2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014 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"
38 #include "ovs-thread.h"
40 #include "poll-loop.h"
43 #if AF_PACKET && LINUX_DATAPATH
44 #include <linux/if_packet.h>
47 #include "netlink-protocol.h"
48 #include "netlink-socket.h"
51 VLOG_DEFINE_THIS_MODULE(socket_util);
53 /* #ifdefs make it a pain to maintain code: you have to try to build both ways.
54 * Thus, this file compiles all of the code regardless of the target, by
55 * writing "if (LINUX_DATAPATH)" instead of "#ifdef __linux__". */
56 #ifndef LINUX_DATAPATH
57 #define LINUX_DATAPATH 0
64 /* Maximum length of the sun_path member in a struct sockaddr_un, excluding
65 * space for a null terminator. */
66 #define MAX_UN_LEN (sizeof(((struct sockaddr_un *) 0)->sun_path) - 1)
68 static int getsockopt_int(int fd, int level, int option, const char *optname,
71 /* Sets 'fd' to non-blocking mode. Returns 0 if successful, otherwise a
72 * positive errno value. */
74 set_nonblocking(int fd)
77 int flags = fcntl(fd, F_GETFL, 0);
79 if (fcntl(fd, F_SETFL, flags | O_NONBLOCK) != -1) {
82 VLOG_ERR("fcntl(F_SETFL) failed: %s", ovs_strerror(errno));
86 VLOG_ERR("fcntl(F_GETFL) failed: %s", ovs_strerror(errno));
90 unsigned long arg = 1;
91 if (ioctlsocket(fd, FIONBIO, &arg)) {
92 int error = sock_errno();
93 VLOG_ERR("set_nonblocking failed: %s", sock_strerror(error));
101 xset_nonblocking(int fd)
103 if (set_nonblocking(fd)) {
109 set_dscp(int fd, uint8_t dscp)
118 if (setsockopt(fd, IPPROTO_IP, IP_TOS, &val, sizeof val)) {
126 rlim_is_finite(rlim_t limit)
128 if (limit == RLIM_INFINITY) {
132 #ifdef RLIM_SAVED_CUR /* FreeBSD 8.0 lacks RLIM_SAVED_CUR. */
133 if (limit == RLIM_SAVED_CUR) {
138 #ifdef RLIM_SAVED_MAX /* FreeBSD 8.0 lacks RLIM_SAVED_MAX. */
139 if (limit == RLIM_SAVED_MAX) {
147 /* Returns the maximum valid FD value, plus 1. */
151 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
154 if (ovsthread_once_start(&once)) {
156 if (!getrlimit(RLIMIT_NOFILE, &r) && rlim_is_finite(r.rlim_cur)) {
157 max_fds = r.rlim_cur;
159 VLOG_WARN("failed to obtain fd limit, defaulting to 1024");
162 ovsthread_once_done(&once);
168 /* Translates 'host_name', which must be a string representation of an IP
169 * address, into a numeric IP address in '*addr'. Returns 0 if successful,
170 * otherwise a positive errno value. */
172 lookup_ip(const char *host_name, struct in_addr *addr)
174 if (!inet_aton(host_name, addr)) {
175 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
176 VLOG_ERR_RL(&rl, "\"%s\" is not a valid IP address", host_name);
182 /* Translates 'host_name', which must be a string representation of an IPv6
183 * address, into a numeric IPv6 address in '*addr'. Returns 0 if successful,
184 * otherwise a positive errno value. */
186 lookup_ipv6(const char *host_name, struct in6_addr *addr)
188 if (inet_pton(AF_INET6, host_name, addr) != 1) {
189 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
190 VLOG_ERR_RL(&rl, "\"%s\" is not a valid IPv6 address", host_name);
196 /* Translates 'host_name', which must be a host name or a string representation
197 * of an IP address, into a numeric IP address in '*addr'. Returns 0 if
198 * successful, otherwise a positive errno value.
200 * Most Open vSwitch code should not use this because it causes deadlocks:
201 * getaddrinfo() sends out a DNS request but that starts a new flow for which
202 * OVS must set up a flow, but it can't because it's waiting for a DNS reply.
203 * The synchronous lookup also delays other activity. (Of course we can solve
204 * this but it doesn't seem worthwhile quite yet.) */
206 lookup_hostname(const char *host_name, struct in_addr *addr)
208 struct addrinfo *result;
209 struct addrinfo hints;
211 if (inet_aton(host_name, addr)) {
215 memset(&hints, 0, sizeof hints);
216 hints.ai_family = AF_INET;
218 switch (getaddrinfo(host_name, NULL, &hints, &result)) {
220 *addr = ALIGNED_CAST(struct sockaddr_in *,
221 result->ai_addr)->sin_addr;
222 freeaddrinfo(result);
225 #ifdef EAI_ADDRFAMILY
246 #if defined (EAI_NODATA) && EAI_NODATA != EAI_NONAME
262 check_connection_completion(int fd)
264 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 10);
269 pfd.events = POLLOUT;
271 retval = poll(&pfd, 1, 0);
272 } while (retval < 0 && errno == EINTR);
274 if (pfd.revents & POLLERR) {
275 ssize_t n = send(fd, "", 1, MSG_DONTWAIT);
279 VLOG_ERR_RL(&rl, "poll return POLLERR but send succeeded");
284 } else if (retval < 0) {
285 VLOG_ERR_RL(&rl, "poll: %s", ovs_strerror(errno));
292 /* Drain all the data currently in the receive queue of a datagram socket (and
293 * possibly additional data). There is no way to know how many packets are in
294 * the receive queue, but we do know that the total number of bytes queued does
295 * not exceed the receive buffer size, so we pull packets until none are left
296 * or we've read that many bytes. */
302 rcvbuf = get_socket_rcvbuf(fd);
308 /* In Linux, specifying MSG_TRUNC in the flags argument causes the
309 * datagram length to be returned, even if that is longer than the
310 * buffer provided. Thus, we can use a 1-byte buffer to discard the
311 * incoming datagram and still be able to account how many bytes were
312 * removed from the receive buffer.
314 * On other Unix-like OSes, MSG_TRUNC has no effect in the flags
316 char buffer[LINUX_DATAPATH ? 1 : 2048];
317 ssize_t n_bytes = recv(fd, buffer, sizeof buffer,
318 MSG_TRUNC | MSG_DONTWAIT);
319 if (n_bytes <= 0 || n_bytes >= rcvbuf) {
327 /* Returns the size of socket 'sock''s receive buffer (SO_RCVBUF), or a
328 * negative errno value if an error occurs. */
330 get_socket_rcvbuf(int sock)
335 error = getsockopt_int(sock, SOL_SOCKET, SO_RCVBUF, "SO_RCVBUF", &rcvbuf);
336 return error ? -error : rcvbuf;
339 /* Reads and discards up to 'n' datagrams from 'fd', stopping as soon as no
340 * more data can be immediately read. ('fd' should therefore be in
341 * non-blocking mode.)*/
343 drain_fd(int fd, size_t n_packets)
345 for (; n_packets > 0; n_packets--) {
346 /* 'buffer' only needs to be 1 byte long in most circumstances. This
347 * size is defensive against the possibility that we someday want to
348 * use a Linux tap device without TUN_NO_PI, in which case a buffer
349 * smaller than sizeof(struct tun_pi) will give EINVAL on read. */
351 if (read(fd, buffer, sizeof buffer) <= 0) {
357 /* Attempts to shorten 'name' by opening a file descriptor for the directory
358 * part of the name and indirecting through /proc/self/fd/<dirfd>/<basename>.
359 * On systems with Linux-like /proc, this works as long as <basename> isn't too
362 * On success, returns 0 and stores the short name in 'short_name' and a
363 * directory file descriptor to eventually be closed in '*dirfpd'. */
365 shorten_name_via_proc(const char *name, char short_name[MAX_UN_LEN + 1],
372 if (!LINUX_DATAPATH) {
376 dir = dir_name(name);
377 dirfd = open(dir, O_DIRECTORY | O_RDONLY);
379 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
382 VLOG_WARN_RL(&rl, "%s: open failed (%s)", dir, ovs_strerror(error));
389 base = base_name(name);
390 len = snprintf(short_name, MAX_UN_LEN + 1,
391 "/proc/self/fd/%d/%s", dirfd, base);
394 if (len >= 0 && len <= MAX_UN_LEN) {
403 /* Attempts to shorten 'name' by creating a symlink for the directory part of
404 * the name and indirecting through <symlink>/<basename>. This works on
405 * systems that support symlinks, as long as <basename> isn't too long.
407 * On success, returns 0 and stores the short name in 'short_name' and the
408 * symbolic link to eventually delete in 'linkname'. */
410 shorten_name_via_symlink(const char *name, char short_name[MAX_UN_LEN + 1],
411 char linkname[MAX_UN_LEN + 1])
413 char *abs, *dir, *base;
418 abs = abs_file_name(NULL, name);
420 base = base_name(abs);
423 tmpdir = getenv("TMPDIR");
424 if (tmpdir == NULL) {
428 for (i = 0; i < 1000; i++) {
431 len = snprintf(linkname, MAX_UN_LEN + 1,
432 "%s/ovs-un-c-%"PRIu32, tmpdir, random_uint32());
433 error = (len < 0 || len > MAX_UN_LEN ? ENAMETOOLONG
434 : symlink(dir, linkname) ? errno
436 if (error != EEXIST) {
444 fatal_signal_add_file_to_unlink(linkname);
446 len = snprintf(short_name, MAX_UN_LEN + 1, "%s/%s", linkname, base);
447 if (len < 0 || len > MAX_UN_LEN) {
448 fatal_signal_unlink_file_now(linkname);
449 error = ENAMETOOLONG;
462 /* Stores in '*un' a sockaddr_un that refers to file 'name'. Stores in
463 * '*un_len' the size of the sockaddr_un.
465 * Returns 0 on success, otherwise a positive errno value.
467 * Uses '*dirfdp' and 'linkname' to store references to data when the caller no
468 * longer needs to use 'un'. On success, freeing these references with
469 * free_sockaddr_un() is mandatory to avoid a leak; on failure, freeing them is
470 * unnecessary but harmless. */
472 make_sockaddr_un(const char *name, struct sockaddr_un *un, socklen_t *un_len,
473 int *dirfdp, char linkname[MAX_UN_LEN + 1])
475 char short_name[MAX_UN_LEN + 1];
479 if (strlen(name) > MAX_UN_LEN) {
480 /* 'name' is too long to fit in a sockaddr_un. Try a workaround. */
481 int error = shorten_name_via_proc(name, short_name, dirfdp);
482 if (error == ENAMETOOLONG) {
483 error = shorten_name_via_symlink(name, short_name, linkname);
486 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1);
488 VLOG_WARN_RL(&rl, "Unix socket name %s is longer than maximum "
489 "%"PRIuSIZE" bytes", name, MAX_UN_LEN);
496 un->sun_family = AF_UNIX;
497 ovs_strzcpy(un->sun_path, name, sizeof un->sun_path);
498 *un_len = (offsetof(struct sockaddr_un, sun_path)
499 + strlen (un->sun_path) + 1);
503 /* Clean up after make_sockaddr_un(). */
505 free_sockaddr_un(int dirfd, const char *linkname)
511 fatal_signal_unlink_file_now(linkname);
515 /* Binds Unix domain socket 'fd' to a file with permissions 0700. */
517 bind_unix_socket(int fd, struct sockaddr *sun, socklen_t sun_len)
519 /* According to _Unix Network Programming_, umask should affect bind(). */
520 mode_t old_umask = umask(0077);
521 int error = bind(fd, sun, sun_len) ? errno : 0;
526 /* Creates a Unix domain socket in the given 'style' (either SOCK_DGRAM or
527 * SOCK_STREAM) that is bound to '*bind_path' (if 'bind_path' is non-null) and
528 * connected to '*connect_path' (if 'connect_path' is non-null). If 'nonblock'
529 * is true, the socket is made non-blocking.
531 * Returns the socket's fd if successful, otherwise a negative errno value. */
533 make_unix_socket(int style, bool nonblock,
534 const char *bind_path, const char *connect_path)
539 fd = socket(PF_UNIX, style, 0);
544 /* Set nonblocking mode right away, if we want it. This prevents blocking
545 * in connect(), if connect_path != NULL. (In turn, that's a corner case:
546 * it will only happen if style is SOCK_STREAM or SOCK_SEQPACKET, and only
547 * if a backlog of un-accepted connections has built up in the kernel.) */
549 error = set_nonblocking(fd);
556 char linkname[MAX_UN_LEN + 1];
557 struct sockaddr_un un;
561 if (unlink(bind_path) && errno != ENOENT) {
562 VLOG_WARN("unlinking \"%s\": %s\n",
563 bind_path, ovs_strerror(errno));
565 fatal_signal_add_file_to_unlink(bind_path);
567 error = make_sockaddr_un(bind_path, &un, &un_len, &dirfd, linkname);
569 error = bind_unix_socket(fd, (struct sockaddr *) &un, un_len);
571 free_sockaddr_un(dirfd, linkname);
579 char linkname[MAX_UN_LEN + 1];
580 struct sockaddr_un un;
584 error = make_sockaddr_un(connect_path, &un, &un_len, &dirfd, linkname);
586 && connect(fd, (struct sockaddr*) &un, un_len)
587 && errno != EINPROGRESS) {
590 free_sockaddr_un(dirfd, linkname);
600 if (error == EAGAIN) {
604 fatal_signal_unlink_file_now(bind_path);
611 get_unix_name_len(socklen_t sun_len)
613 return (sun_len >= offsetof(struct sockaddr_un, sun_path)
614 ? sun_len - offsetof(struct sockaddr_un, sun_path)
619 guess_netmask(ovs_be32 ip_)
621 uint32_t ip = ntohl(ip_);
622 return ((ip >> 31) == 0 ? htonl(0xff000000) /* Class A */
623 : (ip >> 30) == 2 ? htonl(0xffff0000) /* Class B */
624 : (ip >> 29) == 6 ? htonl(0xffffff00) /* Class C */
625 : htonl(0)); /* ??? */
628 /* This is like strsep() except:
630 * - The separator string is ":".
632 * - Square brackets [] quote ":" separators and are removed from the
635 parse_bracketed_token(char **pp)
641 } else if (*p == '\0') {
644 } else if (*p == '[') {
646 char *end = start + strcspn(start, "]");
647 *pp = (*end == '\0' ? NULL
648 : end[1] == ':' ? end + 2
654 char *end = start + strcspn(start, ":");
655 *pp = *end == '\0' ? NULL : end + 1;
662 parse_sockaddr_components(struct sockaddr_storage *ss,
664 const char *port_s, uint16_t default_port,
667 struct sockaddr_in *sin = ALIGNED_CAST(struct sockaddr_in *, ss);
670 if (port_s && port_s[0]) {
671 if (!str_to_int(port_s, 10, &port) || port < 0 || port > 65535) {
672 VLOG_ERR("%s: bad port number \"%s\"", s, port_s);
678 memset(ss, 0, sizeof *ss);
679 if (strchr(host_s, ':')) {
680 struct sockaddr_in6 *sin6
681 = ALIGNED_CAST(struct sockaddr_in6 *, ss);
683 sin6->sin6_family = AF_INET6;
684 sin6->sin6_port = htons(port);
685 if (!inet_pton(AF_INET6, host_s, sin6->sin6_addr.s6_addr)) {
686 VLOG_ERR("%s: bad IPv6 address \"%s\"", s, host_s);
690 sin->sin_family = AF_INET;
691 sin->sin_port = htons(port);
692 if (!inet_pton(AF_INET, host_s, &sin->sin_addr.s_addr)) {
693 VLOG_ERR("%s: bad IPv4 address \"%s\"", s, host_s);
701 memset(ss, 0, sizeof *ss);
705 /* Parses 'target', which should be a string in the format "<host>[:<port>]".
706 * <host>, which is required, may be an IPv4 address or an IPv6 address
707 * enclosed in square brackets. If 'default_port' is nonzero then <port> is
708 * optional and defaults to 'default_port'.
710 * On success, returns true and stores the parsed remote address into '*ss'.
711 * On failure, logs an error, stores zeros into '*ss', and returns false. */
713 inet_parse_active(const char *target_, uint16_t default_port,
714 struct sockaddr_storage *ss)
716 char *target = xstrdup(target_);
723 host = parse_bracketed_token(&p);
724 port = parse_bracketed_token(&p);
726 VLOG_ERR("%s: host must be specified", target_);
728 } else if (!port && !default_port) {
729 VLOG_ERR("%s: port must be specified", target_);
732 ok = parse_sockaddr_components(ss, host, port, default_port, target_);
735 memset(ss, 0, sizeof *ss);
742 /* Opens a non-blocking IPv4 or IPv6 socket of the specified 'style' and
743 * connects to 'target', which should be a string in the format
744 * "<host>[:<port>]". <host>, which is required, may be an IPv4 address or an
745 * IPv6 address enclosed in square brackets. If 'default_port' is nonzero then
746 * <port> is optional and defaults to 'default_port'.
748 * 'style' should be SOCK_STREAM (for TCP) or SOCK_DGRAM (for UDP).
750 * On success, returns 0 (indicating connection complete) or EAGAIN (indicating
751 * connection in progress), in which case the new file descriptor is stored
752 * into '*fdp'. On failure, returns a positive errno value other than EAGAIN
753 * and stores -1 into '*fdp'.
755 * If 'ss' is non-null, then on success stores the target address into '*ss'.
757 * 'dscp' becomes the DSCP bits in the IP headers for the new connection. It
758 * should be in the range [0, 63] and will automatically be shifted to the
759 * appropriately place in the IP tos field. */
761 inet_open_active(int style, const char *target, uint16_t default_port,
762 struct sockaddr_storage *ssp, int *fdp, uint8_t dscp)
764 struct sockaddr_storage ss;
769 if (!inet_parse_active(target, default_port, &ss)) {
770 error = EAFNOSUPPORT;
774 /* Create non-blocking socket. */
775 fd = socket(ss.ss_family, style, 0);
777 error = sock_errno();
778 VLOG_ERR("%s: socket: %s", target, sock_strerror(error));
781 error = set_nonblocking(fd);
786 /* The dscp bits must be configured before connect() to ensure that the
787 * TOS field is set during the connection establishment. If set after
788 * connect(), the handshake SYN frames will be sent with a TOS of 0. */
789 error = set_dscp(fd, dscp);
791 VLOG_ERR("%s: socket: %s", target, sock_strerror(error));
796 error = connect(fd, (struct sockaddr *) &ss, ss_length(&ss)) == 0
799 if (error == EINPROGRESS
801 || error == WSAEALREADY || error == WSAEWOULDBLOCK
808 if (error && error != EAGAIN) {
810 memset(ssp, 0, sizeof *ssp);
825 /* Parses 'target', which should be a string in the format "[<port>][:<host>]":
827 * - If 'default_port' is -1, then <port> is required. Otherwise, if
828 * <port> is omitted, then 'default_port' is used instead.
830 * - If <port> (or 'default_port', if used) is 0, then no port is bound
831 * and the TCP/IP stack will select a port.
833 * - <host> is optional. If supplied, it may be an IPv4 address or an
834 * IPv6 address enclosed in square brackets. If omitted, the IP address
837 * If successful, stores the address into '*ss' and returns true; otherwise
838 * zeros '*ss' and returns false. */
840 inet_parse_passive(const char *target_, int default_port,
841 struct sockaddr_storage *ss)
843 char *target = xstrdup(target_);
850 port = parse_bracketed_token(&p);
851 host = parse_bracketed_token(&p);
852 if (!port && default_port < 0) {
853 VLOG_ERR("%s: port must be specified", target_);
856 ok = parse_sockaddr_components(ss, host ? host : "0.0.0.0",
857 port, default_port, target_);
860 memset(ss, 0, sizeof *ss);
867 /* Opens a non-blocking IPv4 or IPv6 socket of the specified 'style', binds to
868 * 'target', and listens for incoming connections. Parses 'target' in the same
869 * way was inet_parse_passive().
871 * 'style' should be SOCK_STREAM (for TCP) or SOCK_DGRAM (for UDP).
873 * For TCP, the socket will have SO_REUSEADDR turned on.
875 * On success, returns a non-negative file descriptor. On failure, returns a
876 * negative errno value.
878 * If 'ss' is non-null, then on success stores the bound address into '*ss'.
880 * 'dscp' becomes the DSCP bits in the IP headers for the new connection. It
881 * should be in the range [0, 63] and will automatically be shifted to the
882 * appropriately place in the IP tos field. */
884 inet_open_passive(int style, const char *target, int default_port,
885 struct sockaddr_storage *ssp, uint8_t dscp)
887 bool kernel_chooses_port;
888 struct sockaddr_storage ss;
890 unsigned int yes = 1;
892 if (!inet_parse_passive(target, default_port, &ss)) {
893 return -EAFNOSUPPORT;
895 kernel_chooses_port = ss_get_port(&ss) == 0;
897 /* Create non-blocking socket, set SO_REUSEADDR. */
898 fd = socket(ss.ss_family, style, 0);
900 error = sock_errno();
901 VLOG_ERR("%s: socket: %s", target, sock_strerror(error));
904 error = set_nonblocking(fd);
908 if (style == SOCK_STREAM
909 && setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &yes, sizeof yes) < 0) {
910 error = sock_errno();
911 VLOG_ERR("%s: setsockopt(SO_REUSEADDR): %s",
912 target, sock_strerror(error));
917 if (bind(fd, (struct sockaddr *) &ss, ss_length(&ss)) < 0) {
918 error = sock_errno();
919 VLOG_ERR("%s: bind: %s", target, sock_strerror(error));
923 /* The dscp bits must be configured before connect() to ensure that the TOS
924 * field is set during the connection establishment. If set after
925 * connect(), the handshake SYN frames will be sent with a TOS of 0. */
926 error = set_dscp(fd, dscp);
928 VLOG_ERR("%s: socket: %s", target, sock_strerror(error));
933 if (style == SOCK_STREAM && listen(fd, 10) < 0) {
934 error = sock_errno();
935 VLOG_ERR("%s: listen: %s", target, sock_strerror(error));
939 if (ssp || kernel_chooses_port) {
940 socklen_t ss_len = sizeof ss;
941 if (getsockname(fd, (struct sockaddr *) &ss, &ss_len) < 0) {
942 error = sock_errno();
943 VLOG_ERR("%s: getsockname: %s", target, sock_strerror(error));
946 if (kernel_chooses_port) {
947 VLOG_INFO("%s: listening on port %"PRIu16,
948 target, ss_get_port(&ss));
959 memset(ssp, 0, sizeof *ssp);
965 /* Returns a readable and writable fd for /dev/null, if successful, otherwise
966 * a negative errno value. The caller must not close the returned fd (because
967 * the same fd will be handed out to subsequent callers). */
971 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
974 if (ovsthread_once_start(&once)) {
975 null_fd = open("/dev/null", O_RDWR);
978 VLOG_ERR("could not open /dev/null: %s", ovs_strerror(error));
981 ovsthread_once_done(&once);
988 read_fully(int fd, void *p_, size_t size, size_t *bytes_read)
994 ssize_t retval = read(fd, p, size);
996 *bytes_read += retval;
999 } else if (retval == 0) {
1001 } else if (errno != EINTR) {
1009 write_fully(int fd, const void *p_, size_t size, size_t *bytes_written)
1011 const uint8_t *p = p_;
1015 ssize_t retval = write(fd, p, size);
1017 *bytes_written += retval;
1020 } else if (retval == 0) {
1021 VLOG_WARN("write returned 0");
1023 } else if (errno != EINTR) {
1030 /* Given file name 'file_name', fsyncs the directory in which it is contained.
1031 * Returns 0 if successful, otherwise a positive errno value. */
1033 fsync_parent_dir(const char *file_name)
1039 dir = dir_name(file_name);
1040 fd = open(dir, O_RDONLY);
1043 if (errno == EINVAL || errno == EROFS) {
1044 /* This directory does not support synchronization. Not
1045 * really an error. */
1048 VLOG_ERR("%s: fsync failed (%s)", dir, ovs_strerror(error));
1054 VLOG_ERR("%s: open failed (%s)", dir, ovs_strerror(error));
1061 /* Obtains the modification time of the file named 'file_name' to the greatest
1062 * supported precision. If successful, stores the mtime in '*mtime' and
1063 * returns 0. On error, returns a positive errno value and stores zeros in
1066 get_mtime(const char *file_name, struct timespec *mtime)
1070 if (!stat(file_name, &s)) {
1071 mtime->tv_sec = s.st_mtime;
1073 #if HAVE_STRUCT_STAT_ST_MTIM_TV_NSEC
1074 mtime->tv_nsec = s.st_mtim.tv_nsec;
1075 #elif HAVE_STRUCT_STAT_ST_MTIMENSEC
1076 mtime->tv_nsec = s.st_mtimensec;
1083 mtime->tv_sec = mtime->tv_nsec = 0;
1092 VLOG_FATAL("failed to create pipe (%s)", ovs_strerror(errno));
1097 xpipe_nonblocking(int fds[2])
1100 xset_nonblocking(fds[0]);
1101 xset_nonblocking(fds[1]);
1105 getsockopt_int(int fd, int level, int option, const char *optname, int *valuep)
1107 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 10);
1113 if (getsockopt(fd, level, option, &value, &len)) {
1114 error = sock_errno();
1115 VLOG_ERR_RL(&rl, "getsockopt(%s): %s", optname, sock_strerror(error));
1116 } else if (len != sizeof value) {
1118 VLOG_ERR_RL(&rl, "getsockopt(%s): value is %u bytes (expected %"PRIuSIZE")",
1119 optname, (unsigned int) len, sizeof value);
1124 *valuep = error ? 0 : value;
1129 describe_sockaddr(struct ds *string, int fd,
1130 int (*getaddr)(int, struct sockaddr *, socklen_t *))
1132 struct sockaddr_storage ss;
1133 socklen_t len = sizeof ss;
1135 if (!getaddr(fd, (struct sockaddr *) &ss, &len)) {
1136 if (ss.ss_family == AF_INET || ss.ss_family == AF_INET6) {
1137 char addrbuf[SS_NTOP_BUFSIZE];
1139 ds_put_format(string, "%s:%"PRIu16,
1140 ss_format_address(&ss, addrbuf, sizeof addrbuf),
1142 } else if (ss.ss_family == AF_UNIX) {
1143 struct sockaddr_un sun;
1147 memcpy(&sun, &ss, sizeof sun);
1148 maxlen = len - offsetof(struct sockaddr_un, sun_path);
1149 null = memchr(sun.sun_path, '\0', maxlen);
1150 ds_put_buffer(string, sun.sun_path,
1151 null ? null - sun.sun_path : maxlen);
1154 else if (ss.ss_family == AF_NETLINK) {
1157 /* SO_PROTOCOL was introduced in 2.6.32. Support it regardless of the version
1158 * of the Linux kernel headers in use at build time. */
1160 #define SO_PROTOCOL 38
1163 if (!getsockopt_int(fd, SOL_SOCKET, SO_PROTOCOL, "SO_PROTOCOL",
1167 ds_put_cstr(string, "NETLINK_ROUTE");
1170 case NETLINK_GENERIC:
1171 ds_put_cstr(string, "NETLINK_GENERIC");
1175 ds_put_format(string, "AF_NETLINK family %d", protocol);
1179 ds_put_cstr(string, "AF_NETLINK");
1183 #if AF_PACKET && LINUX_DATAPATH
1184 else if (ss.ss_family == AF_PACKET) {
1185 struct sockaddr_ll sll;
1187 memcpy(&sll, &ss, sizeof sll);
1188 ds_put_cstr(string, "AF_PACKET");
1189 if (sll.sll_ifindex) {
1190 char name[IFNAMSIZ];
1192 if (if_indextoname(sll.sll_ifindex, name)) {
1193 ds_put_format(string, "(%s)", name);
1195 ds_put_format(string, "(ifindex=%d)", sll.sll_ifindex);
1198 if (sll.sll_protocol) {
1199 ds_put_format(string, "(protocol=0x%"PRIu16")",
1200 ntohs(sll.sll_protocol));
1204 else if (ss.ss_family == AF_UNSPEC) {
1205 ds_put_cstr(string, "AF_UNSPEC");
1207 ds_put_format(string, "AF_%d", (int) ss.ss_family);
1213 #ifdef LINUX_DATAPATH
1215 put_fd_filename(struct ds *string, int fd)
1221 linkname = xasprintf("/proc/self/fd/%d", fd);
1222 n = readlink(linkname, buf, sizeof buf);
1224 ds_put_char(string, ' ');
1225 ds_put_buffer(string, buf, n);
1226 if (n > sizeof buf) {
1227 ds_put_cstr(string, "...");
1234 /* Returns a malloc()'d string describing 'fd', for use in logging. */
1242 if (fstat(fd, &s)) {
1243 ds_put_format(&string, "fstat failed (%s)", ovs_strerror(errno));
1244 } else if (S_ISSOCK(s.st_mode)) {
1245 describe_sockaddr(&string, fd, getsockname);
1246 ds_put_cstr(&string, "<->");
1247 describe_sockaddr(&string, fd, getpeername);
1249 ds_put_cstr(&string, (isatty(fd) ? "tty"
1250 : S_ISDIR(s.st_mode) ? "directory"
1251 : S_ISCHR(s.st_mode) ? "character device"
1252 : S_ISBLK(s.st_mode) ? "block device"
1253 : S_ISREG(s.st_mode) ? "file"
1254 : S_ISFIFO(s.st_mode) ? "FIFO"
1255 : S_ISLNK(s.st_mode) ? "symbolic link"
1257 #ifdef LINUX_DATAPATH
1258 put_fd_filename(&string, fd);
1261 return ds_steal_cstr(&string);
1264 /* Calls ioctl() on an AF_INET sock, passing the specified 'command' and
1265 * 'arg'. Returns 0 if successful, otherwise a positive errno value. */
1267 af_inet_ioctl(unsigned long int command, const void *arg)
1269 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
1272 if (ovsthread_once_start(&once)) {
1273 sock = socket(AF_INET, SOCK_DGRAM, 0);
1275 int error = sock_errno();
1276 VLOG_ERR("failed to create inet socket: %s", sock_strerror(error));
1279 ovsthread_once_done(&once);
1282 return (sock < 0 ? -sock
1283 : ioctl(sock, command, arg) == -1 ? errno
1288 af_inet_ifreq_ioctl(const char *name, struct ifreq *ifr, unsigned long int cmd,
1289 const char *cmd_name)
1293 ovs_strzcpy(ifr->ifr_name, name, sizeof ifr->ifr_name);
1294 error = af_inet_ioctl(cmd, ifr);
1296 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1297 VLOG_DBG_RL(&rl, "%s: ioctl(%s) failed: %s", name, cmd_name,
1298 ovs_strerror(error));
1303 /* sockaddr_storage helpers. */
1305 /* Returns the IPv4 or IPv6 port in 'ss'. */
1307 ss_get_port(const struct sockaddr_storage *ss)
1309 if (ss->ss_family == AF_INET) {
1310 const struct sockaddr_in *sin
1311 = ALIGNED_CAST(const struct sockaddr_in *, ss);
1312 return ntohs(sin->sin_port);
1313 } else if (ss->ss_family == AF_INET6) {
1314 const struct sockaddr_in6 *sin6
1315 = ALIGNED_CAST(const struct sockaddr_in6 *, ss);
1316 return ntohs(sin6->sin6_port);
1322 /* Formats the IPv4 or IPv6 address in 'ss' into the 'bufsize' bytes in 'buf'.
1323 * If 'ss' is an IPv6 address, puts square brackets around the address.
1324 * 'bufsize' should be at least SS_NTOP_BUFSIZE.
1328 ss_format_address(const struct sockaddr_storage *ss,
1329 char *buf, size_t bufsize)
1331 ovs_assert(bufsize >= SS_NTOP_BUFSIZE);
1332 if (ss->ss_family == AF_INET) {
1333 const struct sockaddr_in *sin
1334 = ALIGNED_CAST(const struct sockaddr_in *, ss);
1336 snprintf(buf, bufsize, IP_FMT, IP_ARGS(sin->sin_addr.s_addr));
1337 } else if (ss->ss_family == AF_INET6) {
1338 const struct sockaddr_in6 *sin6
1339 = ALIGNED_CAST(const struct sockaddr_in6 *, ss);
1342 inet_ntop(AF_INET6, sin6->sin6_addr.s6_addr, buf + 1, bufsize - 1);
1343 strcpy(strchr(buf, '\0'), "]");
1352 ss_length(const struct sockaddr_storage *ss)
1354 switch (ss->ss_family) {
1356 return sizeof(struct sockaddr_in);
1359 return sizeof(struct sockaddr_in6);
1366 /* For Windows socket calls, 'errno' is not set. One has to call
1367 * WSAGetLastError() to get the error number and then pass it to
1368 * this function to get the correct error string.
1370 * ovs_strerror() calls strerror_r() and would not get the correct error
1371 * string for Windows sockets, but is good for POSIX. */
1373 sock_strerror(int error)
1376 return ovs_format_message(error);
1378 return ovs_strerror(error);