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_DATAPATH
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_DATAPATH)" instead of "#ifdef __linux__". */
55 #ifndef LINUX_DATAPATH
56 #define LINUX_DATAPATH 0
63 static int getsockopt_int(int fd, int level, int option, const char *optname,
66 /* Sets 'fd' to non-blocking mode. Returns 0 if successful, otherwise a
67 * positive errno value. */
69 set_nonblocking(int fd)
71 int flags = fcntl(fd, F_GETFL, 0);
73 if (fcntl(fd, F_SETFL, flags | O_NONBLOCK) != -1) {
76 VLOG_ERR("fcntl(F_SETFL) failed: %s", strerror(errno));
80 VLOG_ERR("fcntl(F_GETFL) failed: %s", strerror(errno));
86 xset_nonblocking(int fd)
88 if (set_nonblocking(fd)) {
94 set_dscp(int fd, uint8_t dscp)
103 if (setsockopt(fd, IPPROTO_IP, IP_TOS, &val, sizeof val)) {
111 rlim_is_finite(rlim_t limit)
113 if (limit == RLIM_INFINITY) {
117 #ifdef RLIM_SAVED_CUR /* FreeBSD 8.0 lacks RLIM_SAVED_CUR. */
118 if (limit == RLIM_SAVED_CUR) {
123 #ifdef RLIM_SAVED_MAX /* FreeBSD 8.0 lacks RLIM_SAVED_MAX. */
124 if (limit == RLIM_SAVED_MAX) {
132 /* Returns the maximum valid FD value, plus 1. */
136 static int max_fds = -1;
139 if (!getrlimit(RLIMIT_NOFILE, &r) && rlim_is_finite(r.rlim_cur)) {
140 max_fds = r.rlim_cur;
142 VLOG_WARN("failed to obtain fd limit, defaulting to 1024");
149 /* Translates 'host_name', which must be a string representation of an IP
150 * address, into a numeric IP address in '*addr'. Returns 0 if successful,
151 * otherwise a positive errno value. */
153 lookup_ip(const char *host_name, struct in_addr *addr)
155 if (!inet_aton(host_name, addr)) {
156 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
157 VLOG_ERR_RL(&rl, "\"%s\" is not a valid IP address", host_name);
163 /* Translates 'host_name', which must be a string representation of an IPv6
164 * address, into a numeric IPv6 address in '*addr'. Returns 0 if successful,
165 * otherwise a positive errno value. */
167 lookup_ipv6(const char *host_name, struct in6_addr *addr)
169 if (inet_pton(AF_INET6, host_name, addr) != 1) {
170 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
171 VLOG_ERR_RL(&rl, "\"%s\" is not a valid IPv6 address", host_name);
177 /* Translates 'host_name', which must be a host name or a string representation
178 * of an IP address, into a numeric IP address in '*addr'. Returns 0 if
179 * successful, otherwise a positive errno value.
181 * Most Open vSwitch code should not use this because it causes deadlocks:
182 * gethostbyname() sends out a DNS request but that starts a new flow for which
183 * OVS must set up a flow, but it can't because it's waiting for a DNS reply.
184 * The synchronous lookup also delays other activity. (Of course we can solve
185 * this but it doesn't seem worthwhile quite yet.) */
187 lookup_hostname(const char *host_name, struct in_addr *addr)
191 if (inet_aton(host_name, addr)) {
195 h = gethostbyname(host_name);
197 *addr = *(struct in_addr *) h->h_addr;
201 return (h_errno == HOST_NOT_FOUND ? ENOENT
202 : h_errno == TRY_AGAIN ? EAGAIN
203 : h_errno == NO_RECOVERY ? EIO
204 : h_errno == NO_ADDRESS ? ENXIO
209 check_connection_completion(int fd)
211 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 10);
216 pfd.events = POLLOUT;
218 retval = poll(&pfd, 1, 0);
219 } while (retval < 0 && errno == EINTR);
221 if (pfd.revents & POLLERR) {
222 ssize_t n = send(fd, "", 1, MSG_DONTWAIT);
226 VLOG_ERR_RL(&rl, "poll return POLLERR but send succeeded");
231 } else if (retval < 0) {
232 VLOG_ERR_RL(&rl, "poll: %s", strerror(errno));
239 /* Drain all the data currently in the receive queue of a datagram socket (and
240 * possibly additional data). There is no way to know how many packets are in
241 * the receive queue, but we do know that the total number of bytes queued does
242 * not exceed the receive buffer size, so we pull packets until none are left
243 * or we've read that many bytes. */
249 rcvbuf = get_socket_rcvbuf(fd);
255 /* In Linux, specifying MSG_TRUNC in the flags argument causes the
256 * datagram length to be returned, even if that is longer than the
257 * buffer provided. Thus, we can use a 1-byte buffer to discard the
258 * incoming datagram and still be able to account how many bytes were
259 * removed from the receive buffer.
261 * On other Unix-like OSes, MSG_TRUNC has no effect in the flags
263 char buffer[LINUX_DATAPATH ? 1 : 2048];
264 ssize_t n_bytes = recv(fd, buffer, sizeof buffer,
265 MSG_TRUNC | MSG_DONTWAIT);
266 if (n_bytes <= 0 || n_bytes >= rcvbuf) {
274 /* Returns the size of socket 'sock''s receive buffer (SO_RCVBUF), or a
275 * negative errno value if an error occurs. */
277 get_socket_rcvbuf(int sock)
282 error = getsockopt_int(sock, SOL_SOCKET, SO_RCVBUF, "SO_RCVBUF", &rcvbuf);
283 return error ? -error : rcvbuf;
286 /* Reads and discards up to 'n' datagrams from 'fd', stopping as soon as no
287 * more data can be immediately read. ('fd' should therefore be in
288 * non-blocking mode.)*/
290 drain_fd(int fd, size_t n_packets)
292 for (; n_packets > 0; n_packets--) {
293 /* 'buffer' only needs to be 1 byte long in most circumstances. This
294 * size is defensive against the possibility that we someday want to
295 * use a Linux tap device without TUN_NO_PI, in which case a buffer
296 * smaller than sizeof(struct tun_pi) will give EINVAL on read. */
298 if (read(fd, buffer, sizeof buffer) <= 0) {
304 /* Stores in '*un' a sockaddr_un that refers to file 'name'. Stores in
305 * '*un_len' the size of the sockaddr_un. */
307 make_sockaddr_un__(const char *name, struct sockaddr_un *un, socklen_t *un_len)
309 un->sun_family = AF_UNIX;
310 ovs_strzcpy(un->sun_path, name, sizeof un->sun_path);
311 *un_len = (offsetof(struct sockaddr_un, sun_path)
312 + strlen (un->sun_path) + 1);
315 /* Stores in '*un' a sockaddr_un that refers to file 'name'. Stores in
316 * '*un_len' the size of the sockaddr_un.
318 * Returns 0 on success, otherwise a positive errno value. On success,
319 * '*dirfdp' is either -1 or a nonnegative file descriptor that the caller
320 * should close after using '*un' to bind or connect. On failure, '*dirfdp' is
323 make_sockaddr_un(const char *name, struct sockaddr_un *un, socklen_t *un_len,
326 enum { MAX_UN_LEN = sizeof un->sun_path - 1 };
329 if (strlen(name) > MAX_UN_LEN) {
330 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
332 if (LINUX_DATAPATH) {
333 /* 'name' is too long to fit in a sockaddr_un, but we have a
334 * workaround for that on Linux: shorten it by opening a file
335 * descriptor for the directory part of the name and indirecting
336 * through /proc/self/fd/<dirfd>/<basename>. */
341 dir = dir_name(name);
342 base = base_name(name);
344 dirfd = open(dir, O_DIRECTORY | O_RDONLY);
351 short_name = xasprintf("/proc/self/fd/%d/%s", dirfd, base);
355 if (strlen(short_name) <= MAX_UN_LEN) {
356 make_sockaddr_un__(short_name, un, un_len);
364 VLOG_WARN_RL(&rl, "Unix socket name %s is longer than maximum "
365 "%d bytes (even shortened)", name, MAX_UN_LEN);
367 /* 'name' is too long and we have no workaround. */
368 VLOG_WARN_RL(&rl, "Unix socket name %s is longer than maximum "
369 "%d bytes", name, MAX_UN_LEN);
374 make_sockaddr_un__(name, un, un_len);
379 /* Binds Unix domain socket 'fd' to a file with permissions 0700. */
381 bind_unix_socket(int fd, struct sockaddr *sun, socklen_t sun_len)
383 /* According to _Unix Network Programming_, umask should affect bind(). */
384 mode_t old_umask = umask(0077);
385 int error = bind(fd, sun, sun_len) ? errno : 0;
390 /* Creates a Unix domain socket in the given 'style' (either SOCK_DGRAM or
391 * SOCK_STREAM) that is bound to '*bind_path' (if 'bind_path' is non-null) and
392 * connected to '*connect_path' (if 'connect_path' is non-null). If 'nonblock'
393 * is true, the socket is made non-blocking.
395 * Returns the socket's fd if successful, otherwise a negative errno value. */
397 make_unix_socket(int style, bool nonblock,
398 const char *bind_path, const char *connect_path)
403 fd = socket(PF_UNIX, style, 0);
408 /* Set nonblocking mode right away, if we want it. This prevents blocking
409 * in connect(), if connect_path != NULL. (In turn, that's a corner case:
410 * it will only happen if style is SOCK_STREAM or SOCK_SEQPACKET, and only
411 * if a backlog of un-accepted connections has built up in the kernel.) */
413 int flags = fcntl(fd, F_GETFL, 0);
418 if (fcntl(fd, F_SETFL, flags | O_NONBLOCK) == -1) {
425 struct sockaddr_un un;
429 if (unlink(bind_path) && errno != ENOENT) {
430 VLOG_WARN("unlinking \"%s\": %s\n", bind_path, strerror(errno));
432 fatal_signal_add_file_to_unlink(bind_path);
434 error = make_sockaddr_un(bind_path, &un, &un_len, &dirfd);
436 error = bind_unix_socket(fd, (struct sockaddr *) &un, un_len);
447 struct sockaddr_un un;
451 error = make_sockaddr_un(connect_path, &un, &un_len, &dirfd);
453 && connect(fd, (struct sockaddr*) &un, un_len)
454 && errno != EINPROGRESS) {
468 if (error == EAGAIN) {
472 fatal_signal_unlink_file_now(bind_path);
479 get_unix_name_len(socklen_t sun_len)
481 return (sun_len >= offsetof(struct sockaddr_un, sun_path)
482 ? sun_len - offsetof(struct sockaddr_un, sun_path)
487 guess_netmask(ovs_be32 ip_)
489 uint32_t ip = ntohl(ip_);
490 return ((ip >> 31) == 0 ? htonl(0xff000000) /* Class A */
491 : (ip >> 30) == 2 ? htonl(0xffff0000) /* Class B */
492 : (ip >> 29) == 6 ? htonl(0xffffff00) /* Class C */
493 : htonl(0)); /* ??? */
496 /* Parses 'target', which should be a string in the format "<host>[:<port>]".
497 * <host> is required. If 'default_port' is nonzero then <port> is optional
498 * and defaults to 'default_port'.
500 * On success, returns true and stores the parsed remote address into '*sinp'.
501 * On failure, logs an error, stores zeros into '*sinp', and returns false. */
503 inet_parse_active(const char *target_, uint16_t default_port,
504 struct sockaddr_in *sinp)
506 char *target = xstrdup(target_);
507 char *save_ptr = NULL;
508 const char *host_name;
509 const char *port_string;
513 sinp->sin_family = AF_INET;
514 sinp->sin_port = htons(default_port);
517 host_name = strtok_r(target, ":", &save_ptr);
518 port_string = strtok_r(NULL, ":", &save_ptr);
520 VLOG_ERR("%s: bad peer name format", target_);
524 /* Look up IP, port. */
525 if (lookup_ip(host_name, &sinp->sin_addr)) {
528 if (port_string && atoi(port_string)) {
529 sinp->sin_port = htons(atoi(port_string));
530 } else if (!default_port) {
531 VLOG_ERR("%s: port number must be specified", target_);
539 memset(sinp, 0, sizeof *sinp);
545 /* Opens a non-blocking IPv4 socket of the specified 'style' and connects to
546 * 'target', which should be a string in the format "<host>[:<port>]". <host>
547 * is required. If 'default_port' is nonzero then <port> is optional and
548 * defaults to 'default_port'.
550 * 'style' should be SOCK_STREAM (for TCP) or SOCK_DGRAM (for UDP).
552 * On success, returns 0 (indicating connection complete) or EAGAIN (indicating
553 * connection in progress), in which case the new file descriptor is stored
554 * into '*fdp'. On failure, returns a positive errno value other than EAGAIN
555 * and stores -1 into '*fdp'.
557 * If 'sinp' is non-null, then on success the target address is stored into
560 * 'dscp' becomes the DSCP bits in the IP headers for the new connection. It
561 * should be in the range [0, 63] and will automatically be shifted to the
562 * appropriately place in the IP tos field. */
564 inet_open_active(int style, const char *target, uint16_t default_port,
565 struct sockaddr_in *sinp, int *fdp, uint8_t dscp)
567 struct sockaddr_in sin;
572 if (!inet_parse_active(target, default_port, &sin)) {
573 error = EAFNOSUPPORT;
577 /* Create non-blocking socket. */
578 fd = socket(AF_INET, style, 0);
580 VLOG_ERR("%s: socket: %s", target, strerror(errno));
584 error = set_nonblocking(fd);
589 /* The dscp bits must be configured before connect() to ensure that the TOS
590 * field is set during the connection establishment. If set after
591 * connect(), the handshake SYN frames will be sent with a TOS of 0. */
592 error = set_dscp(fd, dscp);
594 VLOG_ERR("%s: socket: %s", target, strerror(error));
599 error = connect(fd, (struct sockaddr *) &sin, sizeof sin) == 0 ? 0 : errno;
600 if (error == EINPROGRESS) {
605 if (!error || error == EAGAIN) {
609 } else if (fd >= 0) {
617 /* Parses 'target', which should be a string in the format "[<port>][:<ip>]":
619 * - If 'default_port' is -1, then <port> is required. Otherwise, if
620 * <port> is omitted, then 'default_port' is used instead.
622 * - If <port> (or 'default_port', if used) is 0, then no port is bound
623 * and the TCP/IP stack will select a port.
625 * - If <ip> is omitted then the IP address is wildcarded.
627 * If successful, stores the address into '*sinp' and returns true; otherwise
628 * zeros '*sinp' and returns false. */
630 inet_parse_passive(const char *target_, int default_port,
631 struct sockaddr_in *sinp)
633 char *target = xstrdup(target_);
634 char *string_ptr = target;
635 const char *host_name;
636 const char *port_string;
640 /* Address defaults. */
641 memset(sinp, 0, sizeof *sinp);
642 sinp->sin_family = AF_INET;
643 sinp->sin_addr.s_addr = htonl(INADDR_ANY);
644 sinp->sin_port = htons(default_port);
646 /* Parse optional port number. */
647 port_string = strsep(&string_ptr, ":");
648 if (port_string && str_to_int(port_string, 10, &port)) {
649 sinp->sin_port = htons(port);
650 } else if (default_port < 0) {
651 VLOG_ERR("%s: port number must be specified", target_);
655 /* Parse optional bind IP. */
656 host_name = strsep(&string_ptr, ":");
657 if (host_name && host_name[0] && lookup_ip(host_name, &sinp->sin_addr)) {
665 memset(sinp, 0, sizeof *sinp);
672 /* Opens a non-blocking IPv4 socket of the specified 'style', binds to
673 * 'target', and listens for incoming connections. Parses 'target' in the same
674 * way was inet_parse_passive().
676 * 'style' should be SOCK_STREAM (for TCP) or SOCK_DGRAM (for UDP).
678 * For TCP, the socket will have SO_REUSEADDR turned on.
680 * On success, returns a non-negative file descriptor. On failure, returns a
681 * negative errno value.
683 * If 'sinp' is non-null, then on success the bound address is stored into
686 * 'dscp' becomes the DSCP bits in the IP headers for the new connection. It
687 * should be in the range [0, 63] and will automatically be shifted to the
688 * appropriately place in the IP tos field. */
690 inet_open_passive(int style, const char *target, int default_port,
691 struct sockaddr_in *sinp, uint8_t dscp)
693 struct sockaddr_in sin;
695 unsigned int yes = 1;
697 if (!inet_parse_passive(target, default_port, &sin)) {
698 return -EAFNOSUPPORT;
701 /* Create non-blocking socket, set SO_REUSEADDR. */
702 fd = socket(AF_INET, style, 0);
705 VLOG_ERR("%s: socket: %s", target, strerror(error));
708 error = set_nonblocking(fd);
712 if (style == SOCK_STREAM
713 && setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &yes, sizeof yes) < 0) {
715 VLOG_ERR("%s: setsockopt(SO_REUSEADDR): %s", target, strerror(error));
720 if (bind(fd, (struct sockaddr *) &sin, sizeof sin) < 0) {
722 VLOG_ERR("%s: bind: %s", target, strerror(error));
726 /* The dscp bits must be configured before connect() to ensure that the TOS
727 * field is set during the connection establishment. If set after
728 * connect(), the handshake SYN frames will be sent with a TOS of 0. */
729 error = set_dscp(fd, dscp);
731 VLOG_ERR("%s: socket: %s", target, strerror(error));
736 if (style == SOCK_STREAM && listen(fd, 10) < 0) {
738 VLOG_ERR("%s: listen: %s", target, strerror(error));
743 socklen_t sin_len = sizeof sin;
744 if (getsockname(fd, (struct sockaddr *) &sin, &sin_len) < 0){
746 VLOG_ERR("%s: getsockname: %s", target, strerror(error));
749 if (sin.sin_family != AF_INET || sin_len != sizeof sin) {
750 error = EAFNOSUPPORT;
751 VLOG_ERR("%s: getsockname: invalid socket name", target);
764 /* Returns a readable and writable fd for /dev/null, if successful, otherwise
765 * a negative errno value. The caller must not close the returned fd (because
766 * the same fd will be handed out to subsequent callers). */
770 static int null_fd = -1;
772 null_fd = open("/dev/null", O_RDWR);
775 VLOG_ERR("could not open /dev/null: %s", strerror(error));
783 read_fully(int fd, void *p_, size_t size, size_t *bytes_read)
789 ssize_t retval = read(fd, p, size);
791 *bytes_read += retval;
794 } else if (retval == 0) {
796 } else if (errno != EINTR) {
804 write_fully(int fd, const void *p_, size_t size, size_t *bytes_written)
806 const uint8_t *p = p_;
810 ssize_t retval = write(fd, p, size);
812 *bytes_written += retval;
815 } else if (retval == 0) {
816 VLOG_WARN("write returned 0");
818 } else if (errno != EINTR) {
825 /* Given file name 'file_name', fsyncs the directory in which it is contained.
826 * Returns 0 if successful, otherwise a positive errno value. */
828 fsync_parent_dir(const char *file_name)
834 dir = dir_name(file_name);
835 fd = open(dir, O_RDONLY);
838 if (errno == EINVAL || errno == EROFS) {
839 /* This directory does not support synchronization. Not
840 * really an error. */
843 VLOG_ERR("%s: fsync failed (%s)", dir, strerror(error));
849 VLOG_ERR("%s: open failed (%s)", dir, strerror(error));
856 /* Obtains the modification time of the file named 'file_name' to the greatest
857 * supported precision. If successful, stores the mtime in '*mtime' and
858 * returns 0. On error, returns a positive errno value and stores zeros in
861 get_mtime(const char *file_name, struct timespec *mtime)
865 if (!stat(file_name, &s)) {
866 mtime->tv_sec = s.st_mtime;
868 #if HAVE_STRUCT_STAT_ST_MTIM_TV_NSEC
869 mtime->tv_nsec = s.st_mtim.tv_nsec;
870 #elif HAVE_STRUCT_STAT_ST_MTIMENSEC
871 mtime->tv_nsec = s.st_mtimensec;
878 mtime->tv_sec = mtime->tv_nsec = 0;
887 VLOG_FATAL("failed to create pipe (%s)", strerror(errno));
892 xpipe_nonblocking(int fds[2])
895 xset_nonblocking(fds[0]);
896 xset_nonblocking(fds[1]);
900 xsocketpair(int domain, int type, int protocol, int fds[2])
902 if (socketpair(domain, type, protocol, fds)) {
903 VLOG_FATAL("failed to create socketpair (%s)", strerror(errno));
908 getsockopt_int(int fd, int level, int option, const char *optname, int *valuep)
910 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 10);
916 if (getsockopt(fd, level, option, &value, &len)) {
918 VLOG_ERR_RL(&rl, "getsockopt(%s): %s", optname, strerror(error));
919 } else if (len != sizeof value) {
921 VLOG_ERR_RL(&rl, "getsockopt(%s): value is %u bytes (expected %zu)",
922 optname, (unsigned int) len, sizeof value);
927 *valuep = error ? 0 : value;
932 describe_sockaddr(struct ds *string, int fd,
933 int (*getaddr)(int, struct sockaddr *, socklen_t *))
935 struct sockaddr_storage ss;
936 socklen_t len = sizeof ss;
938 if (!getaddr(fd, (struct sockaddr *) &ss, &len)) {
939 if (ss.ss_family == AF_INET) {
940 struct sockaddr_in sin;
942 memcpy(&sin, &ss, sizeof sin);
943 ds_put_format(string, IP_FMT":%"PRIu16,
944 IP_ARGS(&sin.sin_addr.s_addr), ntohs(sin.sin_port));
945 } else if (ss.ss_family == AF_UNIX) {
946 struct sockaddr_un sun;
950 memcpy(&sun, &ss, sizeof sun);
951 maxlen = len - offsetof(struct sockaddr_un, sun_path);
952 null = memchr(sun.sun_path, '\0', maxlen);
953 ds_put_buffer(string, sun.sun_path,
954 null ? null - sun.sun_path : maxlen);
957 else if (ss.ss_family == AF_NETLINK) {
960 /* SO_PROTOCOL was introduced in 2.6.32. Support it regardless of the version
961 * of the Linux kernel headers in use at build time. */
963 #define SO_PROTOCOL 38
966 if (!getsockopt_int(fd, SOL_SOCKET, SO_PROTOCOL, "SO_PROTOCOL",
970 ds_put_cstr(string, "NETLINK_ROUTE");
973 case NETLINK_GENERIC:
974 ds_put_cstr(string, "NETLINK_GENERIC");
978 ds_put_format(string, "AF_NETLINK family %d", protocol);
982 ds_put_cstr(string, "AF_NETLINK");
986 #if AF_PACKET && LINUX_DATAPATH
987 else if (ss.ss_family == AF_PACKET) {
988 struct sockaddr_ll sll;
990 memcpy(&sll, &ss, sizeof sll);
991 ds_put_cstr(string, "AF_PACKET");
992 if (sll.sll_ifindex) {
995 if (if_indextoname(sll.sll_ifindex, name)) {
996 ds_put_format(string, "(%s)", name);
998 ds_put_format(string, "(ifindex=%d)", sll.sll_ifindex);
1001 if (sll.sll_protocol) {
1002 ds_put_format(string, "(protocol=0x%"PRIu16")",
1003 ntohs(sll.sll_protocol));
1007 else if (ss.ss_family == AF_UNSPEC) {
1008 ds_put_cstr(string, "AF_UNSPEC");
1010 ds_put_format(string, "AF_%d", (int) ss.ss_family);
1016 #ifdef LINUX_DATAPATH
1018 put_fd_filename(struct ds *string, int fd)
1024 linkname = xasprintf("/proc/self/fd/%d", fd);
1025 n = readlink(linkname, buf, sizeof buf);
1027 ds_put_char(string, ' ');
1028 ds_put_buffer(string, buf, n);
1029 if (n > sizeof buf) {
1030 ds_put_cstr(string, "...");
1037 /* Returns a malloc()'d string describing 'fd', for use in logging. */
1045 if (fstat(fd, &s)) {
1046 ds_put_format(&string, "fstat failed (%s)", strerror(errno));
1047 } else if (S_ISSOCK(s.st_mode)) {
1048 describe_sockaddr(&string, fd, getsockname);
1049 ds_put_cstr(&string, "<->");
1050 describe_sockaddr(&string, fd, getpeername);
1052 ds_put_cstr(&string, (isatty(fd) ? "tty"
1053 : S_ISDIR(s.st_mode) ? "directory"
1054 : S_ISCHR(s.st_mode) ? "character device"
1055 : S_ISBLK(s.st_mode) ? "block device"
1056 : S_ISREG(s.st_mode) ? "file"
1057 : S_ISFIFO(s.st_mode) ? "FIFO"
1058 : S_ISLNK(s.st_mode) ? "symbolic link"
1060 #ifdef LINUX_DATAPATH
1061 put_fd_filename(&string, fd);
1064 return ds_steal_cstr(&string);
1067 /* Returns the total of the 'iov_len' members of the 'n_iovs' in 'iovs'.
1068 * The caller must ensure that the total does not exceed SIZE_MAX. */
1070 iovec_len(const struct iovec iovs[], size_t n_iovs)
1075 for (i = 0; i < n_iovs; i++) {
1076 len += iovs[i].iov_len;
1081 /* Returns true if all of the 'n_iovs' iovecs in 'iovs' have length zero. */
1083 iovec_is_empty(const struct iovec iovs[], size_t n_iovs)
1087 for (i = 0; i < n_iovs; i++) {
1088 if (iovs[i].iov_len) {
1095 /* Sends the 'n_iovs' iovecs of data in 'iovs' and the 'n_fds' file descriptors
1096 * in 'fds' on Unix domain socket 'sock'. Returns the number of bytes
1097 * successfully sent or -1 if an error occurred. On error, sets errno
1100 send_iovec_and_fds(int sock,
1101 const struct iovec *iovs, size_t n_iovs,
1102 const int fds[], size_t n_fds)
1108 char control[CMSG_SPACE(SOUTIL_MAX_FDS * sizeof *fds)];
1112 assert(!iovec_is_empty(iovs, n_iovs));
1113 assert(n_fds <= SOUTIL_MAX_FDS);
1115 memset(&cmsg, 0, sizeof cmsg);
1116 cmsg.cm.cmsg_len = CMSG_LEN(n_fds * sizeof *fds);
1117 cmsg.cm.cmsg_level = SOL_SOCKET;
1118 cmsg.cm.cmsg_type = SCM_RIGHTS;
1119 memcpy(CMSG_DATA(&cmsg.cm), fds, n_fds * sizeof *fds);
1121 msg.msg_name = NULL;
1122 msg.msg_namelen = 0;
1123 msg.msg_iov = (struct iovec *) iovs;
1124 msg.msg_iovlen = n_iovs;
1125 msg.msg_control = &cmsg.cm;
1126 msg.msg_controllen = CMSG_SPACE(n_fds * sizeof *fds);
1129 return sendmsg(sock, &msg, 0);
1131 return writev(sock, iovs, n_iovs);
1135 /* Sends the 'n_iovs' iovecs of data in 'iovs' and the 'n_fds' file descriptors
1136 * in 'fds' on Unix domain socket 'sock'. If 'skip_bytes' is nonzero, then the
1137 * first 'skip_bytes' of data in the iovecs are not sent, and none of the file
1138 * descriptors are sent. The function continues to retry sending until an
1139 * error (other than EINTR) occurs or all the data and fds are sent.
1141 * Returns 0 if all the data and fds were successfully sent, otherwise a
1142 * positive errno value. Regardless of success, stores the number of bytes
1143 * sent (always at least 'skip_bytes') in '*bytes_sent'. (If at least one byte
1144 * is sent, then all the fds have been sent.)
1146 * 'skip_bytes' must be less than or equal to iovec_len(iovs, n_iovs). */
1148 send_iovec_and_fds_fully(int sock,
1149 const struct iovec iovs[], size_t n_iovs,
1150 const int fds[], size_t n_fds,
1151 size_t skip_bytes, size_t *bytes_sent)
1154 while (n_iovs > 0) {
1158 retval = skip_bytes;
1160 } else if (!*bytes_sent) {
1161 retval = send_iovec_and_fds(sock, iovs, n_iovs, fds, n_fds);
1163 retval = writev(sock, iovs, n_iovs);
1167 *bytes_sent += retval;
1168 while (retval > 0) {
1169 const uint8_t *base = iovs->iov_base;
1170 size_t len = iovs->iov_len;
1176 error = write_fully(sock, base + retval, len - retval,
1178 *bytes_sent += sent;
1188 } else if (retval == 0) {
1189 if (iovec_is_empty(iovs, n_iovs)) {
1192 VLOG_WARN("send returned 0");
1194 } else if (errno != EINTR) {
1202 /* Sends the 'n_iovs' iovecs of data in 'iovs' and the 'n_fds' file descriptors
1203 * in 'fds' on Unix domain socket 'sock'. The function continues to retry
1204 * sending until an error (other than EAGAIN or EINTR) occurs or all the data
1205 * and fds are sent. Upon EAGAIN, the function blocks until the socket is
1206 * ready for more data.
1208 * Returns 0 if all the data and fds were successfully sent, otherwise a
1209 * positive errno value. */
1211 send_iovec_and_fds_fully_block(int sock,
1212 const struct iovec iovs[], size_t n_iovs,
1213 const int fds[], size_t n_fds)
1220 error = send_iovec_and_fds_fully(sock, iovs, n_iovs,
1221 fds, n_fds, sent, &sent);
1222 if (error != EAGAIN) {
1225 poll_fd_wait(sock, POLLOUT);
1230 /* Attempts to receive from Unix domain socket 'sock' up to 'size' bytes of
1231 * data into 'data' and up to SOUTIL_MAX_FDS file descriptors into 'fds'.
1233 * - Upon success, returns the number of bytes of data copied into 'data'
1234 * and stores the number of received file descriptors into '*n_fdsp'.
1236 * - On failure, returns a negative errno value and stores 0 in
1239 * - On EOF, returns 0 and stores 0 in '*n_fdsp'. */
1241 recv_data_and_fds(int sock,
1242 void *data, size_t size,
1243 int fds[SOUTIL_MAX_FDS], size_t *n_fdsp)
1247 char control[CMSG_SPACE(SOUTIL_MAX_FDS * sizeof *fds)];
1259 iov.iov_base = data;
1262 msg.msg_name = NULL;
1263 msg.msg_namelen = 0;
1266 msg.msg_control = &cmsg.cm;
1267 msg.msg_controllen = sizeof cmsg.control;
1270 retval = recvmsg(sock, &msg, 0);
1271 } while (retval < 0 && errno == EINTR);
1273 return retval < 0 ? -errno : 0;
1276 for (p = CMSG_FIRSTHDR(&msg); p; p = CMSG_NXTHDR(&msg, p)) {
1277 if (p->cmsg_level != SOL_SOCKET || p->cmsg_type != SCM_RIGHTS) {
1278 VLOG_ERR("unexpected control message %d:%d",
1279 p->cmsg_level, p->cmsg_type);
1281 } else if (*n_fdsp) {
1282 VLOG_ERR("multiple SCM_RIGHTS received");
1285 size_t n_fds = (p->cmsg_len - CMSG_LEN(0)) / sizeof *fds;
1286 const int *fds_data = (const int *) CMSG_DATA(p);
1289 if (n_fds > SOUTIL_MAX_FDS) {
1290 VLOG_ERR("%zu fds received but only %d supported",
1291 n_fds, SOUTIL_MAX_FDS);
1292 for (i = 0; i < n_fds; i++) {
1299 memcpy(fds, fds_data, n_fds * sizeof *fds);
1306 for (i = 0; i < *n_fdsp; i++) {