1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
8 #include <netinet/ip.h>
14 #include <sys/ioctl.h>
18 #include "alloc-util.h"
19 #include "errno-util.h"
23 #include "format-util.h"
26 #include "memory-util.h"
27 #include "parse-util.h"
28 #include "path-util.h"
29 #include "process-util.h"
30 #include "socket-util.h"
31 #include "string-table.h"
32 #include "string-util.h"
34 #include "sysctl-util.h"
35 #include "user-util.h"
39 # define IDN_FLAGS NI_IDN
44 static const char* const socket_address_type_table
[] = {
45 [SOCK_STREAM
] = "Stream",
46 [SOCK_DGRAM
] = "Datagram",
48 [SOCK_RDM
] = "ReliableDatagram",
49 [SOCK_SEQPACKET
] = "SequentialPacket",
50 [SOCK_DCCP
] = "DatagramCongestionControl",
53 DEFINE_STRING_TABLE_LOOKUP(socket_address_type
, int);
55 int socket_address_verify(const SocketAddress
*a
, bool strict
) {
58 /* With 'strict' we enforce additional sanity constraints which are not set by the standard,
59 * but should only apply to sockets we create ourselves. */
61 switch (socket_address_family(a
)) {
64 if (a
->size
!= sizeof(struct sockaddr_in
))
67 if (a
->sockaddr
.in
.sin_port
== 0)
70 if (!IN_SET(a
->type
, 0, SOCK_STREAM
, SOCK_DGRAM
))
76 if (a
->size
!= sizeof(struct sockaddr_in6
))
79 if (a
->sockaddr
.in6
.sin6_port
== 0)
82 if (!IN_SET(a
->type
, 0, SOCK_STREAM
, SOCK_DGRAM
))
88 if (a
->size
< offsetof(struct sockaddr_un
, sun_path
))
90 if (a
->size
> sizeof(struct sockaddr_un
) + !strict
)
91 /* If !strict, allow one extra byte, since getsockname() on Linux will append
92 * a NUL byte if we have path sockets that are above sun_path's full size. */
95 if (a
->size
> offsetof(struct sockaddr_un
, sun_path
) &&
96 a
->sockaddr
.un
.sun_path
[0] != 0 &&
98 /* Only validate file system sockets here, and only in strict mode */
101 e
= memchr(a
->sockaddr
.un
.sun_path
, 0, sizeof(a
->sockaddr
.un
.sun_path
));
103 /* If there's an embedded NUL byte, make sure the size of the socket address matches it */
104 if (a
->size
!= offsetof(struct sockaddr_un
, sun_path
) + (e
- a
->sockaddr
.un
.sun_path
) + 1)
107 /* If there's no embedded NUL byte, then the size needs to match the whole
108 * structure or the structure with one extra NUL byte suffixed. (Yeah, Linux is awful,
109 * and considers both equivalent: getsockname() even extends sockaddr_un beyond its
110 * size if the path is non NUL terminated.)*/
111 if (!IN_SET(a
->size
, sizeof(a
->sockaddr
.un
.sun_path
), sizeof(a
->sockaddr
.un
.sun_path
)+1))
116 if (!IN_SET(a
->type
, 0, SOCK_STREAM
, SOCK_DGRAM
, SOCK_SEQPACKET
))
123 if (a
->size
!= sizeof(struct sockaddr_nl
))
126 if (!IN_SET(a
->type
, 0, SOCK_RAW
, SOCK_DGRAM
))
132 if (a
->size
!= sizeof(struct sockaddr_vm
))
135 if (!IN_SET(a
->type
, 0, SOCK_STREAM
, SOCK_DGRAM
))
141 return -EAFNOSUPPORT
;
145 int socket_address_print(const SocketAddress
*a
, char **ret
) {
151 r
= socket_address_verify(a
, false); /* We do non-strict validation, because we want to be
152 * able to pretty-print any socket the kernel considers
153 * valid. We still need to do validation to know if we
154 * can meaningfully print the address. */
158 if (socket_address_family(a
) == AF_NETLINK
) {
159 _cleanup_free_
char *sfamily
= NULL
;
161 r
= netlink_family_to_string_alloc(a
->protocol
, &sfamily
);
165 r
= asprintf(ret
, "%s %u", sfamily
, a
->sockaddr
.nl
.nl_groups
);
172 return sockaddr_pretty(&a
->sockaddr
.sa
, a
->size
, false, true, ret
);
175 bool socket_address_can_accept(const SocketAddress
*a
) {
179 IN_SET(a
->type
, SOCK_STREAM
, SOCK_SEQPACKET
);
182 bool socket_address_equal(const SocketAddress
*a
, const SocketAddress
*b
) {
186 /* Invalid addresses are unequal to all */
187 if (socket_address_verify(a
, false) < 0 ||
188 socket_address_verify(b
, false) < 0)
191 if (a
->type
!= b
->type
)
194 if (socket_address_family(a
) != socket_address_family(b
))
197 switch (socket_address_family(a
)) {
200 if (a
->sockaddr
.in
.sin_addr
.s_addr
!= b
->sockaddr
.in
.sin_addr
.s_addr
)
203 if (a
->sockaddr
.in
.sin_port
!= b
->sockaddr
.in
.sin_port
)
209 if (memcmp(&a
->sockaddr
.in6
.sin6_addr
, &b
->sockaddr
.in6
.sin6_addr
, sizeof(a
->sockaddr
.in6
.sin6_addr
)) != 0)
212 if (a
->sockaddr
.in6
.sin6_port
!= b
->sockaddr
.in6
.sin6_port
)
218 if (a
->size
<= offsetof(struct sockaddr_un
, sun_path
) ||
219 b
->size
<= offsetof(struct sockaddr_un
, sun_path
))
222 if ((a
->sockaddr
.un
.sun_path
[0] == 0) != (b
->sockaddr
.un
.sun_path
[0] == 0))
225 if (a
->sockaddr
.un
.sun_path
[0]) {
226 if (!path_equal_or_files_same(a
->sockaddr
.un
.sun_path
, b
->sockaddr
.un
.sun_path
, 0))
229 if (a
->size
!= b
->size
)
232 if (memcmp(a
->sockaddr
.un
.sun_path
, b
->sockaddr
.un
.sun_path
, a
->size
) != 0)
239 if (a
->protocol
!= b
->protocol
)
242 if (a
->sockaddr
.nl
.nl_groups
!= b
->sockaddr
.nl
.nl_groups
)
248 if (a
->sockaddr
.vm
.svm_cid
!= b
->sockaddr
.vm
.svm_cid
)
251 if (a
->sockaddr
.vm
.svm_port
!= b
->sockaddr
.vm
.svm_port
)
257 /* Cannot compare, so we assume the addresses are different */
264 const char* socket_address_get_path(const SocketAddress
*a
) {
267 if (socket_address_family(a
) != AF_UNIX
)
270 if (a
->sockaddr
.un
.sun_path
[0] == 0)
273 /* Note that this is only safe because we know that there's an extra NUL byte after the sockaddr_un
274 * structure. On Linux AF_UNIX file system socket addresses don't have to be NUL terminated if they take up the
275 * full sun_path space. */
276 assert_cc(sizeof(union sockaddr_union
) >= sizeof(struct sockaddr_un
)+1);
277 return a
->sockaddr
.un
.sun_path
;
280 bool socket_ipv6_is_supported(void) {
281 static int cached
= -1;
285 if (access("/proc/net/if_inet6", F_OK
) < 0) {
287 if (errno
!= ENOENT
) {
288 log_debug_errno(errno
, "Unexpected error when checking whether /proc/net/if_inet6 exists: %m");
300 bool socket_ipv6_is_enabled(void) {
301 _cleanup_free_
char *v
= NULL
;
304 /* Much like socket_ipv6_is_supported(), but also checks that the sysctl that disables IPv6 on all
305 * interfaces isn't turned on */
307 if (!socket_ipv6_is_supported())
310 r
= sysctl_read_ip_property(AF_INET6
, "all", "disable_ipv6", &v
);
312 log_debug_errno(r
, "Unexpected error reading 'net.ipv6.conf.all.disable_ipv6' sysctl: %m");
316 r
= parse_boolean(v
);
318 log_debug_errno(r
, "Failed to pare 'net.ipv6.conf.all.disable_ipv6' sysctl: %m");
325 bool socket_address_matches_fd(const SocketAddress
*a
, int fd
) {
332 b
.size
= sizeof(b
.sockaddr
);
333 if (getsockname(fd
, &b
.sockaddr
.sa
, &b
.size
) < 0)
336 if (b
.sockaddr
.sa
.sa_family
!= a
->sockaddr
.sa
.sa_family
)
339 solen
= sizeof(b
.type
);
340 if (getsockopt(fd
, SOL_SOCKET
, SO_TYPE
, &b
.type
, &solen
) < 0)
343 if (b
.type
!= a
->type
)
346 if (a
->protocol
!= 0) {
347 solen
= sizeof(b
.protocol
);
348 if (getsockopt(fd
, SOL_SOCKET
, SO_PROTOCOL
, &b
.protocol
, &solen
) < 0)
351 if (b
.protocol
!= a
->protocol
)
355 return socket_address_equal(a
, &b
);
358 int sockaddr_port(const struct sockaddr
*_sa
, unsigned *ret_port
) {
359 const union sockaddr_union
*sa
= (const union sockaddr_union
*) _sa
;
361 /* Note, this returns the port as 'unsigned' rather than 'uint16_t', as AF_VSOCK knows larger ports */
365 switch (sa
->sa
.sa_family
) {
368 *ret_port
= be16toh(sa
->in
.sin_port
);
372 *ret_port
= be16toh(sa
->in6
.sin6_port
);
376 *ret_port
= sa
->vm
.svm_port
;
380 return -EAFNOSUPPORT
;
384 const union in_addr_union
*sockaddr_in_addr(const struct sockaddr
*_sa
) {
385 const union sockaddr_union
*sa
= (const union sockaddr_union
*) _sa
;
390 switch (sa
->sa
.sa_family
) {
393 return (const union in_addr_union
*) &sa
->in
.sin_addr
;
396 return (const union in_addr_union
*) &sa
->in6
.sin6_addr
;
404 const struct sockaddr
*_sa
,
410 union sockaddr_union
*sa
= (union sockaddr_union
*) _sa
;
415 assert(salen
>= sizeof(sa
->sa
.sa_family
));
417 switch (sa
->sa
.sa_family
) {
422 a
= be32toh(sa
->in
.sin_addr
.s_addr
);
427 a
>> 24, (a
>> 16) & 0xFF, (a
>> 8) & 0xFF, a
& 0xFF,
428 be16toh(sa
->in
.sin_port
));
432 a
>> 24, (a
>> 16) & 0xFF, (a
>> 8) & 0xFF, a
& 0xFF);
439 static const unsigned char ipv4_prefix
[] = {
440 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0xFF, 0xFF
443 if (translate_ipv6
&&
444 memcmp(&sa
->in6
.sin6_addr
, ipv4_prefix
, sizeof(ipv4_prefix
)) == 0) {
445 const uint8_t *a
= sa
->in6
.sin6_addr
.s6_addr
+12;
449 a
[0], a
[1], a
[2], a
[3],
450 be16toh(sa
->in6
.sin6_port
));
454 a
[0], a
[1], a
[2], a
[3]);
458 char a
[INET6_ADDRSTRLEN
], ifname
[IF_NAMESIZE
+ 1];
460 inet_ntop(AF_INET6
, &sa
->in6
.sin6_addr
, a
, sizeof(a
));
461 if (sa
->in6
.sin6_scope_id
!= 0)
462 format_ifname_full(sa
->in6
.sin6_scope_id
, ifname
, FORMAT_IFNAME_IFINDEX
);
468 be16toh(sa
->in6
.sin6_port
),
469 sa
->in6
.sin6_scope_id
!= 0 ? "%" : "",
470 sa
->in6
.sin6_scope_id
!= 0 ? ifname
: "");
474 p
= sa
->in6
.sin6_scope_id
!= 0 ? strjoin(a
, "%", ifname
) : strdup(a
);
484 if (salen
<= offsetof(struct sockaddr_un
, sun_path
) ||
485 (sa
->un
.sun_path
[0] == 0 && salen
== offsetof(struct sockaddr_un
, sun_path
) + 1))
486 /* The name must have at least one character (and the leading NUL does not count) */
487 p
= strdup("<unnamed>");
489 /* Note that we calculate the path pointer here through the .un_buffer[] field, in order to
490 * outtrick bounds checking tools such as ubsan, which are too smart for their own good: on
491 * Linux the kernel may return sun_path[] data one byte longer than the declared size of the
493 char *path
= (char*) sa
->un_buffer
+ offsetof(struct sockaddr_un
, sun_path
);
494 size_t path_len
= salen
- offsetof(struct sockaddr_un
, sun_path
);
497 /* Abstract socket. When parsing address information from, we
498 * explicitly reject overly long paths and paths with embedded NULs.
499 * But we might get such a socket from the outside. Let's return
500 * something meaningful and printable in this case. */
502 _cleanup_free_
char *e
= NULL
;
504 e
= cescape_length(path
+ 1, path_len
- 1);
510 if (path
[path_len
- 1] == '\0')
511 /* We expect a terminating NUL and don't print it */
514 p
= cescape_length(path
, path_len
);
524 if (sa
->vm
.svm_cid
== VMADDR_CID_ANY
)
525 r
= asprintf(&p
, "vsock::%u", sa
->vm
.svm_port
);
527 r
= asprintf(&p
, "vsock:%u:%u", sa
->vm
.svm_cid
, sa
->vm
.svm_port
);
529 r
= asprintf(&p
, "vsock:%u", sa
->vm
.svm_cid
);
542 int getpeername_pretty(int fd
, bool include_port
, char **ret
) {
543 union sockaddr_union sa
;
544 socklen_t salen
= sizeof(sa
);
550 if (getpeername(fd
, &sa
.sa
, &salen
) < 0)
553 if (sa
.sa
.sa_family
== AF_UNIX
) {
554 struct ucred ucred
= {};
556 /* UNIX connection sockets are anonymous, so let's use
557 * PID/UID as pretty credentials instead */
559 r
= getpeercred(fd
, &ucred
);
563 if (asprintf(ret
, "PID "PID_FMT
"/UID "UID_FMT
, ucred
.pid
, ucred
.uid
) < 0)
569 /* For remote sockets we translate IPv6 addresses back to IPv4
570 * if applicable, since that's nicer. */
572 return sockaddr_pretty(&sa
.sa
, salen
, true, include_port
, ret
);
575 int getsockname_pretty(int fd
, char **ret
) {
576 union sockaddr_union sa
;
577 socklen_t salen
= sizeof(sa
);
582 if (getsockname(fd
, &sa
.sa
, &salen
) < 0)
585 /* For local sockets we do not translate IPv6 addresses back
586 * to IPv6 if applicable, since this is usually used for
587 * listening sockets where the difference between IPv4 and
590 return sockaddr_pretty(&sa
.sa
, salen
, false, true, ret
);
593 int socknameinfo_pretty(union sockaddr_union
*sa
, socklen_t salen
, char **_ret
) {
595 char host
[NI_MAXHOST
], *ret
;
599 r
= getnameinfo(&sa
->sa
, salen
, host
, sizeof(host
), NULL
, 0, IDN_FLAGS
);
601 int saved_errno
= errno
;
603 r
= sockaddr_pretty(&sa
->sa
, salen
, true, true, &ret
);
607 log_debug_errno(saved_errno
, "getnameinfo(%s) failed: %m", ret
);
618 static const char* const netlink_family_table
[] = {
619 [NETLINK_ROUTE
] = "route",
620 [NETLINK_FIREWALL
] = "firewall",
621 [NETLINK_INET_DIAG
] = "inet-diag",
622 [NETLINK_NFLOG
] = "nflog",
623 [NETLINK_XFRM
] = "xfrm",
624 [NETLINK_SELINUX
] = "selinux",
625 [NETLINK_ISCSI
] = "iscsi",
626 [NETLINK_AUDIT
] = "audit",
627 [NETLINK_FIB_LOOKUP
] = "fib-lookup",
628 [NETLINK_CONNECTOR
] = "connector",
629 [NETLINK_NETFILTER
] = "netfilter",
630 [NETLINK_IP6_FW
] = "ip6-fw",
631 [NETLINK_DNRTMSG
] = "dnrtmsg",
632 [NETLINK_KOBJECT_UEVENT
] = "kobject-uevent",
633 [NETLINK_GENERIC
] = "generic",
634 [NETLINK_SCSITRANSPORT
] = "scsitransport",
635 [NETLINK_ECRYPTFS
] = "ecryptfs",
636 [NETLINK_RDMA
] = "rdma",
639 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(netlink_family
, int, INT_MAX
);
641 static const char* const socket_address_bind_ipv6_only_table
[_SOCKET_ADDRESS_BIND_IPV6_ONLY_MAX
] = {
642 [SOCKET_ADDRESS_DEFAULT
] = "default",
643 [SOCKET_ADDRESS_BOTH
] = "both",
644 [SOCKET_ADDRESS_IPV6_ONLY
] = "ipv6-only"
647 DEFINE_STRING_TABLE_LOOKUP(socket_address_bind_ipv6_only
, SocketAddressBindIPv6Only
);
649 SocketAddressBindIPv6Only
socket_address_bind_ipv6_only_or_bool_from_string(const char *n
) {
652 r
= parse_boolean(n
);
654 return SOCKET_ADDRESS_IPV6_ONLY
;
656 return SOCKET_ADDRESS_BOTH
;
658 return socket_address_bind_ipv6_only_from_string(n
);
661 bool sockaddr_equal(const union sockaddr_union
*a
, const union sockaddr_union
*b
) {
665 if (a
->sa
.sa_family
!= b
->sa
.sa_family
)
668 if (a
->sa
.sa_family
== AF_INET
)
669 return a
->in
.sin_addr
.s_addr
== b
->in
.sin_addr
.s_addr
;
671 if (a
->sa
.sa_family
== AF_INET6
)
672 return memcmp(&a
->in6
.sin6_addr
, &b
->in6
.sin6_addr
, sizeof(a
->in6
.sin6_addr
)) == 0;
674 if (a
->sa
.sa_family
== AF_VSOCK
)
675 return a
->vm
.svm_cid
== b
->vm
.svm_cid
;
680 int fd_set_sndbuf(int fd
, size_t n
, bool increase
) {
682 socklen_t l
= sizeof(value
);
687 r
= getsockopt(fd
, SOL_SOCKET
, SO_SNDBUF
, &value
, &l
);
688 if (r
>= 0 && l
== sizeof(value
) && increase
? (size_t) value
>= n
*2 : (size_t) value
== n
*2)
691 /* First, try to set the buffer size with SO_SNDBUF. */
692 r
= setsockopt_int(fd
, SOL_SOCKET
, SO_SNDBUF
, n
);
696 /* SO_SNDBUF above may set to the kernel limit, instead of the requested size.
697 * So, we need to check the actual buffer size here. */
699 r
= getsockopt(fd
, SOL_SOCKET
, SO_SNDBUF
, &value
, &l
);
700 if (r
>= 0 && l
== sizeof(value
) && increase
? (size_t) value
>= n
*2 : (size_t) value
== n
*2)
703 /* If we have the privileges we will ignore the kernel limit. */
704 r
= setsockopt_int(fd
, SOL_SOCKET
, SO_SNDBUFFORCE
, n
);
711 int fd_set_rcvbuf(int fd
, size_t n
, bool increase
) {
713 socklen_t l
= sizeof(value
);
718 r
= getsockopt(fd
, SOL_SOCKET
, SO_RCVBUF
, &value
, &l
);
719 if (r
>= 0 && l
== sizeof(value
) && increase
? (size_t) value
>= n
*2 : (size_t) value
== n
*2)
722 /* First, try to set the buffer size with SO_RCVBUF. */
723 r
= setsockopt_int(fd
, SOL_SOCKET
, SO_RCVBUF
, n
);
727 /* SO_RCVBUF above may set to the kernel limit, instead of the requested size.
728 * So, we need to check the actual buffer size here. */
730 r
= getsockopt(fd
, SOL_SOCKET
, SO_RCVBUF
, &value
, &l
);
731 if (r
>= 0 && l
== sizeof(value
) && increase
? (size_t) value
>= n
*2 : (size_t) value
== n
*2)
734 /* If we have the privileges we will ignore the kernel limit. */
735 r
= setsockopt_int(fd
, SOL_SOCKET
, SO_RCVBUFFORCE
, n
);
742 static const char* const ip_tos_table
[] = {
743 [IPTOS_LOWDELAY
] = "low-delay",
744 [IPTOS_THROUGHPUT
] = "throughput",
745 [IPTOS_RELIABILITY
] = "reliability",
746 [IPTOS_LOWCOST
] = "low-cost",
749 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(ip_tos
, int, 0xff);
751 bool ifname_valid_full(const char *p
, IfnameValidFlags flags
) {
754 /* Checks whether a network interface name is valid. This is inspired by dev_valid_name() in the kernel sources
755 * but slightly stricter, as we only allow non-control, non-space ASCII characters in the interface name. We
756 * also don't permit names that only container numbers, to avoid confusion with numeric interface indexes. */
758 assert(!(flags
& ~_IFNAME_VALID_ALL
));
763 /* A valid ifindex? If so, it's valid iff IFNAME_VALID_NUMERIC is set */
764 if (parse_ifindex(p
) >= 0)
765 return flags
& IFNAME_VALID_NUMERIC
;
767 if (flags
& IFNAME_VALID_ALTERNATIVE
) {
768 if (strlen(p
) >= ALTIFNAMSIZ
)
771 if (strlen(p
) >= IFNAMSIZ
)
775 if (dot_or_dot_dot(p
))
778 /* Let's refuse "all" and "default" as interface name, to avoid collisions with the special sysctl
779 * directories /proc/sys/net/{ipv4,ipv6}/conf/{all,default} */
780 if (STR_IN_SET(p
, "all", "default"))
783 for (const char *t
= p
; *t
; t
++) {
784 if ((unsigned char) *t
>= 127U)
787 if ((unsigned char) *t
<= 32U)
791 ':', /* colons are used by the legacy "alias" interface logic */
792 '/', /* slashes cannot work, since we need to use network interfaces in sysfs paths, and in paths slashes are separators */
793 '%')) /* %d is used in the kernel's weird foo%d format string naming feature which we really really don't want to ever run into by accident */
796 numeric
= numeric
&& (*t
>= '0' && *t
<= '9');
799 /* It's fully numeric but didn't parse as valid ifindex above? if so, it must be too large or zero or
800 * so, let's refuse that. */
807 bool address_label_valid(const char *p
) {
812 if (strlen(p
) >= IFNAMSIZ
)
816 if ((uint8_t) *p
>= 127U)
819 if ((uint8_t) *p
<= 31U)
827 int getpeercred(int fd
, struct ucred
*ucred
) {
828 socklen_t n
= sizeof(struct ucred
);
835 r
= getsockopt(fd
, SOL_SOCKET
, SO_PEERCRED
, &u
, &n
);
839 if (n
!= sizeof(struct ucred
))
842 /* Check if the data is actually useful and not suppressed due to namespacing issues */
843 if (!pid_is_valid(u
.pid
))
846 /* Note that we don't check UID/GID here, as namespace translation works differently there: instead of
847 * receiving in "invalid" user/group we get the overflow UID/GID. */
853 int getpeersec(int fd
, char **ret
) {
854 _cleanup_free_
char *s
= NULL
;
865 if (getsockopt(fd
, SOL_SOCKET
, SO_PEERSEC
, s
, &n
) >= 0)
882 int getpeergroups(int fd
, gid_t
**ret
) {
883 socklen_t n
= sizeof(gid_t
) * 64;
884 _cleanup_free_ gid_t
*d
= NULL
;
894 if (getsockopt(fd
, SOL_SOCKET
, SO_PEERGROUPS
, d
, &n
) >= 0)
903 assert_se(n
% sizeof(gid_t
) == 0);
906 if ((socklen_t
) (int) n
!= n
)
914 ssize_t
send_one_fd_iov_sa(
917 struct iovec
*iov
, size_t iovlen
,
918 const struct sockaddr
*sa
, socklen_t len
,
921 CMSG_BUFFER_TYPE(CMSG_SPACE(sizeof(int))) control
= {};
923 .msg_name
= (struct sockaddr
*) sa
,
926 .msg_iovlen
= iovlen
,
930 assert(transport_fd
>= 0);
933 * We need either an FD or data to send.
934 * If there's nothing, return an error.
940 struct cmsghdr
*cmsg
;
942 mh
.msg_control
= &control
;
943 mh
.msg_controllen
= sizeof(control
);
945 cmsg
= CMSG_FIRSTHDR(&mh
);
946 cmsg
->cmsg_level
= SOL_SOCKET
;
947 cmsg
->cmsg_type
= SCM_RIGHTS
;
948 cmsg
->cmsg_len
= CMSG_LEN(sizeof(int));
949 memcpy(CMSG_DATA(cmsg
), &fd
, sizeof(int));
951 k
= sendmsg(transport_fd
, &mh
, MSG_NOSIGNAL
| flags
);
953 return (ssize_t
) -errno
;
961 const struct sockaddr
*sa
, socklen_t len
,
966 return (int) send_one_fd_iov_sa(transport_fd
, fd
, NULL
, 0, sa
, len
, flags
);
969 ssize_t
receive_one_fd_iov(
971 struct iovec
*iov
, size_t iovlen
,
975 CMSG_BUFFER_TYPE(CMSG_SPACE(sizeof(int))) control
;
977 .msg_control
= &control
,
978 .msg_controllen
= sizeof(control
),
980 .msg_iovlen
= iovlen
,
982 struct cmsghdr
*found
;
985 assert(transport_fd
>= 0);
989 * Receive a single FD via @transport_fd. We don't care for
990 * the transport-type. We retrieve a single FD at most, so for
991 * packet-based transports, the caller must ensure to send
992 * only a single FD per packet. This is best used in
993 * combination with send_one_fd().
996 k
= recvmsg_safe(transport_fd
, &mh
, MSG_CMSG_CLOEXEC
| flags
);
1000 found
= cmsg_find(&mh
, SOL_SOCKET
, SCM_RIGHTS
, CMSG_LEN(sizeof(int)));
1002 cmsg_close_all(&mh
);
1004 /* If didn't receive an FD or any data, return an error. */
1010 *ret_fd
= *(int*) CMSG_DATA(found
);
1017 int receive_one_fd(int transport_fd
, int flags
) {
1021 k
= receive_one_fd_iov(transport_fd
, NULL
, 0, flags
, &fd
);
1025 /* k must be negative, since receive_one_fd_iov() only returns
1026 * a positive value if data was received through the iov. */
1031 ssize_t
next_datagram_size_fd(int fd
) {
1035 /* This is a bit like FIONREAD/SIOCINQ, however a bit more powerful. The difference being: recv(MSG_PEEK) will
1036 * actually cause the next datagram in the queue to be validated regarding checksums, which FIONREAD doesn't
1037 * do. This difference is actually of major importance as we need to be sure that the size returned here
1038 * actually matches what we will read with recvmsg() next, as otherwise we might end up allocating a buffer of
1039 * the wrong size. */
1041 l
= recv(fd
, NULL
, 0, MSG_PEEK
|MSG_TRUNC
);
1043 if (IN_SET(errno
, EOPNOTSUPP
, EFAULT
))
1056 /* Some sockets (AF_PACKET) do not support null-sized recv() with MSG_TRUNC set, let's fall back to FIONREAD
1057 * for them. Checksums don't matter for raw sockets anyway, hence this should be fine. */
1059 if (ioctl(fd
, FIONREAD
, &k
) < 0)
1065 /* Put a limit on how many times will attempt to call accept4(). We loop
1066 * only on "transient" errors, but let's make sure we don't loop forever. */
1067 #define MAX_FLUSH_ITERATIONS 1024
1069 int flush_accept(int fd
) {
1072 socklen_t l
= sizeof(b
);
1074 /* Similar to flush_fd() but flushes all incoming connections by accepting and immediately closing
1077 if (getsockopt(fd
, SOL_SOCKET
, SO_ACCEPTCONN
, &b
, &l
) < 0)
1080 assert(l
== sizeof(b
));
1081 if (!b
) /* Let's check if this socket accepts connections before calling accept(). accept4() can
1082 * return EOPNOTSUPP if the fd is not a listening socket, which we should treat as a fatal
1083 * error, or in case the incoming TCP connection triggered a network issue, which we want to
1084 * treat as a transient error. Thus, let's rule out the first reason for EOPNOTSUPP early, so
1085 * we can loop safely on transient errors below. */
1088 for (unsigned iteration
= 0;; iteration
++) {
1091 r
= fd_wait_for_event(fd
, POLLIN
, 0);
1101 if (iteration
>= MAX_FLUSH_ITERATIONS
)
1102 return log_debug_errno(SYNTHETIC_ERRNO(EBUSY
),
1103 "Failed to flush connections within " STRINGIFY(MAX_FLUSH_ITERATIONS
) " iterations.");
1105 cfd
= accept4(fd
, NULL
, NULL
, SOCK_NONBLOCK
|SOCK_CLOEXEC
);
1107 if (errno
== EAGAIN
)
1110 if (ERRNO_IS_ACCEPT_AGAIN(errno
))
1120 struct cmsghdr
* cmsg_find(struct msghdr
*mh
, int level
, int type
, socklen_t length
) {
1121 struct cmsghdr
*cmsg
;
1125 CMSG_FOREACH(cmsg
, mh
)
1126 if (cmsg
->cmsg_level
== level
&&
1127 cmsg
->cmsg_type
== type
&&
1128 (length
== (socklen_t
) -1 || length
== cmsg
->cmsg_len
))
1134 int socket_ioctl_fd(void) {
1137 /* Create a socket to invoke the various network interface ioctl()s on. Traditionally only AF_INET was good for
1138 * that. Since kernel 4.6 AF_NETLINK works for this too. We first try to use AF_INET hence, but if that's not
1139 * available (for example, because it is made unavailable via SECCOMP or such), we'll fall back to the more
1140 * generic AF_NETLINK. */
1142 fd
= socket(AF_INET
, SOCK_DGRAM
|SOCK_CLOEXEC
, 0);
1144 fd
= socket(AF_NETLINK
, SOCK_RAW
|SOCK_CLOEXEC
, NETLINK_GENERIC
);
1151 int sockaddr_un_unlink(const struct sockaddr_un
*sa
) {
1152 const char *p
, * nul
;
1156 if (sa
->sun_family
!= AF_UNIX
)
1159 if (sa
->sun_path
[0] == 0) /* Nothing to do for abstract sockets */
1162 /* The path in .sun_path is not necessarily NUL terminated. Let's fix that. */
1163 nul
= memchr(sa
->sun_path
, 0, sizeof(sa
->sun_path
));
1167 p
= memdupa_suffix0(sa
->sun_path
, sizeof(sa
->sun_path
));
1175 int sockaddr_un_set_path(struct sockaddr_un
*ret
, const char *path
) {
1181 /* Initialize ret->sun_path from the specified argument. This will interpret paths starting with '@' as
1182 * abstract namespace sockets, and those starting with '/' as regular filesystem sockets. It won't accept
1183 * anything else (i.e. no relative paths), to avoid ambiguities. Note that this function cannot be used to
1184 * reference paths in the abstract namespace that include NUL bytes in the name. */
1189 if (!IN_SET(path
[0], '/', '@'))
1192 /* Don't allow paths larger than the space in sockaddr_un. Note that we are a tiny bit more restrictive than
1193 * the kernel is: we insist on NUL termination (both for abstract namespace and regular file system socket
1194 * addresses!), which the kernel doesn't. We do this to reduce chance of incompatibility with other apps that
1195 * do not expect non-NUL terminated file system path*/
1196 if (l
+1 > sizeof(ret
->sun_path
))
1199 *ret
= (struct sockaddr_un
) {
1200 .sun_family
= AF_UNIX
,
1203 if (path
[0] == '@') {
1204 /* Abstract namespace socket */
1205 memcpy(ret
->sun_path
+ 1, path
+ 1, l
); /* copy *with* trailing NUL byte */
1206 return (int) (offsetof(struct sockaddr_un
, sun_path
) + l
); /* 🔥 *don't* 🔥 include trailing NUL in size */
1209 assert(path
[0] == '/');
1211 /* File system socket */
1212 memcpy(ret
->sun_path
, path
, l
+ 1); /* copy *with* trailing NUL byte */
1213 return (int) (offsetof(struct sockaddr_un
, sun_path
) + l
+ 1); /* include trailing NUL in size */
1217 int socket_bind_to_ifname(int fd
, const char *ifname
) {
1220 /* Call with NULL to drop binding */
1222 if (setsockopt(fd
, SOL_SOCKET
, SO_BINDTODEVICE
, ifname
, strlen_ptr(ifname
)) < 0)
1228 int socket_bind_to_ifindex(int fd
, int ifindex
) {
1229 char ifname
[IF_NAMESIZE
+ 1];
1236 if (setsockopt(fd
, SOL_SOCKET
, SO_BINDTODEVICE
, NULL
, 0) < 0)
1242 r
= setsockopt_int(fd
, SOL_SOCKET
, SO_BINDTOIFINDEX
, ifindex
);
1243 if (r
!= -ENOPROTOOPT
)
1246 /* Fall back to SO_BINDTODEVICE on kernels < 5.0 which didn't have SO_BINDTOIFINDEX */
1247 if (!format_ifname(ifindex
, ifname
))
1250 return socket_bind_to_ifname(fd
, ifname
);
1253 ssize_t
recvmsg_safe(int sockfd
, struct msghdr
*msg
, int flags
) {
1256 /* A wrapper around recvmsg() that checks for MSG_CTRUNC, and turns it into an error, in a reasonably
1257 * safe way, closing any SCM_RIGHTS fds in the error path.
1259 * Note that unlike our usual coding style this might modify *msg on failure. */
1261 n
= recvmsg(sockfd
, msg
, flags
);
1265 if (FLAGS_SET(msg
->msg_flags
, MSG_CTRUNC
)) {
1266 cmsg_close_all(msg
);
1267 return -EXFULL
; /* a recognizable error code */
1273 int socket_get_family(int fd
, int *ret
) {
1275 socklen_t sl
= sizeof(af
);
1277 if (getsockopt(fd
, SOL_SOCKET
, SO_DOMAIN
, &af
, &sl
) < 0)
1280 if (sl
!= sizeof(af
))
1286 int socket_set_recvpktinfo(int fd
, int af
, bool b
) {
1289 if (af
== AF_UNSPEC
) {
1290 r
= socket_get_family(fd
, &af
);
1298 return setsockopt_int(fd
, IPPROTO_IP
, IP_PKTINFO
, b
);
1301 return setsockopt_int(fd
, IPPROTO_IPV6
, IPV6_RECVPKTINFO
, b
);
1304 return setsockopt_int(fd
, SOL_NETLINK
, NETLINK_PKTINFO
, b
);
1307 return setsockopt_int(fd
, SOL_PACKET
, PACKET_AUXDATA
, b
);
1310 return -EAFNOSUPPORT
;
1314 int socket_set_unicast_if(int fd
, int af
, int ifi
) {
1315 be32_t ifindex_be
= htobe32(ifi
);
1318 if (af
== AF_UNSPEC
) {
1319 r
= socket_get_family(fd
, &af
);
1327 if (setsockopt(fd
, IPPROTO_IP
, IP_UNICAST_IF
, &ifindex_be
, sizeof(ifindex_be
)) < 0)
1333 if (setsockopt(fd
, IPPROTO_IPV6
, IPV6_UNICAST_IF
, &ifindex_be
, sizeof(ifindex_be
)) < 0)
1339 return -EAFNOSUPPORT
;
1343 int socket_set_option(int fd
, int af
, int opt_ipv4
, int opt_ipv6
, int val
) {
1346 if (af
== AF_UNSPEC
) {
1347 r
= socket_get_family(fd
, &af
);
1355 return setsockopt_int(fd
, IPPROTO_IP
, opt_ipv4
, val
);
1358 return setsockopt_int(fd
, IPPROTO_IPV6
, opt_ipv6
, val
);
1361 return -EAFNOSUPPORT
;
1365 int socket_get_mtu(int fd
, int af
, size_t *ret
) {
1368 if (af
== AF_UNSPEC
) {
1369 r
= socket_get_family(fd
, &af
);
1377 r
= getsockopt_int(fd
, IPPROTO_IP
, IP_MTU
, &mtu
);
1381 r
= getsockopt_int(fd
, IPPROTO_IPV6
, IPV6_MTU
, &mtu
);
1385 return -EAFNOSUPPORT
;
1393 *ret
= (size_t) mtu
;