Merge pull request #19134 from poettering/outbound-special-hostname

introduce a new synthetic hostname "_outbound" that maps to "the" local IP address

src/shared/local-addresses.c

@@ -1,8 +1,11 @@
 /* SPDX-License-Identifier: LGPL-2.1-or-later */
 
+#include <net/if_arp.h>
+
 #include "sd-netlink.h"
 
 #include "alloc-util.h"
+#include "fd-util.h"
 #include "local-addresses.h"
 #include "macro.h"
 #include "netlink-util.h"
@@ -33,7 +36,34 @@ static int address_compare(const struct local_address *a, const struct local_add
         return memcmp(&a->address, &b->address, FAMILY_ADDRESS_SIZE(a->family));
 }
 
-int local_addresses(sd_netlink *context, int ifindex, int af, struct local_address **ret) {
+static void suppress_duplicates(struct local_address *list, size_t *n_list) {
+        size_t old_size, new_size;
+
+        /* Removes duplicate entries, assumes the list of addresses is already sorted. Updates in-place. */
+
+        if (*n_list < 2) /* list with less than two entries can't have duplicates */
+                return;
+
+        old_size = *n_list;
+        new_size = 1;
+
+        for (size_t i = 1; i < old_size; i++) {
+
+                if (address_compare(list + i, list + new_size - 1) == 0)
+                        continue;
+
+                list[new_size++] = list[i];
+        }
+
+        *n_list = new_size;
+}
+
+int local_addresses(
+                sd_netlink *context,
+                int ifindex,
+                int af,
+                struct local_address **ret) {
+
         _cleanup_(sd_netlink_message_unrefp) sd_netlink_message *req = NULL, *reply = NULL;
         _cleanup_(sd_netlink_unrefp) sd_netlink *rtnl = NULL;
         _cleanup_free_ struct local_address *list = NULL;
@@ -135,6 +165,7 @@ int local_addresses(sd_netlink *context, int ifindex, int af, struct local_addre
 
         if (ret) {
                 typesafe_qsort(list, n_list, address_compare);
+                suppress_duplicates(list, &n_list);
                 *ret = TAKE_PTR(list);
         }
 
@@ -171,7 +202,12 @@ static int add_local_gateway(
         return 0;
 }
 
-int local_gateways(sd_netlink *context, int ifindex, int af, struct local_address **ret) {
+int local_gateways(
+                sd_netlink *context,
+                int ifindex,
+                int af,
+                struct local_address **ret) {
+
         _cleanup_(sd_netlink_message_unrefp) sd_netlink_message *req = NULL, *reply = NULL;
         _cleanup_(sd_netlink_unrefp) sd_netlink *rtnl = NULL;
         _cleanup_free_ struct local_address *list = NULL;
@@ -308,6 +344,151 @@ int local_gateways(sd_netlink *context, int ifindex, int af, struct local_addres
 
         if (ret) {
                 typesafe_qsort(list, n_list, address_compare);
+                suppress_duplicates(list, &n_list);
+                *ret = TAKE_PTR(list);
+        }
+
+        return (int) n_list;
+}
+
+int local_outbounds(
+                sd_netlink *context,
+                int ifindex,
+                int af,
+                struct local_address **ret) {
+
+        _cleanup_free_ struct local_address *list = NULL, *gateways = NULL;
+        size_t n_list = 0, n_allocated = 0;
+        int r, n_gateways;
+
+        /* Determines our default outbound addresses, i.e. the "primary" local addresses we use to talk to IP
+         * addresses behind the default routes. This is still an address of the local host (i.e. this doesn't
+         * resolve NAT or so), but it's the set of addresses the local IP stack most likely uses to talk to
+         * other hosts.
+         *
+         * This works by connect()ing a SOCK_DGRAM socket to the local gateways, and then reading the IP
+         * address off the socket that was chosen for the routing decision. */
+
+        n_gateways = local_gateways(context, ifindex, af, &gateways);
+        if (n_gateways < 0)
+                return n_gateways;
+        if (n_gateways == 0) {
+                /* No gateways? Then we have no outbound addresses either. */
+                if (ret)
+                        *ret = NULL;
+
+                return 0;
+        }
+
+        for (int i = 0; i < n_gateways; i++) {
+                _cleanup_close_ int fd = -1;
+                union sockaddr_union sa;
+                socklen_t salen;
+
+                fd = socket(gateways[i].family, SOCK_DGRAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0);
+                if (fd < 0)
+                        return -errno;
+
+                switch (gateways[i].family) {
+
+                case AF_INET:
+                        sa.in = (struct sockaddr_in) {
+                                .sin_family = AF_INET,
+                                .sin_addr = gateways[i].address.in,
+                                .sin_port = htobe16(53), /* doesn't really matter which port we pick — we just care about the routing decision */
+                        };
+
+                        break;
+
+                case AF_INET6:
+                        sa.in6 = (struct sockaddr_in6) {
+                                .sin6_family = AF_INET6,
+                                .sin6_addr = gateways[i].address.in6,
+                                .sin6_port = htobe16(53),
+                                .sin6_scope_id = gateways[i].ifindex,
+                        };
+
+                        break;
+
+                default:
+                        assert_not_reached("Unexpected protocol");
+                }
+
+                /* So ideally we'd just use IP_UNICAST_IF here to pass the ifindex info to the kernel before
+                 * connect()ing, sot that it influences the routing decision. However, on current kernels
+                 * IP_UNICAST_IF doesn't actually influence the routing decision for UDP — which I think
+                 * should probably just be considered a bug. Once that bug is fixed this is the best API to
+                 * use, since it is the most lightweight. */
+                r = socket_set_unicast_if(fd, gateways[i].family, gateways[i].ifindex);
+                if (r < 0)
+                        log_debug_errno(r, "Failed to set unicast interface index %i, ignoring: %m", gateways[i].ifindex);
+
+                /* We'll also use SO_BINDTOINDEX. This requires CAP_NET_RAW on old kernels, hence there's a
+                 * good chance this fails. Since 5.7 this restriction was dropped and the first
+                 * SO_BINDTOINDEX on a socket may be done without privileges. This one has the benefit of
+                 * really influencing the routing decision, i.e. this one definitely works for us — as long
+                 * as we have the privileges for it.*/
+                r = socket_bind_to_ifindex(fd, gateways[i].ifindex);
+                if (r < 0)
+                        log_debug_errno(r, "Failed to bind socket to interface %i, ignoring: %m", gateways[i].ifindex);
+
+                /* Let's now connect() to the UDP socket, forcing the kernel to make a routing decision and
+                 * auto-bind the socket. We ignore failures on this, since that failure might happen for a
+                 * multitude of reasons (policy/firewall issues, who knows?) and some of them might be
+                 * *after* the routing decision and the auto-binding already took place. If so we can still
+                 * make use of the binding and return it. Hence, let's not unnecessarily fail early here: we
+                 * can still easily detect if the auto-binding worked or not, by comparing the bound IP
+                 * address with zero — which we do below.  */
+                if (connect(fd, &sa.sa, SOCKADDR_LEN(sa)) < 0)
+                        log_debug_errno(errno, "Failed to connect SOCK_DGRAM socket to gateway, ignoring: %m");
+
+                /* Let's now read the socket address of the socket. A routing decision should have been
+                 * made. Let's verify that and use the data. */
+                salen = SOCKADDR_LEN(sa);
+                if (getsockname(fd, &sa.sa, &salen) < 0)
+                        return -errno;
+                assert(sa.sa.sa_family == gateways[i].family);
+                assert(salen == SOCKADDR_LEN(sa));
+
+                switch (gateways[i].family) {
+
+                case AF_INET:
+                        if (in4_addr_is_null(&sa.in.sin_addr)) /* Auto-binding didn't work. :-( */
+                                continue;
+
+                        if (!GREEDY_REALLOC(list, n_allocated, n_list+1))
+                                return -ENOMEM;
+
+                        list[n_list++] = (struct local_address) {
+                                .family = gateways[i].family,
+                                .ifindex = gateways[i].ifindex,
+                                .address.in = sa.in.sin_addr,
+                        };
+
+                        break;
+
+                case AF_INET6:
+                        if (in6_addr_is_null(&sa.in6.sin6_addr))
+                                continue;
+
+                        if (!GREEDY_REALLOC(list, n_allocated, n_list+1))
+                                return -ENOMEM;
+
+                        list[n_list++] = (struct local_address) {
+                                .family = gateways[i].family,
+                                .ifindex = gateways[i].ifindex,
+                                .address.in6 = sa.in6.sin6_addr,
+                        };
+                        break;
+
+                default:
+                        assert_not_reached("Unexpected protocol");
+                }
+        }
+
+        if (ret) {
+                typesafe_qsort(list, n_list, address_compare);
+                suppress_duplicates(list, &n_list);
                 *ret = TAKE_PTR(list);
         }
 

src/shared/local-addresses.h

@@ -15,3 +15,5 @@ struct local_address {
 int local_addresses(sd_netlink *rtnl, int ifindex, int af, struct local_address **ret);
 
 int local_gateways(sd_netlink *rtnl, int ifindex, int af, struct local_address **ret);
+
+int local_outbounds(sd_netlink *rtnl, int ifindex, int af, struct local_address **ret);