2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013 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>
29 #include <sys/ioctl.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 /* Maximum length of the sun_path member in a struct sockaddr_un, excluding
64 * space for a null terminator. */
65 #define MAX_UN_LEN (sizeof(((struct sockaddr_un *) 0)->sun_path) - 1)
67 static int getsockopt_int(int fd, int level, int option, const char *optname,
70 /* Sets 'fd' to non-blocking mode. Returns 0 if successful, otherwise a
71 * positive errno value. */
73 set_nonblocking(int fd)
75 int flags = fcntl(fd, F_GETFL, 0);
77 if (fcntl(fd, F_SETFL, flags | O_NONBLOCK) != -1) {
80 VLOG_ERR("fcntl(F_SETFL) failed: %s", ovs_strerror(errno));
84 VLOG_ERR("fcntl(F_GETFL) failed: %s", ovs_strerror(errno));
90 xset_nonblocking(int fd)
92 if (set_nonblocking(fd)) {
98 set_dscp(int fd, uint8_t dscp)
107 if (setsockopt(fd, IPPROTO_IP, IP_TOS, &val, sizeof val)) {
115 rlim_is_finite(rlim_t limit)
117 if (limit == RLIM_INFINITY) {
121 #ifdef RLIM_SAVED_CUR /* FreeBSD 8.0 lacks RLIM_SAVED_CUR. */
122 if (limit == RLIM_SAVED_CUR) {
127 #ifdef RLIM_SAVED_MAX /* FreeBSD 8.0 lacks RLIM_SAVED_MAX. */
128 if (limit == RLIM_SAVED_MAX) {
136 /* Returns the maximum valid FD value, plus 1. */
140 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
143 if (ovsthread_once_start(&once)) {
145 if (!getrlimit(RLIMIT_NOFILE, &r) && rlim_is_finite(r.rlim_cur)) {
146 max_fds = r.rlim_cur;
148 VLOG_WARN("failed to obtain fd limit, defaulting to 1024");
151 ovsthread_once_done(&once);
157 /* Translates 'host_name', which must be a string representation of an IP
158 * address, into a numeric IP address in '*addr'. Returns 0 if successful,
159 * otherwise a positive errno value. */
161 lookup_ip(const char *host_name, struct in_addr *addr)
163 if (!inet_aton(host_name, addr)) {
164 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
165 VLOG_ERR_RL(&rl, "\"%s\" is not a valid IP address", host_name);
171 /* Translates 'host_name', which must be a string representation of an IPv6
172 * address, into a numeric IPv6 address in '*addr'. Returns 0 if successful,
173 * otherwise a positive errno value. */
175 lookup_ipv6(const char *host_name, struct in6_addr *addr)
177 if (inet_pton(AF_INET6, host_name, addr) != 1) {
178 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
179 VLOG_ERR_RL(&rl, "\"%s\" is not a valid IPv6 address", host_name);
185 /* Translates 'host_name', which must be a host name or a string representation
186 * of an IP address, into a numeric IP address in '*addr'. Returns 0 if
187 * successful, otherwise a positive errno value.
189 * Most Open vSwitch code should not use this because it causes deadlocks:
190 * getaddrinfo() sends out a DNS request but that starts a new flow for which
191 * OVS must set up a flow, but it can't because it's waiting for a DNS reply.
192 * The synchronous lookup also delays other activity. (Of course we can solve
193 * this but it doesn't seem worthwhile quite yet.) */
195 lookup_hostname(const char *host_name, struct in_addr *addr)
197 struct addrinfo *result;
198 struct addrinfo hints;
200 if (inet_aton(host_name, addr)) {
204 memset(&hints, 0, sizeof hints);
205 hints.ai_family = AF_INET;
207 switch (getaddrinfo(host_name, NULL, &hints, &result)) {
209 *addr = ALIGNED_CAST(struct sockaddr_in *,
210 result->ai_addr)->sin_addr;
211 freeaddrinfo(result);
214 #ifdef EAI_ADDRFAMILY
249 check_connection_completion(int fd)
251 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 10);
256 pfd.events = POLLOUT;
258 retval = poll(&pfd, 1, 0);
259 } while (retval < 0 && errno == EINTR);
261 if (pfd.revents & POLLERR) {
262 ssize_t n = send(fd, "", 1, MSG_DONTWAIT);
266 VLOG_ERR_RL(&rl, "poll return POLLERR but send succeeded");
271 } else if (retval < 0) {
272 VLOG_ERR_RL(&rl, "poll: %s", ovs_strerror(errno));
279 /* Drain all the data currently in the receive queue of a datagram socket (and
280 * possibly additional data). There is no way to know how many packets are in
281 * the receive queue, but we do know that the total number of bytes queued does
282 * not exceed the receive buffer size, so we pull packets until none are left
283 * or we've read that many bytes. */
289 rcvbuf = get_socket_rcvbuf(fd);
295 /* In Linux, specifying MSG_TRUNC in the flags argument causes the
296 * datagram length to be returned, even if that is longer than the
297 * buffer provided. Thus, we can use a 1-byte buffer to discard the
298 * incoming datagram and still be able to account how many bytes were
299 * removed from the receive buffer.
301 * On other Unix-like OSes, MSG_TRUNC has no effect in the flags
303 char buffer[LINUX_DATAPATH ? 1 : 2048];
304 ssize_t n_bytes = recv(fd, buffer, sizeof buffer,
305 MSG_TRUNC | MSG_DONTWAIT);
306 if (n_bytes <= 0 || n_bytes >= rcvbuf) {
314 /* Returns the size of socket 'sock''s receive buffer (SO_RCVBUF), or a
315 * negative errno value if an error occurs. */
317 get_socket_rcvbuf(int sock)
322 error = getsockopt_int(sock, SOL_SOCKET, SO_RCVBUF, "SO_RCVBUF", &rcvbuf);
323 return error ? -error : rcvbuf;
326 /* Reads and discards up to 'n' datagrams from 'fd', stopping as soon as no
327 * more data can be immediately read. ('fd' should therefore be in
328 * non-blocking mode.)*/
330 drain_fd(int fd, size_t n_packets)
332 for (; n_packets > 0; n_packets--) {
333 /* 'buffer' only needs to be 1 byte long in most circumstances. This
334 * size is defensive against the possibility that we someday want to
335 * use a Linux tap device without TUN_NO_PI, in which case a buffer
336 * smaller than sizeof(struct tun_pi) will give EINVAL on read. */
338 if (read(fd, buffer, sizeof buffer) <= 0) {
344 /* Attempts to shorten 'name' by opening a file descriptor for the directory
345 * part of the name and indirecting through /proc/self/fd/<dirfd>/<basename>.
346 * On systems with Linux-like /proc, this works as long as <basename> isn't too
349 * On success, returns 0 and stores the short name in 'short_name' and a
350 * directory file descriptor to eventually be closed in '*dirfpd'. */
352 shorten_name_via_proc(const char *name, char short_name[MAX_UN_LEN + 1],
359 if (!LINUX_DATAPATH) {
363 dir = dir_name(name);
364 dirfd = open(dir, O_DIRECTORY | O_RDONLY);
366 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
369 VLOG_WARN_RL(&rl, "%s: open failed (%s)", dir, ovs_strerror(error));
376 base = base_name(name);
377 len = snprintf(short_name, MAX_UN_LEN + 1,
378 "/proc/self/fd/%d/%s", dirfd, base);
381 if (len >= 0 && len <= MAX_UN_LEN) {
390 /* Attempts to shorten 'name' by creating a symlink for the directory part of
391 * the name and indirecting through <symlink>/<basename>. This works on
392 * systems that support symlinks, as long as <basename> isn't too long.
394 * On success, returns 0 and stores the short name in 'short_name' and the
395 * symbolic link to eventually delete in 'linkname'. */
397 shorten_name_via_symlink(const char *name, char short_name[MAX_UN_LEN + 1],
398 char linkname[MAX_UN_LEN + 1])
400 char *abs, *dir, *base;
405 abs = abs_file_name(NULL, name);
407 base = base_name(abs);
410 tmpdir = getenv("TMPDIR");
411 if (tmpdir == NULL) {
415 for (i = 0; i < 1000; i++) {
418 len = snprintf(linkname, MAX_UN_LEN + 1,
419 "%s/ovs-un-c-%"PRIu32, tmpdir, random_uint32());
420 error = (len < 0 || len > MAX_UN_LEN ? ENAMETOOLONG
421 : symlink(dir, linkname) ? errno
423 if (error != EEXIST) {
431 fatal_signal_add_file_to_unlink(linkname);
433 len = snprintf(short_name, MAX_UN_LEN + 1, "%s/%s", linkname, base);
434 if (len < 0 || len > MAX_UN_LEN) {
435 fatal_signal_unlink_file_now(linkname);
436 error = ENAMETOOLONG;
449 /* Stores in '*un' a sockaddr_un that refers to file 'name'. Stores in
450 * '*un_len' the size of the sockaddr_un.
452 * Returns 0 on success, otherwise a positive errno value.
454 * Uses '*dirfdp' and 'linkname' to store references to data when the caller no
455 * longer needs to use 'un'. On success, freeing these references with
456 * free_sockaddr_un() is mandatory to avoid a leak; on failure, freeing them is
457 * unnecessary but harmless. */
459 make_sockaddr_un(const char *name, struct sockaddr_un *un, socklen_t *un_len,
460 int *dirfdp, char linkname[MAX_UN_LEN + 1])
462 char short_name[MAX_UN_LEN + 1];
466 if (strlen(name) > MAX_UN_LEN) {
467 /* 'name' is too long to fit in a sockaddr_un. Try a workaround. */
468 int error = shorten_name_via_proc(name, short_name, dirfdp);
469 if (error == ENAMETOOLONG) {
470 error = shorten_name_via_symlink(name, short_name, linkname);
473 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
475 VLOG_WARN_RL(&rl, "Unix socket name %s is longer than maximum "
476 "%zu bytes", name, MAX_UN_LEN);
483 un->sun_family = AF_UNIX;
484 ovs_strzcpy(un->sun_path, name, sizeof un->sun_path);
485 *un_len = (offsetof(struct sockaddr_un, sun_path)
486 + strlen (un->sun_path) + 1);
490 /* Clean up after make_sockaddr_un(). */
492 free_sockaddr_un(int dirfd, const char *linkname)
498 fatal_signal_unlink_file_now(linkname);
502 /* Binds Unix domain socket 'fd' to a file with permissions 0700. */
504 bind_unix_socket(int fd, struct sockaddr *sun, socklen_t sun_len)
506 /* According to _Unix Network Programming_, umask should affect bind(). */
507 mode_t old_umask = umask(0077);
508 int error = bind(fd, sun, sun_len) ? errno : 0;
513 /* Creates a Unix domain socket in the given 'style' (either SOCK_DGRAM or
514 * SOCK_STREAM) that is bound to '*bind_path' (if 'bind_path' is non-null) and
515 * connected to '*connect_path' (if 'connect_path' is non-null). If 'nonblock'
516 * is true, the socket is made non-blocking.
518 * Returns the socket's fd if successful, otherwise a negative errno value. */
520 make_unix_socket(int style, bool nonblock,
521 const char *bind_path, const char *connect_path)
526 fd = socket(PF_UNIX, style, 0);
531 /* Set nonblocking mode right away, if we want it. This prevents blocking
532 * in connect(), if connect_path != NULL. (In turn, that's a corner case:
533 * it will only happen if style is SOCK_STREAM or SOCK_SEQPACKET, and only
534 * if a backlog of un-accepted connections has built up in the kernel.) */
536 error = set_nonblocking(fd);
543 char linkname[MAX_UN_LEN + 1];
544 struct sockaddr_un un;
548 if (unlink(bind_path) && errno != ENOENT) {
549 VLOG_WARN("unlinking \"%s\": %s\n",
550 bind_path, ovs_strerror(errno));
552 fatal_signal_add_file_to_unlink(bind_path);
554 error = make_sockaddr_un(bind_path, &un, &un_len, &dirfd, linkname);
556 error = bind_unix_socket(fd, (struct sockaddr *) &un, un_len);
558 free_sockaddr_un(dirfd, linkname);
566 char linkname[MAX_UN_LEN + 1];
567 struct sockaddr_un un;
571 error = make_sockaddr_un(connect_path, &un, &un_len, &dirfd, linkname);
573 && connect(fd, (struct sockaddr*) &un, un_len)
574 && errno != EINPROGRESS) {
577 free_sockaddr_un(dirfd, linkname);
587 if (error == EAGAIN) {
591 fatal_signal_unlink_file_now(bind_path);
598 get_unix_name_len(socklen_t sun_len)
600 return (sun_len >= offsetof(struct sockaddr_un, sun_path)
601 ? sun_len - offsetof(struct sockaddr_un, sun_path)
606 guess_netmask(ovs_be32 ip_)
608 uint32_t ip = ntohl(ip_);
609 return ((ip >> 31) == 0 ? htonl(0xff000000) /* Class A */
610 : (ip >> 30) == 2 ? htonl(0xffff0000) /* Class B */
611 : (ip >> 29) == 6 ? htonl(0xffffff00) /* Class C */
612 : htonl(0)); /* ??? */
615 /* Parses 'target', which should be a string in the format "<host>[:<port>]".
616 * <host> is required. If 'default_port' is nonzero then <port> is optional
617 * and defaults to 'default_port'.
619 * On success, returns true and stores the parsed remote address into '*sinp'.
620 * On failure, logs an error, stores zeros into '*sinp', and returns false. */
622 inet_parse_active(const char *target_, uint16_t default_port,
623 struct sockaddr_in *sinp)
625 char *target = xstrdup(target_);
626 char *save_ptr = NULL;
627 const char *host_name;
628 const char *port_string;
632 sinp->sin_family = AF_INET;
633 sinp->sin_port = htons(default_port);
636 host_name = strtok_r(target, ":", &save_ptr);
637 port_string = strtok_r(NULL, ":", &save_ptr);
639 VLOG_ERR("%s: bad peer name format", target_);
643 /* Look up IP, port. */
644 if (lookup_ip(host_name, &sinp->sin_addr)) {
647 if (port_string && atoi(port_string)) {
648 sinp->sin_port = htons(atoi(port_string));
649 } else if (!default_port) {
650 VLOG_ERR("%s: port number must be specified", target_);
658 memset(sinp, 0, sizeof *sinp);
664 /* Opens a non-blocking IPv4 socket of the specified 'style' and connects to
665 * 'target', which should be a string in the format "<host>[:<port>]". <host>
666 * is required. If 'default_port' is nonzero then <port> is optional and
667 * defaults to 'default_port'.
669 * 'style' should be SOCK_STREAM (for TCP) or SOCK_DGRAM (for UDP).
671 * On success, returns 0 (indicating connection complete) or EAGAIN (indicating
672 * connection in progress), in which case the new file descriptor is stored
673 * into '*fdp'. On failure, returns a positive errno value other than EAGAIN
674 * and stores -1 into '*fdp'.
676 * If 'sinp' is non-null, then on success the target address is stored into
679 * 'dscp' becomes the DSCP bits in the IP headers for the new connection. It
680 * should be in the range [0, 63] and will automatically be shifted to the
681 * appropriately place in the IP tos field. */
683 inet_open_active(int style, const char *target, uint16_t default_port,
684 struct sockaddr_in *sinp, int *fdp, uint8_t dscp)
686 struct sockaddr_in sin;
691 if (!inet_parse_active(target, default_port, &sin)) {
692 error = EAFNOSUPPORT;
696 /* Create non-blocking socket. */
697 fd = socket(AF_INET, style, 0);
699 VLOG_ERR("%s: socket: %s", target, ovs_strerror(errno));
703 error = set_nonblocking(fd);
708 /* The dscp bits must be configured before connect() to ensure that the TOS
709 * field is set during the connection establishment. If set after
710 * connect(), the handshake SYN frames will be sent with a TOS of 0. */
711 error = set_dscp(fd, dscp);
713 VLOG_ERR("%s: socket: %s", target, ovs_strerror(error));
718 error = connect(fd, (struct sockaddr *) &sin, sizeof sin) == 0 ? 0 : errno;
719 if (error == EINPROGRESS) {
724 if (!error || error == EAGAIN) {
728 } else if (fd >= 0) {
736 /* Parses 'target', which should be a string in the format "[<port>][:<ip>]":
738 * - If 'default_port' is -1, then <port> is required. Otherwise, if
739 * <port> is omitted, then 'default_port' is used instead.
741 * - If <port> (or 'default_port', if used) is 0, then no port is bound
742 * and the TCP/IP stack will select a port.
744 * - If <ip> is omitted then the IP address is wildcarded.
746 * If successful, stores the address into '*sinp' and returns true; otherwise
747 * zeros '*sinp' and returns false. */
749 inet_parse_passive(const char *target_, int default_port,
750 struct sockaddr_in *sinp)
752 char *target = xstrdup(target_);
753 char *string_ptr = target;
754 const char *host_name;
755 const char *port_string;
759 /* Address defaults. */
760 memset(sinp, 0, sizeof *sinp);
761 sinp->sin_family = AF_INET;
762 sinp->sin_addr.s_addr = htonl(INADDR_ANY);
763 sinp->sin_port = htons(default_port);
765 /* Parse optional port number. */
766 port_string = strsep(&string_ptr, ":");
767 if (port_string && str_to_int(port_string, 10, &port)) {
768 sinp->sin_port = htons(port);
769 } else if (default_port < 0) {
770 VLOG_ERR("%s: port number must be specified", target_);
774 /* Parse optional bind IP. */
775 host_name = strsep(&string_ptr, ":");
776 if (host_name && host_name[0] && lookup_ip(host_name, &sinp->sin_addr)) {
784 memset(sinp, 0, sizeof *sinp);
791 /* Opens a non-blocking IPv4 socket of the specified 'style', binds to
792 * 'target', and listens for incoming connections. Parses 'target' in the same
793 * way was inet_parse_passive().
795 * 'style' should be SOCK_STREAM (for TCP) or SOCK_DGRAM (for UDP).
797 * For TCP, the socket will have SO_REUSEADDR turned on.
799 * On success, returns a non-negative file descriptor. On failure, returns a
800 * negative errno value.
802 * If 'sinp' is non-null, then on success the bound address is stored into
805 * 'dscp' becomes the DSCP bits in the IP headers for the new connection. It
806 * should be in the range [0, 63] and will automatically be shifted to the
807 * appropriately place in the IP tos field. */
809 inet_open_passive(int style, const char *target, int default_port,
810 struct sockaddr_in *sinp, uint8_t dscp)
812 bool kernel_chooses_port;
813 struct sockaddr_in sin;
815 unsigned int yes = 1;
817 if (!inet_parse_passive(target, default_port, &sin)) {
818 return -EAFNOSUPPORT;
821 /* Create non-blocking socket, set SO_REUSEADDR. */
822 fd = socket(AF_INET, style, 0);
825 VLOG_ERR("%s: socket: %s", target, ovs_strerror(error));
828 error = set_nonblocking(fd);
832 if (style == SOCK_STREAM
833 && setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &yes, sizeof yes) < 0) {
835 VLOG_ERR("%s: setsockopt(SO_REUSEADDR): %s",
836 target, ovs_strerror(error));
841 if (bind(fd, (struct sockaddr *) &sin, sizeof sin) < 0) {
843 VLOG_ERR("%s: bind: %s", target, ovs_strerror(error));
847 /* The dscp bits must be configured before connect() to ensure that the TOS
848 * field is set during the connection establishment. If set after
849 * connect(), the handshake SYN frames will be sent with a TOS of 0. */
850 error = set_dscp(fd, dscp);
852 VLOG_ERR("%s: socket: %s", target, ovs_strerror(error));
857 if (style == SOCK_STREAM && listen(fd, 10) < 0) {
859 VLOG_ERR("%s: listen: %s", target, ovs_strerror(error));
863 kernel_chooses_port = sin.sin_port == htons(0);
864 if (sinp || kernel_chooses_port) {
865 socklen_t sin_len = sizeof sin;
866 if (getsockname(fd, (struct sockaddr *) &sin, &sin_len) < 0) {
868 VLOG_ERR("%s: getsockname: %s", target, ovs_strerror(error));
871 if (sin.sin_family != AF_INET || sin_len != sizeof sin) {
872 error = EAFNOSUPPORT;
873 VLOG_ERR("%s: getsockname: invalid socket name", target);
879 if (kernel_chooses_port) {
880 VLOG_INFO("%s: listening on port %"PRIu16,
881 target, ntohs(sin.sin_port));
892 /* Returns a readable and writable fd for /dev/null, if successful, otherwise
893 * a negative errno value. The caller must not close the returned fd (because
894 * the same fd will be handed out to subsequent callers). */
898 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
901 if (ovsthread_once_start(&once)) {
902 null_fd = open("/dev/null", O_RDWR);
905 VLOG_ERR("could not open /dev/null: %s", ovs_strerror(error));
908 ovsthread_once_done(&once);
915 read_fully(int fd, void *p_, size_t size, size_t *bytes_read)
921 ssize_t retval = read(fd, p, size);
923 *bytes_read += retval;
926 } else if (retval == 0) {
928 } else if (errno != EINTR) {
936 write_fully(int fd, const void *p_, size_t size, size_t *bytes_written)
938 const uint8_t *p = p_;
942 ssize_t retval = write(fd, p, size);
944 *bytes_written += retval;
947 } else if (retval == 0) {
948 VLOG_WARN("write returned 0");
950 } else if (errno != EINTR) {
957 /* Given file name 'file_name', fsyncs the directory in which it is contained.
958 * Returns 0 if successful, otherwise a positive errno value. */
960 fsync_parent_dir(const char *file_name)
966 dir = dir_name(file_name);
967 fd = open(dir, O_RDONLY);
970 if (errno == EINVAL || errno == EROFS) {
971 /* This directory does not support synchronization. Not
972 * really an error. */
975 VLOG_ERR("%s: fsync failed (%s)", dir, ovs_strerror(error));
981 VLOG_ERR("%s: open failed (%s)", dir, ovs_strerror(error));
988 /* Obtains the modification time of the file named 'file_name' to the greatest
989 * supported precision. If successful, stores the mtime in '*mtime' and
990 * returns 0. On error, returns a positive errno value and stores zeros in
993 get_mtime(const char *file_name, struct timespec *mtime)
997 if (!stat(file_name, &s)) {
998 mtime->tv_sec = s.st_mtime;
1000 #if HAVE_STRUCT_STAT_ST_MTIM_TV_NSEC
1001 mtime->tv_nsec = s.st_mtim.tv_nsec;
1002 #elif HAVE_STRUCT_STAT_ST_MTIMENSEC
1003 mtime->tv_nsec = s.st_mtimensec;
1010 mtime->tv_sec = mtime->tv_nsec = 0;
1019 VLOG_FATAL("failed to create pipe (%s)", ovs_strerror(errno));
1024 xpipe_nonblocking(int fds[2])
1027 xset_nonblocking(fds[0]);
1028 xset_nonblocking(fds[1]);
1032 xsocketpair(int domain, int type, int protocol, int fds[2])
1034 if (socketpair(domain, type, protocol, fds)) {
1035 VLOG_FATAL("failed to create socketpair (%s)", ovs_strerror(errno));
1040 getsockopt_int(int fd, int level, int option, const char *optname, int *valuep)
1042 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 10);
1048 if (getsockopt(fd, level, option, &value, &len)) {
1050 VLOG_ERR_RL(&rl, "getsockopt(%s): %s", optname, ovs_strerror(error));
1051 } else if (len != sizeof value) {
1053 VLOG_ERR_RL(&rl, "getsockopt(%s): value is %u bytes (expected %zu)",
1054 optname, (unsigned int) len, sizeof value);
1059 *valuep = error ? 0 : value;
1064 describe_sockaddr(struct ds *string, int fd,
1065 int (*getaddr)(int, struct sockaddr *, socklen_t *))
1067 struct sockaddr_storage ss;
1068 socklen_t len = sizeof ss;
1070 if (!getaddr(fd, (struct sockaddr *) &ss, &len)) {
1071 if (ss.ss_family == AF_INET) {
1072 struct sockaddr_in sin;
1074 memcpy(&sin, &ss, sizeof sin);
1075 ds_put_format(string, IP_FMT":%"PRIu16,
1076 IP_ARGS(sin.sin_addr.s_addr), ntohs(sin.sin_port));
1077 } else if (ss.ss_family == AF_UNIX) {
1078 struct sockaddr_un sun;
1082 memcpy(&sun, &ss, sizeof sun);
1083 maxlen = len - offsetof(struct sockaddr_un, sun_path);
1084 null = memchr(sun.sun_path, '\0', maxlen);
1085 ds_put_buffer(string, sun.sun_path,
1086 null ? null - sun.sun_path : maxlen);
1089 else if (ss.ss_family == AF_NETLINK) {
1092 /* SO_PROTOCOL was introduced in 2.6.32. Support it regardless of the version
1093 * of the Linux kernel headers in use at build time. */
1095 #define SO_PROTOCOL 38
1098 if (!getsockopt_int(fd, SOL_SOCKET, SO_PROTOCOL, "SO_PROTOCOL",
1102 ds_put_cstr(string, "NETLINK_ROUTE");
1105 case NETLINK_GENERIC:
1106 ds_put_cstr(string, "NETLINK_GENERIC");
1110 ds_put_format(string, "AF_NETLINK family %d", protocol);
1114 ds_put_cstr(string, "AF_NETLINK");
1118 #if AF_PACKET && LINUX_DATAPATH
1119 else if (ss.ss_family == AF_PACKET) {
1120 struct sockaddr_ll sll;
1122 memcpy(&sll, &ss, sizeof sll);
1123 ds_put_cstr(string, "AF_PACKET");
1124 if (sll.sll_ifindex) {
1125 char name[IFNAMSIZ];
1127 if (if_indextoname(sll.sll_ifindex, name)) {
1128 ds_put_format(string, "(%s)", name);
1130 ds_put_format(string, "(ifindex=%d)", sll.sll_ifindex);
1133 if (sll.sll_protocol) {
1134 ds_put_format(string, "(protocol=0x%"PRIu16")",
1135 ntohs(sll.sll_protocol));
1139 else if (ss.ss_family == AF_UNSPEC) {
1140 ds_put_cstr(string, "AF_UNSPEC");
1142 ds_put_format(string, "AF_%d", (int) ss.ss_family);
1148 #ifdef LINUX_DATAPATH
1150 put_fd_filename(struct ds *string, int fd)
1156 linkname = xasprintf("/proc/self/fd/%d", fd);
1157 n = readlink(linkname, buf, sizeof buf);
1159 ds_put_char(string, ' ');
1160 ds_put_buffer(string, buf, n);
1161 if (n > sizeof buf) {
1162 ds_put_cstr(string, "...");
1169 /* Returns a malloc()'d string describing 'fd', for use in logging. */
1177 if (fstat(fd, &s)) {
1178 ds_put_format(&string, "fstat failed (%s)", ovs_strerror(errno));
1179 } else if (S_ISSOCK(s.st_mode)) {
1180 describe_sockaddr(&string, fd, getsockname);
1181 ds_put_cstr(&string, "<->");
1182 describe_sockaddr(&string, fd, getpeername);
1184 ds_put_cstr(&string, (isatty(fd) ? "tty"
1185 : S_ISDIR(s.st_mode) ? "directory"
1186 : S_ISCHR(s.st_mode) ? "character device"
1187 : S_ISBLK(s.st_mode) ? "block device"
1188 : S_ISREG(s.st_mode) ? "file"
1189 : S_ISFIFO(s.st_mode) ? "FIFO"
1190 : S_ISLNK(s.st_mode) ? "symbolic link"
1192 #ifdef LINUX_DATAPATH
1193 put_fd_filename(&string, fd);
1196 return ds_steal_cstr(&string);
1199 /* Returns the total of the 'iov_len' members of the 'n_iovs' in 'iovs'.
1200 * The caller must ensure that the total does not exceed SIZE_MAX. */
1202 iovec_len(const struct iovec iovs[], size_t n_iovs)
1207 for (i = 0; i < n_iovs; i++) {
1208 len += iovs[i].iov_len;
1213 /* Returns true if all of the 'n_iovs' iovecs in 'iovs' have length zero. */
1215 iovec_is_empty(const struct iovec iovs[], size_t n_iovs)
1219 for (i = 0; i < n_iovs; i++) {
1220 if (iovs[i].iov_len) {
1227 /* Sends the 'n_iovs' iovecs of data in 'iovs' and the 'n_fds' file descriptors
1228 * in 'fds' on Unix domain socket 'sock'. Returns the number of bytes
1229 * successfully sent or -1 if an error occurred. On error, sets errno
1232 send_iovec_and_fds(int sock,
1233 const struct iovec *iovs, size_t n_iovs,
1234 const int fds[], size_t n_fds)
1236 ovs_assert(sock >= 0);
1240 char control[CMSG_SPACE(SOUTIL_MAX_FDS * sizeof *fds)];
1244 ovs_assert(!iovec_is_empty(iovs, n_iovs));
1245 ovs_assert(n_fds <= SOUTIL_MAX_FDS);
1247 memset(&cmsg, 0, sizeof cmsg);
1248 cmsg.cm.cmsg_len = CMSG_LEN(n_fds * sizeof *fds);
1249 cmsg.cm.cmsg_level = SOL_SOCKET;
1250 cmsg.cm.cmsg_type = SCM_RIGHTS;
1251 memcpy(CMSG_DATA(&cmsg.cm), fds, n_fds * sizeof *fds);
1253 msg.msg_name = NULL;
1254 msg.msg_namelen = 0;
1255 msg.msg_iov = CONST_CAST(struct iovec *, iovs);
1256 msg.msg_iovlen = n_iovs;
1257 msg.msg_control = &cmsg.cm;
1258 msg.msg_controllen = CMSG_SPACE(n_fds * sizeof *fds);
1261 return sendmsg(sock, &msg, 0);
1263 return writev(sock, iovs, n_iovs);
1267 /* Sends the 'n_iovs' iovecs of data in 'iovs' and the 'n_fds' file descriptors
1268 * in 'fds' on Unix domain socket 'sock'. If 'skip_bytes' is nonzero, then the
1269 * first 'skip_bytes' of data in the iovecs are not sent, and none of the file
1270 * descriptors are sent. The function continues to retry sending until an
1271 * error (other than EINTR) occurs or all the data and fds are sent.
1273 * Returns 0 if all the data and fds were successfully sent, otherwise a
1274 * positive errno value. Regardless of success, stores the number of bytes
1275 * sent (always at least 'skip_bytes') in '*bytes_sent'. (If at least one byte
1276 * is sent, then all the fds have been sent.)
1278 * 'skip_bytes' must be less than or equal to iovec_len(iovs, n_iovs). */
1280 send_iovec_and_fds_fully(int sock,
1281 const struct iovec iovs[], size_t n_iovs,
1282 const int fds[], size_t n_fds,
1283 size_t skip_bytes, size_t *bytes_sent)
1286 while (n_iovs > 0) {
1290 retval = skip_bytes;
1292 } else if (!*bytes_sent) {
1293 retval = send_iovec_and_fds(sock, iovs, n_iovs, fds, n_fds);
1295 retval = writev(sock, iovs, n_iovs);
1299 *bytes_sent += retval;
1300 while (retval > 0) {
1301 const uint8_t *base = iovs->iov_base;
1302 size_t len = iovs->iov_len;
1308 error = write_fully(sock, base + retval, len - retval,
1310 *bytes_sent += sent;
1320 } else if (retval == 0) {
1321 if (iovec_is_empty(iovs, n_iovs)) {
1324 VLOG_WARN("send returned 0");
1326 } else if (errno != EINTR) {
1334 /* Sends the 'n_iovs' iovecs of data in 'iovs' and the 'n_fds' file descriptors
1335 * in 'fds' on Unix domain socket 'sock'. The function continues to retry
1336 * sending until an error (other than EAGAIN or EINTR) occurs or all the data
1337 * and fds are sent. Upon EAGAIN, the function blocks until the socket is
1338 * ready for more data.
1340 * Returns 0 if all the data and fds were successfully sent, otherwise a
1341 * positive errno value. */
1343 send_iovec_and_fds_fully_block(int sock,
1344 const struct iovec iovs[], size_t n_iovs,
1345 const int fds[], size_t n_fds)
1352 error = send_iovec_and_fds_fully(sock, iovs, n_iovs,
1353 fds, n_fds, sent, &sent);
1354 if (error != EAGAIN) {
1357 poll_fd_wait(sock, POLLOUT);
1362 /* Attempts to receive from Unix domain socket 'sock' up to 'size' bytes of
1363 * data into 'data' and up to SOUTIL_MAX_FDS file descriptors into 'fds'.
1365 * - Upon success, returns the number of bytes of data copied into 'data'
1366 * and stores the number of received file descriptors into '*n_fdsp'.
1368 * - On failure, returns a negative errno value and stores 0 in
1371 * - On EOF, returns 0 and stores 0 in '*n_fdsp'. */
1373 recv_data_and_fds(int sock,
1374 void *data, size_t size,
1375 int fds[SOUTIL_MAX_FDS], size_t *n_fdsp)
1379 char control[CMSG_SPACE(SOUTIL_MAX_FDS * sizeof *fds)];
1391 iov.iov_base = data;
1394 msg.msg_name = NULL;
1395 msg.msg_namelen = 0;
1398 msg.msg_control = &cmsg.cm;
1399 msg.msg_controllen = sizeof cmsg.control;
1402 retval = recvmsg(sock, &msg, 0);
1403 } while (retval < 0 && errno == EINTR);
1405 return retval < 0 ? -errno : 0;
1408 for (p = CMSG_FIRSTHDR(&msg); p; p = CMSG_NXTHDR(&msg, p)) {
1409 if (p->cmsg_level != SOL_SOCKET || p->cmsg_type != SCM_RIGHTS) {
1410 VLOG_ERR("unexpected control message %d:%d",
1411 p->cmsg_level, p->cmsg_type);
1413 } else if (*n_fdsp) {
1414 VLOG_ERR("multiple SCM_RIGHTS received");
1417 size_t n_fds = (p->cmsg_len - CMSG_LEN(0)) / sizeof *fds;
1418 const int *fds_data = ALIGNED_CAST(const int *, CMSG_DATA(p));
1420 ovs_assert(n_fds > 0);
1421 if (n_fds > SOUTIL_MAX_FDS) {
1422 VLOG_ERR("%zu fds received but only %d supported",
1423 n_fds, SOUTIL_MAX_FDS);
1424 for (i = 0; i < n_fds; i++) {
1431 memcpy(fds, fds_data, n_fds * sizeof *fds);
1438 for (i = 0; i < *n_fdsp; i++) {
1445 /* Calls ioctl() on an AF_INET sock, passing the specified 'command' and
1446 * 'arg'. Returns 0 if successful, otherwise a positive errno value. */
1448 af_inet_ioctl(unsigned long int command, const void *arg)
1450 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
1453 if (ovsthread_once_start(&once)) {
1454 sock = socket(AF_INET, SOCK_DGRAM, 0);
1457 VLOG_ERR("failed to create inet socket: %s", ovs_strerror(errno));
1459 ovsthread_once_done(&once);
1462 return (sock < 0 ? -sock
1463 : ioctl(sock, command, arg) == -1 ? errno
1468 af_inet_ifreq_ioctl(const char *name, struct ifreq *ifr, unsigned long int cmd,
1469 const char *cmd_name)
1473 ovs_strzcpy(ifr->ifr_name, name, sizeof ifr->ifr_name);
1474 error = af_inet_ioctl(cmd, ifr);
1476 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1477 VLOG_DBG_RL(&rl, "%s: ioctl(%s) failed: %s", name, cmd_name,
1478 ovs_strerror(error));