+ memcpy (&inev, eptr, nb);
+ else
+ break;
+ }
+ continue;
+ }
+ /* Actually perform the cache clearing. */
+ clear_db_cache (to_clear);
+ return 0;
+}
+
+#endif
static void
__attribute__ ((__noreturn__))
size_t nused = 1;
size_t firstfree = 1;
+#ifdef HAVE_INOTIFY
+ if (inotify_fd != -1)
+ {
+ conns[1].fd = inotify_fd;
+ conns[1].events = POLLRDNORM;
+ nused = 2;
+ firstfree = 2;
+ }
+#endif
+
+#ifdef HAVE_NETLINK
+ size_t idx_nl_status_fd = 0;
+ if (nl_status_fd != -1)
+ {
+ idx_nl_status_fd = nused;
+ conns[nused].fd = nl_status_fd;
+ conns[nused].events = POLLRDNORM;
+ ++nused;
+ firstfree = nused;
+ }
+#endif
+
while (1)
{
/* Wait for any event. We wait at most a couple of seconds so
if (conns[0].revents != 0)
{
/* We have a new incoming connection. Accept the connection. */
- int fd = TEMP_FAILURE_RETRY (accept (sock, NULL, NULL));
+ int fd;
- /* use the descriptor if we have not reached the limit. */
- if (fd >= 0 && firstfree < nconns)
+#ifndef __ASSUME_ACCEPT4
+ fd = -1;
+ if (have_accept4 >= 0)
+#endif
{
- conns[firstfree].fd = fd;
- conns[firstfree].events = POLLRDNORM;
- starttime[firstfree] = now;
- if (firstfree >= nused)
- nused = firstfree + 1;
+ fd = TEMP_FAILURE_RETRY (accept4 (sock, NULL, NULL,
+ SOCK_NONBLOCK));
+#ifndef __ASSUME_ACCEPT4
+ if (have_accept4 == 0)
+ have_accept4 = fd != -1 || errno != ENOSYS ? 1 : -1;
+#endif
+ }
+#ifndef __ASSUME_ACCEPT4
+ if (have_accept4 < 0)
+ fd = TEMP_FAILURE_RETRY (accept (sock, NULL, NULL));
+#endif
- do
- ++firstfree;
- while (firstfree < nused && conns[firstfree].fd != -1);
+ /* Use the descriptor if we have not reached the limit. */
+ if (fd >= 0)
+ {
+ if (firstfree < nconns)
+ {
+ conns[firstfree].fd = fd;
+ conns[firstfree].events = POLLRDNORM;
+ starttime[firstfree] = now;
+ if (firstfree >= nused)
+ nused = firstfree + 1;
+
+ do
+ ++firstfree;
+ while (firstfree < nused && conns[firstfree].fd != -1);
+ }
+ else
+ /* We cannot use the connection so close it. */
+ close (fd);
}
--n;
}
- for (size_t cnt = 1; cnt < nused && n > 0; ++cnt)
+ size_t first = 1;
+#ifdef HAVE_INOTIFY
+ if (inotify_fd != -1 && conns[1].fd == inotify_fd)
+ {
+ if (conns[1].revents != 0)
+ {
+ int ret;
+ ret = handle_inotify_events ();
+ if (ret == -1)
+ {
+ /* Something went wrong when reading the inotify
+ data. Better disable inotify. */
+ dbg_log (_("disabled inotify-based monitoring after read error %d"), errno);
+ conns[1].fd = -1;
+ firstfree = 1;
+ if (nused == 2)
+ nused = 1;
+ close (inotify_fd);
+ inotify_fd = -1;
+ }
+ --n;
+ }
+
+ first = 2;
+ }
+#endif
+
+#ifdef HAVE_NETLINK
+ if (idx_nl_status_fd != 0 && conns[idx_nl_status_fd].revents != 0)
+ {
+ char buf[4096];
+ /* Read all the data. We do not interpret it here. */
+ while (TEMP_FAILURE_RETRY (read (nl_status_fd, buf,
+ sizeof (buf))) != -1)
+ ;
+
+ dbs[hstdb].head->extra_data[NSCD_HST_IDX_CONF_TIMESTAMP]
+ = __bump_nl_timestamp ();
+ }
+#endif
+
+ for (size_t cnt = first; cnt < nused && n > 0; ++cnt)
if (conns[cnt].revents != 0)
{
fd_ready (conns[cnt].fd);
/* We cannot use epoll. */
return;
+# ifdef HAVE_INOTIFY
+ if (inotify_fd != -1)
+ {
+ ev.events = EPOLLRDNORM;
+ ev.data.fd = inotify_fd;
+ if (epoll_ctl (efd, EPOLL_CTL_ADD, inotify_fd, &ev) == -1)
+ /* We cannot use epoll. */
+ return;
+ nused = 2;
+ }
+# endif
+
+# ifdef HAVE_NETLINK
+ if (nl_status_fd != -1)
+ {
+ ev.events = EPOLLRDNORM;
+ ev.data.fd = nl_status_fd;
+ if (epoll_ctl (efd, EPOLL_CTL_ADD, nl_status_fd, &ev) == -1)
+ /* We cannot use epoll. */
+ return;
+ }
+# endif
+
while (1)
{
struct epoll_event revs[100];
if (revs[cnt].data.fd == sock)
{
/* A new connection. */
- int fd = TEMP_FAILURE_RETRY (accept (sock, NULL, NULL));
+ int fd;
+
+# ifndef __ASSUME_ACCEPT4
+ fd = -1;
+ if (have_accept4 >= 0)
+# endif
+ {
+ fd = TEMP_FAILURE_RETRY (accept4 (sock, NULL, NULL,
+ SOCK_NONBLOCK));
+# ifndef __ASSUME_ACCEPT4
+ if (have_accept4 == 0)
+ have_accept4 = fd != -1 || errno != ENOSYS ? 1 : -1;
+# endif
+ }
+# ifndef __ASSUME_ACCEPT4
+ if (have_accept4 < 0)
+ fd = TEMP_FAILURE_RETRY (accept (sock, NULL, NULL));
+# endif
+ /* Use the descriptor if we have not reached the limit. */
if (fd >= 0)
{
/* Try to add the new descriptor. */
}
}
}
+# ifdef HAVE_INOTIFY
+ else if (revs[cnt].data.fd == inotify_fd)
+ {
+ int ret;
+ ret = handle_inotify_events ();
+ if (ret == -1)
+ {
+ /* Something went wrong when reading the inotify
+ data. Better disable inotify. */
+ dbg_log (_("disabled inotify-based monitoring after read error %d"), errno);
+ (void) epoll_ctl (efd, EPOLL_CTL_DEL, inotify_fd, NULL);
+ close (inotify_fd);
+ inotify_fd = -1;
+ break;
+ }
+ }
+# endif
+# ifdef HAVE_NETLINK
+ else if (revs[cnt].data.fd == nl_status_fd)
+ {
+ char buf[4096];
+ /* Read all the data. We do not interpret it here. */
+ while (TEMP_FAILURE_RETRY (read (nl_status_fd, buf,
+ sizeof (buf))) != -1)
+ ;
+
+ __bump_nl_timestamp ();
+ }
+# endif
else
{
/* Remove the descriptor from the epoll descriptor. */
- struct epoll_event ev = { 0, };
- (void) epoll_ctl (efd, EPOLL_CTL_DEL, revs[cnt].data.fd, &ev);
+ (void) epoll_ctl (efd, EPOLL_CTL_DEL, revs[cnt].data.fd, NULL);
- /* Get a worked to handle the request. */
+ /* Get a worker to handle the request. */
fd_ready (revs[cnt].data.fd);
/* Reset the time. */
/* Now look for descriptors for accepted connections which have
no reply in too long of a time. */
time_t laststart = now - ACCEPT_TIMEOUT;
+ assert (starttime[sock] == 0);
+# ifdef HAVE_INOTIFY
+ assert (inotify_fd == -1 || starttime[inotify_fd] == 0);
+# endif
+ assert (nl_status_fd == -1 || starttime[nl_status_fd] == 0);
for (int cnt = highest; cnt > STDERR_FILENO; --cnt)
- if (cnt != sock && starttime[cnt] != 0 && starttime[cnt] < laststart)
+ if (starttime[cnt] != 0 && starttime[cnt] < laststart)
{
/* We are waiting for this one for too long. Close it. */
- struct epoll_event ev = {0, };
- (void) epoll_ctl (efd, EPOLL_CTL_DEL, cnt, &ev);
+ (void) epoll_ctl (efd, EPOLL_CTL_DEL, cnt, NULL);
(void) close (cnt);
timeout_clock = CLOCK_MONOTONIC;
#endif
- pthread_cond_init (&readylist_cond, &condattr);
- pthread_condattr_destroy (&condattr);
-
-
/* Create the attribute for the threads. They are all created
detached. */
pthread_attr_init (&attr);
pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
/* Use 1MB stacks, twice as much for 64-bit architectures. */
- pthread_attr_setstacksize (&attr, 1024 * 1024 * (sizeof (void *) / 4));
+ pthread_attr_setstacksize (&attr, NSCD_THREAD_STACKSIZE);
/* We allow less than LASTDB threads only for debugging. */
if (debug_level == 0)
nthreads = MAX (nthreads, lastdb);
- int nfailed = 0;
- for (long int i = 0; i < nthreads; ++i)
+ /* Create the threads which prune the databases. */
+ // XXX Ideally this work would be done by some of the worker threads.
+ // XXX But this is problematic since we would need to be able to wake
+ // XXX them up explicitly as well as part of the group handling the
+ // XXX ready-list. This requires an operation where we can wait on
+ // XXX two conditional variables at the same time. This operation
+ // XXX does not exist (yet).
+ for (long int i = 0; i < lastdb; ++i)
{
+ /* Initialize the conditional variable. */
+ if (pthread_cond_init (&dbs[i].prune_cond, &condattr) != 0)
+ {
+ dbg_log (_("could not initialize conditional variable"));
+ do_exit (1, 0, NULL);
+ }
+
pthread_t th;
- if (pthread_create (&th, &attr, nscd_run, (void *) (i - nfailed)) != 0)
- ++nfailed;
+ if (dbs[i].enabled
+ && pthread_create (&th, &attr, nscd_run_prune, (void *) i) != 0)
+ {
+ dbg_log (_("could not start clean-up thread; terminating"));
+ do_exit (1, 0, NULL);
+ }
}
- if (nthreads - nfailed < lastdb)
+
+ pthread_condattr_destroy (&condattr);
+
+ for (long int i = 0; i < nthreads; ++i)
{
- /* We could not start enough threads. */
- dbg_log (_("could only start %d threads; terminating"),
- nthreads - nfailed);
- exit (1);
+ pthread_t th;
+ if (pthread_create (&th, &attr, nscd_run_worker, NULL) != 0)
+ {
+ if (i == 0)
+ {
+ dbg_log (_("could not start any worker thread; terminating"));
+ do_exit (1, 0, NULL);
+ }
+
+ break;
+ }
}
+ /* Now it is safe to let the parent know that we're doing fine and it can
+ exit. */
+ notify_parent (0);
+
/* Determine how much room for descriptors we should initially
allocate. This might need to change later if we cap the number
with MAXCONN. */
if (pwd == NULL)
{
dbg_log (_("Failed to run nscd as user '%s'"), server_user);
- error (EXIT_FAILURE, 0, _("Failed to run nscd as user '%s'"),
- server_user);
+ do_exit (EXIT_FAILURE, 0,
+ _("Failed to run nscd as user '%s'"), server_user);
}
server_uid = pwd->pw_uid;
{
/* This really must never happen. */
dbg_log (_("Failed to run nscd as user '%s'"), server_user);
- error (EXIT_FAILURE, errno, _("initial getgrouplist failed"));
+ do_exit (EXIT_FAILURE, errno,
+ _("initial getgrouplist failed"));
}
server_groups = (gid_t *) xmalloc (server_ngroups * sizeof (gid_t));
== -1)
{
dbg_log (_("Failed to run nscd as user '%s'"), server_user);
- error (EXIT_FAILURE, errno, _("getgrouplist failed"));
+ do_exit (EXIT_FAILURE, errno, _("getgrouplist failed"));
}
}
static void
finish_drop_privileges (void)
{
+#if defined HAVE_LIBAUDIT && defined HAVE_LIBCAP
+ /* We need to preserve the capabilities to connect to the audit daemon. */
+ cap_t new_caps = preserve_capabilities ();
+#endif
+
if (setgroups (server_ngroups, server_groups) == -1)
{
dbg_log (_("Failed to run nscd as user '%s'"), server_user);
- error (EXIT_FAILURE, errno, _("setgroups failed"));
+ do_exit (EXIT_FAILURE, errno, _("setgroups failed"));
}
- if (setgid (server_gid) == -1)
+ int res;
+ if (paranoia)
+ res = setresgid (server_gid, server_gid, old_gid);
+ else
+ res = setgid (server_gid);
+ if (res == -1)
{
dbg_log (_("Failed to run nscd as user '%s'"), server_user);
- perror ("setgid");
- exit (1);
+ do_exit (4, errno, "setgid");
}
- if (setuid (server_uid) == -1)
+ if (paranoia)
+ res = setresuid (server_uid, server_uid, old_uid);
+ else
+ res = setuid (server_uid);
+ if (res == -1)
{
dbg_log (_("Failed to run nscd as user '%s'"), server_user);
- perror ("setuid");
- exit (1);
+ do_exit (4, errno, "setuid");
}
+
+#if defined HAVE_LIBAUDIT && defined HAVE_LIBCAP
+ /* Remove the temporary capabilities. */
+ install_real_capabilities (new_caps);
+#endif
}