2 * Copyright (c) 2008, 2009, 2010, 2011, 2012 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "socket-util.h"
19 #include <arpa/inet.h>
30 #include <sys/resource.h>
31 #include <sys/socket.h>
36 #include "dynamic-string.h"
37 #include "fatal-signal.h"
39 #include "poll-loop.h"
42 #if AF_PACKET && __linux__
43 #include <linux/if_packet.h>
46 #include "netlink-protocol.h"
47 #include "netlink-socket.h"
50 VLOG_DEFINE_THIS_MODULE(socket_util);
52 /* #ifdefs make it a pain to maintain code: you have to try to build both ways.
53 * Thus, this file compiles all of the code regardless of the target, by
54 * writing "if (LINUX)" instead of "#ifdef __linux__". */
65 static int getsockopt_int(int fd, int level, int option, const char *optname,
68 /* Sets 'fd' to non-blocking mode. Returns 0 if successful, otherwise a
69 * positive errno value. */
71 set_nonblocking(int fd)
73 int flags = fcntl(fd, F_GETFL, 0);
75 if (fcntl(fd, F_SETFL, flags | O_NONBLOCK) != -1) {
78 VLOG_ERR("fcntl(F_SETFL) failed: %s", strerror(errno));
82 VLOG_ERR("fcntl(F_GETFL) failed: %s", strerror(errno));
88 xset_nonblocking(int fd)
90 if (set_nonblocking(fd)) {
96 set_dscp(int fd, uint8_t dscp)
105 if (setsockopt(fd, IPPROTO_IP, IP_TOS, &val, sizeof val)) {
113 rlim_is_finite(rlim_t limit)
115 if (limit == RLIM_INFINITY) {
119 #ifdef RLIM_SAVED_CUR /* FreeBSD 8.0 lacks RLIM_SAVED_CUR. */
120 if (limit == RLIM_SAVED_CUR) {
125 #ifdef RLIM_SAVED_MAX /* FreeBSD 8.0 lacks RLIM_SAVED_MAX. */
126 if (limit == RLIM_SAVED_MAX) {
134 /* Returns the maximum valid FD value, plus 1. */
138 static int max_fds = -1;
141 if (!getrlimit(RLIMIT_NOFILE, &r) && rlim_is_finite(r.rlim_cur)) {
142 max_fds = r.rlim_cur;
144 VLOG_WARN("failed to obtain fd limit, defaulting to 1024");
151 /* Translates 'host_name', which must be a string representation of an IP
152 * address, into a numeric IP address in '*addr'. Returns 0 if successful,
153 * otherwise a positive errno value. */
155 lookup_ip(const char *host_name, struct in_addr *addr)
157 if (!inet_aton(host_name, addr)) {
158 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
159 VLOG_ERR_RL(&rl, "\"%s\" is not a valid IP address", host_name);
165 /* Translates 'host_name', which must be a string representation of an IPv6
166 * address, into a numeric IPv6 address in '*addr'. Returns 0 if successful,
167 * otherwise a positive errno value. */
169 lookup_ipv6(const char *host_name, struct in6_addr *addr)
171 if (inet_pton(AF_INET6, host_name, addr) != 1) {
172 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
173 VLOG_ERR_RL(&rl, "\"%s\" is not a valid IPv6 address", host_name);
179 /* Translates 'host_name', which must be a host name or a string representation
180 * of an IP address, into a numeric IP address in '*addr'. Returns 0 if
181 * successful, otherwise a positive errno value.
183 * Most Open vSwitch code should not use this because it causes deadlocks:
184 * gethostbyname() sends out a DNS request but that starts a new flow for which
185 * OVS must set up a flow, but it can't because it's waiting for a DNS reply.
186 * The synchronous lookup also delays other activity. (Of course we can solve
187 * this but it doesn't seem worthwhile quite yet.) */
189 lookup_hostname(const char *host_name, struct in_addr *addr)
193 if (inet_aton(host_name, addr)) {
197 h = gethostbyname(host_name);
199 *addr = *(struct in_addr *) h->h_addr;
203 return (h_errno == HOST_NOT_FOUND ? ENOENT
204 : h_errno == TRY_AGAIN ? EAGAIN
205 : h_errno == NO_RECOVERY ? EIO
206 : h_errno == NO_ADDRESS ? ENXIO
210 /* Returns the error condition associated with socket 'fd' and resets the
211 * socket's error status. */
213 get_socket_error(int fd)
217 if (getsockopt_int(fd, SOL_SOCKET, SO_ERROR, "SO_ERROR", &error)) {
224 check_connection_completion(int fd)
230 pfd.events = POLLOUT;
232 retval = poll(&pfd, 1, 0);
233 } while (retval < 0 && errno == EINTR);
235 return get_socket_error(fd);
236 } else if (retval < 0) {
237 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 10);
238 VLOG_ERR_RL(&rl, "poll: %s", strerror(errno));
245 /* Drain all the data currently in the receive queue of a datagram socket (and
246 * possibly additional data). There is no way to know how many packets are in
247 * the receive queue, but we do know that the total number of bytes queued does
248 * not exceed the receive buffer size, so we pull packets until none are left
249 * or we've read that many bytes. */
255 rcvbuf = get_socket_rcvbuf(fd);
261 /* In Linux, specifying MSG_TRUNC in the flags argument causes the
262 * datagram length to be returned, even if that is longer than the
263 * buffer provided. Thus, we can use a 1-byte buffer to discard the
264 * incoming datagram and still be able to account how many bytes were
265 * removed from the receive buffer.
267 * On other Unix-like OSes, MSG_TRUNC has no effect in the flags
269 char buffer[LINUX ? 1 : 2048];
270 ssize_t n_bytes = recv(fd, buffer, sizeof buffer,
271 MSG_TRUNC | MSG_DONTWAIT);
272 if (n_bytes <= 0 || n_bytes >= rcvbuf) {
280 /* Returns the size of socket 'sock''s receive buffer (SO_RCVBUF), or a
281 * negative errno value if an error occurs. */
283 get_socket_rcvbuf(int sock)
288 error = getsockopt_int(sock, SOL_SOCKET, SO_RCVBUF, "SO_RCVBUF", &rcvbuf);
289 return error ? -error : rcvbuf;
292 /* Reads and discards up to 'n' datagrams from 'fd', stopping as soon as no
293 * more data can be immediately read. ('fd' should therefore be in
294 * non-blocking mode.)*/
296 drain_fd(int fd, size_t n_packets)
298 for (; n_packets > 0; n_packets--) {
299 /* 'buffer' only needs to be 1 byte long in most circumstances. This
300 * size is defensive against the possibility that we someday want to
301 * use a Linux tap device without TUN_NO_PI, in which case a buffer
302 * smaller than sizeof(struct tun_pi) will give EINVAL on read. */
304 if (read(fd, buffer, sizeof buffer) <= 0) {
310 /* Stores in '*un' a sockaddr_un that refers to file 'name'. Stores in
311 * '*un_len' the size of the sockaddr_un. */
313 make_sockaddr_un__(const char *name, struct sockaddr_un *un, socklen_t *un_len)
315 un->sun_family = AF_UNIX;
316 ovs_strzcpy(un->sun_path, name, sizeof un->sun_path);
317 *un_len = (offsetof(struct sockaddr_un, sun_path)
318 + strlen (un->sun_path) + 1);
321 /* Stores in '*un' a sockaddr_un that refers to file 'name'. Stores in
322 * '*un_len' the size of the sockaddr_un.
324 * Returns 0 on success, otherwise a positive errno value. On success,
325 * '*dirfdp' is either -1 or a nonnegative file descriptor that the caller
326 * should close after using '*un' to bind or connect. On failure, '*dirfdp' is
329 make_sockaddr_un(const char *name, struct sockaddr_un *un, socklen_t *un_len,
332 enum { MAX_UN_LEN = sizeof un->sun_path - 1 };
335 if (strlen(name) > MAX_UN_LEN) {
336 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
339 /* 'name' is too long to fit in a sockaddr_un, but we have a
340 * workaround for that on Linux: shorten it by opening a file
341 * descriptor for the directory part of the name and indirecting
342 * through /proc/self/fd/<dirfd>/<basename>. */
347 dir = dir_name(name);
348 base = base_name(name);
350 dirfd = open(dir, O_DIRECTORY | O_RDONLY);
357 short_name = xasprintf("/proc/self/fd/%d/%s", dirfd, base);
361 if (strlen(short_name) <= MAX_UN_LEN) {
362 make_sockaddr_un__(short_name, un, un_len);
370 VLOG_WARN_RL(&rl, "Unix socket name %s is longer than maximum "
371 "%d bytes (even shortened)", name, MAX_UN_LEN);
373 /* 'name' is too long and we have no workaround. */
374 VLOG_WARN_RL(&rl, "Unix socket name %s is longer than maximum "
375 "%d bytes", name, MAX_UN_LEN);
380 make_sockaddr_un__(name, un, un_len);
385 /* Binds Unix domain socket 'fd' to a file with permissions 0700. */
387 bind_unix_socket(int fd, struct sockaddr *sun, socklen_t sun_len)
389 /* According to _Unix Network Programming_, umask should affect bind(). */
390 mode_t old_umask = umask(0077);
391 int error = bind(fd, sun, sun_len) ? errno : 0;
396 /* Creates a Unix domain socket in the given 'style' (either SOCK_DGRAM or
397 * SOCK_STREAM) that is bound to '*bind_path' (if 'bind_path' is non-null) and
398 * connected to '*connect_path' (if 'connect_path' is non-null). If 'nonblock'
399 * is true, the socket is made non-blocking.
401 * Returns the socket's fd if successful, otherwise a negative errno value. */
403 make_unix_socket(int style, bool nonblock,
404 const char *bind_path, const char *connect_path)
409 fd = socket(PF_UNIX, style, 0);
414 /* Set nonblocking mode right away, if we want it. This prevents blocking
415 * in connect(), if connect_path != NULL. (In turn, that's a corner case:
416 * it will only happen if style is SOCK_STREAM or SOCK_SEQPACKET, and only
417 * if a backlog of un-accepted connections has built up in the kernel.) */
419 int flags = fcntl(fd, F_GETFL, 0);
424 if (fcntl(fd, F_SETFL, flags | O_NONBLOCK) == -1) {
431 struct sockaddr_un un;
435 if (unlink(bind_path) && errno != ENOENT) {
436 VLOG_WARN("unlinking \"%s\": %s\n", bind_path, strerror(errno));
438 fatal_signal_add_file_to_unlink(bind_path);
440 error = make_sockaddr_un(bind_path, &un, &un_len, &dirfd);
442 error = bind_unix_socket(fd, (struct sockaddr *) &un, un_len);
453 struct sockaddr_un un;
457 error = make_sockaddr_un(connect_path, &un, &un_len, &dirfd);
459 && connect(fd, (struct sockaddr*) &un, un_len)
460 && errno != EINPROGRESS) {
474 if (error == EAGAIN) {
478 fatal_signal_unlink_file_now(bind_path);
485 get_unix_name_len(socklen_t sun_len)
487 return (sun_len >= offsetof(struct sockaddr_un, sun_path)
488 ? sun_len - offsetof(struct sockaddr_un, sun_path)
493 guess_netmask(ovs_be32 ip_)
495 uint32_t ip = ntohl(ip_);
496 return ((ip >> 31) == 0 ? htonl(0xff000000) /* Class A */
497 : (ip >> 30) == 2 ? htonl(0xffff0000) /* Class B */
498 : (ip >> 29) == 6 ? htonl(0xffffff00) /* Class C */
499 : htonl(0)); /* ??? */
502 /* Parses 'target', which should be a string in the format "<host>[:<port>]".
503 * <host> is required. If 'default_port' is nonzero then <port> is optional
504 * and defaults to 'default_port'.
506 * On success, returns true and stores the parsed remote address into '*sinp'.
507 * On failure, logs an error, stores zeros into '*sinp', and returns false. */
509 inet_parse_active(const char *target_, uint16_t default_port,
510 struct sockaddr_in *sinp)
512 char *target = xstrdup(target_);
513 char *save_ptr = NULL;
514 const char *host_name;
515 const char *port_string;
519 sinp->sin_family = AF_INET;
520 sinp->sin_port = htons(default_port);
523 host_name = strtok_r(target, ":", &save_ptr);
524 port_string = strtok_r(NULL, ":", &save_ptr);
526 VLOG_ERR("%s: bad peer name format", target_);
530 /* Look up IP, port. */
531 if (lookup_ip(host_name, &sinp->sin_addr)) {
534 if (port_string && atoi(port_string)) {
535 sinp->sin_port = htons(atoi(port_string));
536 } else if (!default_port) {
537 VLOG_ERR("%s: port number must be specified", target_);
545 memset(sinp, 0, sizeof *sinp);
551 /* Opens a non-blocking IPv4 socket of the specified 'style' and connects to
552 * 'target', which should be a string in the format "<host>[:<port>]". <host>
553 * is required. If 'default_port' is nonzero then <port> is optional and
554 * defaults to 'default_port'.
556 * 'style' should be SOCK_STREAM (for TCP) or SOCK_DGRAM (for UDP).
558 * On success, returns 0 (indicating connection complete) or EAGAIN (indicating
559 * connection in progress), in which case the new file descriptor is stored
560 * into '*fdp'. On failure, returns a positive errno value other than EAGAIN
561 * and stores -1 into '*fdp'.
563 * If 'sinp' is non-null, then on success the target address is stored into
566 * 'dscp' becomes the DSCP bits in the IP headers for the new connection. It
567 * should be in the range [0, 63] and will automatically be shifted to the
568 * appropriately place in the IP tos field. */
570 inet_open_active(int style, const char *target, uint16_t default_port,
571 struct sockaddr_in *sinp, int *fdp, uint8_t dscp)
573 struct sockaddr_in sin;
578 if (!inet_parse_active(target, default_port, &sin)) {
579 error = EAFNOSUPPORT;
583 /* Create non-blocking socket. */
584 fd = socket(AF_INET, style, 0);
586 VLOG_ERR("%s: socket: %s", target, strerror(errno));
590 error = set_nonblocking(fd);
595 /* The dscp bits must be configured before connect() to ensure that the TOS
596 * field is set during the connection establishment. If set after
597 * connect(), the handshake SYN frames will be sent with a TOS of 0. */
598 error = set_dscp(fd, dscp);
600 VLOG_ERR("%s: socket: %s", target, strerror(error));
605 error = connect(fd, (struct sockaddr *) &sin, sizeof sin) == 0 ? 0 : errno;
606 if (error == EINPROGRESS) {
611 if (!error || error == EAGAIN) {
615 } else if (fd >= 0) {
622 /* Parses 'target', which should be a string in the format "[<port>][:<ip>]":
624 * - If 'default_port' is -1, then <port> is required. Otherwise, if
625 * <port> is omitted, then 'default_port' is used instead.
627 * - If <port> (or 'default_port', if used) is 0, then no port is bound
628 * and the TCP/IP stack will select a port.
630 * - If <ip> is omitted then the IP address is wildcarded.
632 * If successful, stores the address into '*sinp' and returns true; otherwise
633 * zeros '*sinp' and returns false. */
635 inet_parse_passive(const char *target_, int default_port,
636 struct sockaddr_in *sinp)
638 char *target = xstrdup(target_);
639 char *string_ptr = target;
640 const char *host_name;
641 const char *port_string;
645 /* Address defaults. */
646 memset(sinp, 0, sizeof *sinp);
647 sinp->sin_family = AF_INET;
648 sinp->sin_addr.s_addr = htonl(INADDR_ANY);
649 sinp->sin_port = htons(default_port);
651 /* Parse optional port number. */
652 port_string = strsep(&string_ptr, ":");
653 if (port_string && str_to_int(port_string, 10, &port)) {
654 sinp->sin_port = htons(port);
655 } else if (default_port < 0) {
656 VLOG_ERR("%s: port number must be specified", target_);
660 /* Parse optional bind IP. */
661 host_name = strsep(&string_ptr, ":");
662 if (host_name && host_name[0] && lookup_ip(host_name, &sinp->sin_addr)) {
670 memset(sinp, 0, sizeof *sinp);
677 /* Opens a non-blocking IPv4 socket of the specified 'style', binds to
678 * 'target', and listens for incoming connections. Parses 'target' in the same
679 * way was inet_parse_passive().
681 * 'style' should be SOCK_STREAM (for TCP) or SOCK_DGRAM (for UDP).
683 * For TCP, the socket will have SO_REUSEADDR turned on.
685 * On success, returns a non-negative file descriptor. On failure, returns a
686 * negative errno value.
688 * If 'sinp' is non-null, then on success the bound address is stored into
691 * 'dscp' becomes the DSCP bits in the IP headers for the new connection. It
692 * should be in the range [0, 63] and will automatically be shifted to the
693 * appropriately place in the IP tos field. */
695 inet_open_passive(int style, const char *target, int default_port,
696 struct sockaddr_in *sinp, uint8_t dscp)
698 struct sockaddr_in sin;
700 unsigned int yes = 1;
702 if (!inet_parse_passive(target, default_port, &sin)) {
703 return -EAFNOSUPPORT;
706 /* Create non-blocking socket, set SO_REUSEADDR. */
707 fd = socket(AF_INET, style, 0);
710 VLOG_ERR("%s: socket: %s", target, strerror(error));
713 error = set_nonblocking(fd);
717 if (style == SOCK_STREAM
718 && setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &yes, sizeof yes) < 0) {
720 VLOG_ERR("%s: setsockopt(SO_REUSEADDR): %s", target, strerror(error));
725 if (bind(fd, (struct sockaddr *) &sin, sizeof sin) < 0) {
727 VLOG_ERR("%s: bind: %s", target, strerror(error));
731 /* The dscp bits must be configured before connect() to ensure that the TOS
732 * field is set during the connection establishment. If set after
733 * connect(), the handshake SYN frames will be sent with a TOS of 0. */
734 error = set_dscp(fd, dscp);
736 VLOG_ERR("%s: socket: %s", target, strerror(error));
741 if (style == SOCK_STREAM && listen(fd, 10) < 0) {
743 VLOG_ERR("%s: listen: %s", target, strerror(error));
748 socklen_t sin_len = sizeof sin;
749 if (getsockname(fd, (struct sockaddr *) &sin, &sin_len) < 0){
751 VLOG_ERR("%s: getsockname: %s", target, strerror(error));
754 if (sin.sin_family != AF_INET || sin_len != sizeof sin) {
755 error = EAFNOSUPPORT;
756 VLOG_ERR("%s: getsockname: invalid socket name", target);
769 /* Returns a readable and writable fd for /dev/null, if successful, otherwise
770 * a negative errno value. The caller must not close the returned fd (because
771 * the same fd will be handed out to subsequent callers). */
775 static int null_fd = -1;
777 null_fd = open("/dev/null", O_RDWR);
780 VLOG_ERR("could not open /dev/null: %s", strerror(error));
788 read_fully(int fd, void *p_, size_t size, size_t *bytes_read)
794 ssize_t retval = read(fd, p, size);
796 *bytes_read += retval;
799 } else if (retval == 0) {
801 } else if (errno != EINTR) {
809 write_fully(int fd, const void *p_, size_t size, size_t *bytes_written)
811 const uint8_t *p = p_;
815 ssize_t retval = write(fd, p, size);
817 *bytes_written += retval;
820 } else if (retval == 0) {
821 VLOG_WARN("write returned 0");
823 } else if (errno != EINTR) {
830 /* Given file name 'file_name', fsyncs the directory in which it is contained.
831 * Returns 0 if successful, otherwise a positive errno value. */
833 fsync_parent_dir(const char *file_name)
839 dir = dir_name(file_name);
840 fd = open(dir, O_RDONLY);
843 if (errno == EINVAL || errno == EROFS) {
844 /* This directory does not support synchronization. Not
845 * really an error. */
848 VLOG_ERR("%s: fsync failed (%s)", dir, strerror(error));
854 VLOG_ERR("%s: open failed (%s)", dir, strerror(error));
861 /* Obtains the modification time of the file named 'file_name' to the greatest
862 * supported precision. If successful, stores the mtime in '*mtime' and
863 * returns 0. On error, returns a positive errno value and stores zeros in
866 get_mtime(const char *file_name, struct timespec *mtime)
870 if (!stat(file_name, &s)) {
871 mtime->tv_sec = s.st_mtime;
873 #if HAVE_STRUCT_STAT_ST_MTIM_TV_NSEC
874 mtime->tv_nsec = s.st_mtim.tv_nsec;
875 #elif HAVE_STRUCT_STAT_ST_MTIMENSEC
876 mtime->tv_nsec = s.st_mtimensec;
883 mtime->tv_sec = mtime->tv_nsec = 0;
892 VLOG_FATAL("failed to create pipe (%s)", strerror(errno));
897 xsocketpair(int domain, int type, int protocol, int fds[2])
899 if (socketpair(domain, type, protocol, fds)) {
900 VLOG_FATAL("failed to create socketpair (%s)", strerror(errno));
905 getsockopt_int(int fd, int level, int option, const char *optname, int *valuep)
907 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 10);
913 if (getsockopt(fd, level, option, &value, &len)) {
915 VLOG_ERR_RL(&rl, "getsockopt(%s): %s", optname, strerror(error));
916 } else if (len != sizeof value) {
918 VLOG_ERR_RL(&rl, "getsockopt(%s): value is %u bytes (expected %zu)",
919 optname, (unsigned int) len, sizeof value);
924 *valuep = error ? 0 : value;
929 describe_sockaddr(struct ds *string, int fd,
930 int (*getaddr)(int, struct sockaddr *, socklen_t *))
932 struct sockaddr_storage ss;
933 socklen_t len = sizeof ss;
935 if (!getaddr(fd, (struct sockaddr *) &ss, &len)) {
936 if (ss.ss_family == AF_INET) {
937 struct sockaddr_in sin;
939 memcpy(&sin, &ss, sizeof sin);
940 ds_put_format(string, IP_FMT":%"PRIu16,
941 IP_ARGS(&sin.sin_addr.s_addr), ntohs(sin.sin_port));
942 } else if (ss.ss_family == AF_UNIX) {
943 struct sockaddr_un sun;
947 memcpy(&sun, &ss, sizeof sun);
948 maxlen = len - offsetof(struct sockaddr_un, sun_path);
949 null = memchr(sun.sun_path, '\0', maxlen);
950 ds_put_buffer(string, sun.sun_path,
951 null ? null - sun.sun_path : maxlen);
954 else if (ss.ss_family == AF_NETLINK) {
957 /* SO_PROTOCOL was introduced in 2.6.32. Support it regardless of the version
958 * of the Linux kernel headers in use at build time. */
960 #define SO_PROTOCOL 38
963 if (!getsockopt_int(fd, SOL_SOCKET, SO_PROTOCOL, "SO_PROTOCOL",
967 ds_put_cstr(string, "NETLINK_ROUTE");
970 case NETLINK_GENERIC:
971 ds_put_cstr(string, "NETLINK_GENERIC");
975 ds_put_format(string, "AF_NETLINK family %d", protocol);
979 ds_put_cstr(string, "AF_NETLINK");
983 #if AF_PACKET && __linux__
984 else if (ss.ss_family == AF_PACKET) {
985 struct sockaddr_ll sll;
987 memcpy(&sll, &ss, sizeof sll);
988 ds_put_cstr(string, "AF_PACKET");
989 if (sll.sll_ifindex) {
992 if (if_indextoname(sll.sll_ifindex, name)) {
993 ds_put_format(string, "(%s)", name);
995 ds_put_format(string, "(ifindex=%d)", sll.sll_ifindex);
998 if (sll.sll_protocol) {
999 ds_put_format(string, "(protocol=0x%"PRIu16")",
1000 ntohs(sll.sll_protocol));
1004 else if (ss.ss_family == AF_UNSPEC) {
1005 ds_put_cstr(string, "AF_UNSPEC");
1007 ds_put_format(string, "AF_%d", (int) ss.ss_family);
1015 put_fd_filename(struct ds *string, int fd)
1021 linkname = xasprintf("/proc/self/fd/%d", fd);
1022 n = readlink(linkname, buf, sizeof buf);
1024 ds_put_char(string, ' ');
1025 ds_put_buffer(string, buf, n);
1026 if (n > sizeof buf) {
1027 ds_put_cstr(string, "...");
1034 /* Returns a malloc()'d string describing 'fd', for use in logging. */
1042 if (fstat(fd, &s)) {
1043 ds_put_format(&string, "fstat failed (%s)", strerror(errno));
1044 } else if (S_ISSOCK(s.st_mode)) {
1045 describe_sockaddr(&string, fd, getsockname);
1046 ds_put_cstr(&string, "<->");
1047 describe_sockaddr(&string, fd, getpeername);
1049 ds_put_cstr(&string, (isatty(fd) ? "tty"
1050 : S_ISDIR(s.st_mode) ? "directory"
1051 : S_ISCHR(s.st_mode) ? "character device"
1052 : S_ISBLK(s.st_mode) ? "block device"
1053 : S_ISREG(s.st_mode) ? "file"
1054 : S_ISFIFO(s.st_mode) ? "FIFO"
1055 : S_ISLNK(s.st_mode) ? "symbolic link"
1058 put_fd_filename(&string, fd);
1061 return ds_steal_cstr(&string);
1064 /* Returns the total of the 'iov_len' members of the 'n_iovs' in 'iovs'.
1065 * The caller must ensure that the total does not exceed SIZE_MAX. */
1067 iovec_len(const struct iovec iovs[], size_t n_iovs)
1072 for (i = 0; i < n_iovs; i++) {
1073 len += iovs[i].iov_len;
1078 /* Returns true if all of the 'n_iovs' iovecs in 'iovs' have length zero. */
1080 iovec_is_empty(const struct iovec iovs[], size_t n_iovs)
1084 for (i = 0; i < n_iovs; i++) {
1085 if (iovs[i].iov_len) {
1092 /* Sends the 'n_iovs' iovecs of data in 'iovs' and the 'n_fds' file descriptors
1093 * in 'fds' on Unix domain socket 'sock'. Returns the number of bytes
1094 * successfully sent or -1 if an error occurred. On error, sets errno
1097 send_iovec_and_fds(int sock,
1098 const struct iovec *iovs, size_t n_iovs,
1099 const int fds[], size_t n_fds)
1105 char control[CMSG_SPACE(SOUTIL_MAX_FDS * sizeof *fds)];
1109 assert(!iovec_is_empty(iovs, n_iovs));
1110 assert(n_fds <= SOUTIL_MAX_FDS);
1112 memset(&cmsg, 0, sizeof cmsg);
1113 cmsg.cm.cmsg_len = CMSG_LEN(n_fds * sizeof *fds);
1114 cmsg.cm.cmsg_level = SOL_SOCKET;
1115 cmsg.cm.cmsg_type = SCM_RIGHTS;
1116 memcpy(CMSG_DATA(&cmsg.cm), fds, n_fds * sizeof *fds);
1118 msg.msg_name = NULL;
1119 msg.msg_namelen = 0;
1120 msg.msg_iov = (struct iovec *) iovs;
1121 msg.msg_iovlen = n_iovs;
1122 msg.msg_control = &cmsg.cm;
1123 msg.msg_controllen = CMSG_SPACE(n_fds * sizeof *fds);
1126 return sendmsg(sock, &msg, 0);
1128 return writev(sock, iovs, n_iovs);
1132 /* Sends the 'n_iovs' iovecs of data in 'iovs' and the 'n_fds' file descriptors
1133 * in 'fds' on Unix domain socket 'sock'. If 'skip_bytes' is nonzero, then the
1134 * first 'skip_bytes' of data in the iovecs are not sent, and none of the file
1135 * descriptors are sent. The function continues to retry sending until an
1136 * error (other than EINTR) occurs or all the data and fds are sent.
1138 * Returns 0 if all the data and fds were successfully sent, otherwise a
1139 * positive errno value. Regardless of success, stores the number of bytes
1140 * sent (always at least 'skip_bytes') in '*bytes_sent'. (If at least one byte
1141 * is sent, then all the fds have been sent.)
1143 * 'skip_bytes' must be less than or equal to iovec_len(iovs, n_iovs). */
1145 send_iovec_and_fds_fully(int sock,
1146 const struct iovec iovs[], size_t n_iovs,
1147 const int fds[], size_t n_fds,
1148 size_t skip_bytes, size_t *bytes_sent)
1151 while (n_iovs > 0) {
1155 retval = skip_bytes;
1157 } else if (!*bytes_sent) {
1158 retval = send_iovec_and_fds(sock, iovs, n_iovs, fds, n_fds);
1160 retval = writev(sock, iovs, n_iovs);
1164 *bytes_sent += retval;
1165 while (retval > 0) {
1166 const uint8_t *base = iovs->iov_base;
1167 size_t len = iovs->iov_len;
1173 error = write_fully(sock, base + retval, len - retval,
1175 *bytes_sent += sent;
1185 } else if (retval == 0) {
1186 if (iovec_is_empty(iovs, n_iovs)) {
1189 VLOG_WARN("send returned 0");
1191 } else if (errno != EINTR) {
1199 /* Sends the 'n_iovs' iovecs of data in 'iovs' and the 'n_fds' file descriptors
1200 * in 'fds' on Unix domain socket 'sock'. The function continues to retry
1201 * sending until an error (other than EAGAIN or EINTR) occurs or all the data
1202 * and fds are sent. Upon EAGAIN, the function blocks until the socket is
1203 * ready for more data.
1205 * Returns 0 if all the data and fds were successfully sent, otherwise a
1206 * positive errno value. */
1208 send_iovec_and_fds_fully_block(int sock,
1209 const struct iovec iovs[], size_t n_iovs,
1210 const int fds[], size_t n_fds)
1217 error = send_iovec_and_fds_fully(sock, iovs, n_iovs,
1218 fds, n_fds, sent, &sent);
1219 if (error != EAGAIN) {
1222 poll_fd_wait(sock, POLLOUT);
1227 /* Attempts to receive from Unix domain socket 'sock' up to 'size' bytes of
1228 * data into 'data' and up to SOUTIL_MAX_FDS file descriptors into 'fds'.
1230 * - Upon success, returns the number of bytes of data copied into 'data'
1231 * and stores the number of received file descriptors into '*n_fdsp'.
1233 * - On failure, returns a negative errno value and stores 0 in
1236 * - On EOF, returns 0 and stores 0 in '*n_fdsp'. */
1238 recv_data_and_fds(int sock,
1239 void *data, size_t size,
1240 int fds[SOUTIL_MAX_FDS], size_t *n_fdsp)
1244 char control[CMSG_SPACE(SOUTIL_MAX_FDS * sizeof *fds)];
1256 iov.iov_base = data;
1259 msg.msg_name = NULL;
1260 msg.msg_namelen = 0;
1263 msg.msg_control = &cmsg.cm;
1264 msg.msg_controllen = sizeof cmsg.control;
1267 retval = recvmsg(sock, &msg, 0);
1268 } while (retval < 0 && errno == EINTR);
1270 return retval < 0 ? -errno : 0;
1273 for (p = CMSG_FIRSTHDR(&msg); p; p = CMSG_NXTHDR(&msg, p)) {
1274 if (p->cmsg_level != SOL_SOCKET || p->cmsg_type != SCM_RIGHTS) {
1275 VLOG_ERR("unexpected control message %d:%d",
1276 p->cmsg_level, p->cmsg_type);
1278 } else if (*n_fdsp) {
1279 VLOG_ERR("multiple SCM_RIGHTS received");
1282 size_t n_fds = (p->cmsg_len - CMSG_LEN(0)) / sizeof *fds;
1283 const int *fds_data = (const int *) CMSG_DATA(p);
1286 if (n_fds > SOUTIL_MAX_FDS) {
1287 VLOG_ERR("%zu fds received but only %d supported",
1288 n_fds, SOUTIL_MAX_FDS);
1289 for (i = 0; i < n_fds; i++) {
1296 memcpy(fds, fds_data, n_fds * sizeof *fds);
1303 for (i = 0; i < *n_fdsp; i++) {