resolved: support for TCP DNS queries

7 files changed, 454 insertions, 219 deletions

diff --git a/src/resolve/resolved-bus.c b/src/resolve/resolved-bus.c
index 02e971dbf3..64bb9ac889 100644
--- a/src/resolve/resolved-bus.c
+++ b/src/resolve/resolved-bus.c
@@ -25,141 +25,161 @@
 #include "resolved.h"
 #include "resolved-dns-domain.h"
 
-static void bus_method_resolve_hostname_complete(DnsQuery *q) {
+static int reply_query_state(DnsQuery *q) {
+        _cleanup_free_ char *ip = NULL;
+        const char *name;
         int r;
 
-        assert(q);
+        if (q->request_hostname)
+                name = q->request_hostname;
+        else {
+                r = in_addr_to_string(q->request_family, &q->request_address, &ip);
+                if (r < 0)
+                        return r;
+
+                name = ip;
+        }
 
-        switch(q->state) {
+        switch (q->state) {
 
-        case DNS_QUERY_SKIPPED:
-                r = sd_bus_reply_method_errorf(q->request, BUS_ERROR_NO_NAME_SERVERS, "Not appropriate name servers or networks found");
-                break;
+        case DNS_QUERY_NO_SERVERS:
+                return sd_bus_reply_method_errorf(q->request, BUS_ERROR_NO_NAME_SERVERS, "Not appropriate name servers or networks found");
 
         case DNS_QUERY_TIMEOUT:
-                r = sd_bus_reply_method_errorf(q->request, SD_BUS_ERROR_TIMEOUT, "Query timed out");
-                break;
+                return sd_bus_reply_method_errorf(q->request, SD_BUS_ERROR_TIMEOUT, "Query timed out");
 
         case DNS_QUERY_ATTEMPTS_MAX:
-                r = sd_bus_reply_method_errorf(q->request, SD_BUS_ERROR_TIMEOUT, "All attempts to contact name servers or networks failed");
-                break;
+                return sd_bus_reply_method_errorf(q->request, SD_BUS_ERROR_TIMEOUT, "All attempts to contact name servers or networks failed");
+
+        case DNS_QUERY_RESOURCES:
+                return sd_bus_reply_method_errorf(q->request, BUS_ERROR_NO_RESOURCES, "Not enough resources");
+
+        case DNS_QUERY_INVALID_REPLY:
+                return sd_bus_reply_method_errorf(q->request, BUS_ERROR_INVALID_REPLY, "Received invalid reply");
 
         case DNS_QUERY_FAILURE: {
                 _cleanup_bus_error_free_ sd_bus_error error = SD_BUS_ERROR_NULL;
 
-                if (q->rcode == DNS_RCODE_NXDOMAIN)
-                        sd_bus_error_setf(&error, _BUS_ERROR_DNS "NXDOMAIN", "Hostname %s does not exist", q->request_hostname);
+                assert(q->received);
+
+                if (DNS_PACKET_RCODE(q->received) == DNS_RCODE_NXDOMAIN)
+                        sd_bus_error_setf(&error, _BUS_ERROR_DNS "NXDOMAIN", "'%s' not found", name);
                 else {
                         const char *rc, *n;
-                        char p[DECIMAL_STR_MAX(q->rcode)];
+                        char p[3]; /* the rcode is 4 bits long */
 
-                        rc = dns_rcode_to_string(q->rcode);
+                        rc = dns_rcode_to_string(DNS_PACKET_RCODE(q->received));
                         if (!rc) {
-                                sprintf(p, "%i", q->rcode);
+                                sprintf(p, "%i", DNS_PACKET_RCODE(q->received));
                                 rc = p;
                         }
 
                         n = strappenda(_BUS_ERROR_DNS, rc);
-
-                        sd_bus_error_setf(&error, n, "Could not resolve hostname %s, server or network returned error %s", q->request_hostname, rc);
+                        sd_bus_error_setf(&error, n, "Could not resolve '%s', server or network returned error %s", name, rc);
                 }
 
-                r = sd_bus_reply_method_error(q->request, &error);
-                break;
+                return sd_bus_reply_method_error(q->request, &error);
         }
 
-        case DNS_QUERY_SUCCESS: {
-                _cleanup_bus_message_unref_ sd_bus_message *reply = NULL;
-                unsigned i, n, added = 0;
+        case DNS_QUERY_NULL:
+        case DNS_QUERY_SENT:
+        case DNS_QUERY_SUCCESS:
+                assert_not_reached("Impossible state");
+        }
+}
 
-                assert(q->packet);
+static void bus_method_resolve_hostname_complete(DnsQuery *q) {
+        _cleanup_bus_message_unref_ sd_bus_message *reply = NULL;
+        unsigned i, n, added = 0;
+        int r;
 
-                r = dns_packet_skip_question(q->packet);
-                if (r < 0)
-                        goto parse_fail;
+        assert(q);
 
-                r = sd_bus_message_new_method_return(q->request, &reply);
-                if (r < 0)
-                        goto finish;
+        if (q->state != DNS_QUERY_SUCCESS) {
+                r = reply_query_state(q);
+                goto finish;
+        }
 
-                r = sd_bus_message_open_container(reply, 'a', "(yayi)");
-                if (r < 0)
-                        goto finish;
+        assert(q->received);
 
-                n = DNS_PACKET_ANCOUNT(q->packet) +
-                    DNS_PACKET_NSCOUNT(q->packet) +
-                    DNS_PACKET_ARCOUNT(q->packet);
-                for (i = 0; i < n; i++) {
-                        _cleanup_(dns_resource_record_unrefp) DnsResourceRecord *rr = NULL;
+        r = dns_packet_skip_question(q->received);
+        if (r < 0)
+                goto parse_fail;
 
-                        r = dns_packet_read_rr(q->packet, &rr, NULL);
-                        if (r < 0)
-                                goto parse_fail;
+        r = sd_bus_message_new_method_return(q->request, &reply);
+        if (r < 0)
+                goto finish;
 
-                        if (rr->key.class != DNS_CLASS_IN)
-                                continue;
+        r = sd_bus_message_open_container(reply, 'a', "(yayi)");
+        if (r < 0)
+                goto finish;
 
-                        if (!(q->request_family != AF_INET6 && rr->key.type == DNS_TYPE_A) &&
-                            !(q->request_family != AF_INET && rr->key.type == DNS_TYPE_AAAA))
-                                continue;
+        n = DNS_PACKET_ANCOUNT(q->received) +
+            DNS_PACKET_NSCOUNT(q->received) +
+            DNS_PACKET_ARCOUNT(q->received);
+        for (i = 0; i < n; i++) {
+                _cleanup_(dns_resource_record_unrefp) DnsResourceRecord *rr = NULL;
 
-                        if (!dns_name_equal(rr->key.name, q->request_hostname))
-                                continue;
+                r = dns_packet_read_rr(q->received, &rr, NULL);
+                if (r < 0)
+                        goto parse_fail;
 
-                        r = sd_bus_message_open_container(reply, 'r', "yayi");
-                        if (r < 0)
-                                goto finish;
+                if (rr->key.class != DNS_CLASS_IN)
+                        continue;
 
-                        if (rr->key.type == DNS_TYPE_A) {
-                                r = sd_bus_message_append(reply, "y", AF_INET);
-                                if (r < 0)
-                                        goto finish;
+                if (!(q->request_family != AF_INET6 && rr->key.type == DNS_TYPE_A) &&
+                    !(q->request_family != AF_INET && rr->key.type == DNS_TYPE_AAAA))
+                        continue;
 
-                                r = sd_bus_message_append_array(reply, 'y', &rr->a.in_addr, sizeof(struct in_addr));
-                        } else {
-                                r = sd_bus_message_append(reply, "y", AF_INET6);
-                                if (r < 0)
-                                        goto finish;
+                if (!dns_name_equal(rr->key.name, q->request_hostname))
+                        continue;
 
-                                r = sd_bus_message_append_array(reply, 'y', &rr->aaaa.in6_addr, sizeof(struct in6_addr));
-                        }
-                        if (r < 0)
-                                goto finish;
+                r = sd_bus_message_open_container(reply, 'r', "yayi");
+                if (r < 0)
+                        goto finish;
 
-                        r = sd_bus_message_append(reply, "i", q->packet->ifindex);
+                if (rr->key.type == DNS_TYPE_A) {
+                        r = sd_bus_message_append(reply, "y", AF_INET);
                         if (r < 0)
                                 goto finish;
 
-                        r = sd_bus_message_close_container(reply);
+                        r = sd_bus_message_append_array(reply, 'y', &rr->a.in_addr, sizeof(struct in_addr));
+                } else {
+                        r = sd_bus_message_append(reply, "y", AF_INET6);
                         if (r < 0)
                                 goto finish;
 
-                        added ++;
+                        r = sd_bus_message_append_array(reply, 'y', &rr->aaaa.in6_addr, sizeof(struct in6_addr));
                 }
+                if (r < 0)
+                        goto finish;
 
-                if (added <= 0) {
-                        r = sd_bus_reply_method_errorf(q->request, BUS_ERROR_NO_SUCH_RR, "Hostname %s does not have RR of this type", q->request_hostname);
-                        break;
-                }
+                r = sd_bus_message_append(reply, "i", q->received->ifindex);
+                if (r < 0)
+                        goto finish;
 
                 r = sd_bus_message_close_container(reply);
                 if (r < 0)
                         goto finish;
 
-                r = sd_bus_send(q->manager->bus, reply, NULL);
-                break;
+                added ++;
         }
 
-        parse_fail:
-        case DNS_QUERY_INVALID_REPLY:
-                r = sd_bus_reply_method_errorf(q->request, BUS_ERROR_INVALID_REPLY, "Received invalid reply");
-                break;
-
-        case DNS_QUERY_NULL:
-        case DNS_QUERY_SENT:
-                assert_not_reached("Unexpected query state");
+        if (added <= 0) {
+                r = sd_bus_reply_method_errorf(q->request, BUS_ERROR_NO_SUCH_RR, "'%s' does not have RR of this type", q->request_hostname);
+                goto finish;
         }
 
+        r = sd_bus_message_close_container(reply);
+        if (r < 0)
+                goto finish;
+
+        r = sd_bus_send(q->manager->bus, reply, NULL);
+        goto finish;
+
+parse_fail:
+        r = sd_bus_reply_method_errorf(q->request, BUS_ERROR_INVALID_REPLY, "Received invalid reply");
+
 finish:
         if (r < 0)
                 log_error("Failed to send bus reply: %s", strerror(-r));
@@ -223,120 +243,79 @@ static int bus_method_resolve_hostname(sd_bus *bus, sd_bus_message *message, voi
 }
 
 static void bus_method_resolve_address_complete(DnsQuery *q) {
-        _cleanup_free_ char *ip = NULL;
+        _cleanup_bus_message_unref_ sd_bus_message *reply = NULL;
+        unsigned i, n, added = 0;
+        _cleanup_free_ char *reverse = NULL;
         int r;
 
         assert(q);
 
-        in_addr_to_string(q->request_family, &q->request_address, &ip);
-
-        switch(q->state) {
-
-        case DNS_QUERY_SKIPPED:
-                r = sd_bus_reply_method_errorf(q->request, BUS_ERROR_NO_NAME_SERVERS, "Not appropriate name servers or networks found");
-                break;
-
-        case DNS_QUERY_TIMEOUT:
-                r = sd_bus_reply_method_errorf(q->request, SD_BUS_ERROR_TIMEOUT, "Query timed out");
-                break;
-
-        case DNS_QUERY_ATTEMPTS_MAX:
-                r = sd_bus_reply_method_errorf(q->request, SD_BUS_ERROR_TIMEOUT, "All attempts to contact name servers or networks failed");
-                break;
-
-        case DNS_QUERY_FAILURE: {
-                _cleanup_bus_error_free_ sd_bus_error error = SD_BUS_ERROR_NULL;
-
-                if (q->rcode == DNS_RCODE_NXDOMAIN)
-                        sd_bus_error_setf(&error, _BUS_ERROR_DNS "NXDOMAIN", "No hostname known for address %s ", ip);
-                else {
-                        const char *rc, *n;
-                        char p[DECIMAL_STR_MAX(q->rcode)];
+        if (q->state != DNS_QUERY_SUCCESS) {
+                r = reply_query_state(q);
+                goto finish;
+        }
 
-                        rc = dns_rcode_to_string(q->rcode);
-                        if (!rc) {
-                                sprintf(p, "%i", q->rcode);
-                                rc = p;
-                        }
+        assert(q->received);
 
-                        n = strappenda(_BUS_ERROR_DNS, rc);
+        r = dns_name_reverse(q->request_family, &q->request_address, &reverse);
+        if (r < 0)
+                goto finish;
 
-                        sd_bus_error_setf(&error, n, "Could not resolve address %s, server or network returned error %s", ip, rc);
-                }
+        r = dns_packet_skip_question(q->received);
+        if (r < 0)
+                goto parse_fail;
 
-                r = sd_bus_reply_method_error(q->request, &error);
-                break;
-        }
+        r = sd_bus_message_new_method_return(q->request, &reply);
+        if (r < 0)
+                goto finish;
 
-        case DNS_QUERY_SUCCESS: {
-                _cleanup_bus_message_unref_ sd_bus_message *reply = NULL;
-                unsigned i, n, added = 0;
-                _cleanup_free_ char *reverse = NULL;
+        r = sd_bus_message_open_container(reply, 'a', "s");
+        if (r < 0)
+                goto finish;
 
-                assert(q->packet);
+        n = DNS_PACKET_ANCOUNT(q->received) +
+            DNS_PACKET_NSCOUNT(q->received) +
+            DNS_PACKET_ARCOUNT(q->received);
 
-                r = dns_name_reverse(q->request_family, &q->request_address, &reverse);
-                if (r < 0)
-                        goto finish;
+        for (i = 0; i < n; i++) {
+                _cleanup_(dns_resource_record_unrefp) DnsResourceRecord *rr = NULL;
 
-                r = dns_packet_skip_question(q->packet);
+                r = dns_packet_read_rr(q->received, &rr, NULL);
                 if (r < 0)
                         goto parse_fail;
 
-                r = sd_bus_message_new_method_return(q->request, &reply);
-                if (r < 0)
-                        goto finish;
+                if (rr->key.class != DNS_CLASS_IN)
+                        continue;
+                if (rr->key.type != DNS_TYPE_PTR)
+                        continue;
+                if (!dns_name_equal(rr->key.name, reverse))
+                        continue;
 
-                r = sd_bus_message_open_container(reply, 'a', "s");
+                r = sd_bus_message_append(reply, "s", rr->ptr.name);
                 if (r < 0)
                         goto finish;
 
-                n = DNS_PACKET_ANCOUNT(q->packet) +
-                    DNS_PACKET_NSCOUNT(q->packet) +
-                    DNS_PACKET_ARCOUNT(q->packet);
-                for (i = 0; i < n; i++) {
-                        _cleanup_(dns_resource_record_unrefp) DnsResourceRecord *rr = NULL;
-
-                        r = dns_packet_read_rr(q->packet, &rr, NULL);
-                        if (r < 0)
-                                goto parse_fail;
-
-                        if (rr->key.class != DNS_CLASS_IN)
-                                continue;
-                        if (rr->key.type != DNS_TYPE_PTR)
-                                continue;
-                        if (!dns_name_equal(rr->key.name, reverse))
-                                continue;
-
-                        r = sd_bus_message_append(reply, "s", rr->ptr.name);
-                        if (r < 0)
-                                goto finish;
-
-                        added ++;
-                }
+                added ++;
+        }
 
-                if (added <= 0) {
-                        r = sd_bus_reply_method_errorf(q->request, BUS_ERROR_NO_SUCH_RR, "Address %s does not have RR of this type", ip);
-                        break;
-                }
+        if (added <= 0) {
+                _cleanup_free_ char *ip = NULL;
 
-                r = sd_bus_message_close_container(reply);
-                if (r < 0)
-                        goto finish;
+                in_addr_to_string(q->request_family, &q->request_address, &ip);
 
-                r = sd_bus_send(q->manager->bus, reply, NULL);
-                break;
+                r = sd_bus_reply_method_errorf(q->request, BUS_ERROR_NO_SUCH_RR, "Address '%s' does not have RR of this type", ip);
+                goto finish;
         }
 
-        parse_fail:
-        case DNS_QUERY_INVALID_REPLY:
-                r = sd_bus_reply_method_errorf(q->request, BUS_ERROR_INVALID_REPLY, "Received invalid reply");
-                break;
+        r = sd_bus_message_close_container(reply);
+        if (r < 0)
+                goto finish;
 
-        case DNS_QUERY_NULL:
-        case DNS_QUERY_SENT:
-                assert_not_reached("Unexpected query state");
-        }
+        r = sd_bus_send(q->manager->bus, reply, NULL);
+        goto finish;
+
+parse_fail:
+        r = sd_bus_reply_method_errorf(q->request, BUS_ERROR_INVALID_REPLY, "Received invalid reply");
 
 finish:
         if (r < 0)
diff --git a/src/resolve/resolved-dns-packet.h b/src/resolve/resolved-dns-packet.h
index 10f8f96915..d4c9c475f5 100644
--- a/src/resolve/resolved-dns-packet.h
+++ b/src/resolve/resolved-dns-packet.h
@@ -71,6 +71,7 @@ static inline uint8_t* DNS_PACKET_DATA(DnsPacket *p) {
 #define DNS_PACKET_QR(p) ((be16toh(DNS_PACKET_HEADER(p)->flags) >> 15) & 1)
 #define DNS_PACKET_OPCODE(p) ((be16toh(DNS_PACKET_HEADER(p)->flags) >> 11) & 15)
 #define DNS_PACKET_RCODE(p) (be16toh(DNS_PACKET_HEADER(p)->flags) & 15)
+#define DNS_PACKET_TC(p) ((be16toh(DNS_PACKET_HEADER(p)->flags) >> 9) & 1)
 #define DNS_PACKET_QDCOUNT(p) be16toh(DNS_PACKET_HEADER(p)->qdcount)
 #define DNS_PACKET_ANCOUNT(p) be16toh(DNS_PACKET_HEADER(p)->ancount)
 #define DNS_PACKET_NSCOUNT(p) be16toh(DNS_PACKET_HEADER(p)->nscount)
diff --git a/src/resolve/resolved-dns-query.c b/src/resolve/resolved-dns-query.c
index dd812150cf..5bd5920254 100644
--- a/src/resolve/resolved-dns-query.c
+++ b/src/resolve/resolved-dns-query.c
@@ -31,7 +31,12 @@ DnsQueryTransaction* dns_query_transaction_free(DnsQueryTransaction *t) {
                 return NULL;
 
         sd_event_source_unref(t->timeout_event_source);
-        dns_packet_unref(t->packet);
+
+        dns_packet_unref(t->sent);
+        dns_packet_unref(t->received);
+
+        sd_event_source_unref(t->tcp_event_source);
+        safe_close(t->tcp_fd);
 
         if (t->query) {
                 LIST_REMOVE(transactions_by_query, t->query->transactions, t);
@@ -62,6 +67,8 @@ static int dns_query_transaction_new(DnsQuery *q, DnsQueryTransaction **ret, Dns
         if (!t)
                 return -ENOMEM;
 
+        t->tcp_fd = -1;
+
         do
                 random_bytes(&t->id, sizeof(t->id));
         while (t->id == 0 ||
@@ -87,6 +94,14 @@ static int dns_query_transaction_new(DnsQuery *q, DnsQueryTransaction **ret, Dns
         return 0;
 }
 
+static void dns_query_transaction_stop(DnsQueryTransaction *t) {
+        assert(t);
+
+        t->timeout_event_source = sd_event_source_unref(t->timeout_event_source);
+        t->tcp_event_source = sd_event_source_unref(t->tcp_event_source);
+        t->tcp_fd = safe_close(t->tcp_fd);
+}
+
 static void dns_query_transaction_set_state(DnsQueryTransaction *t, DnsQueryState state) {
         assert(t);
 
@@ -95,24 +110,179 @@ static void dns_query_transaction_set_state(DnsQueryTransaction *t, DnsQueryStat
 
         t->state = state;
 
-        if (state != DNS_QUERY_SENT)
-                t->timeout_event_source = sd_event_source_unref(t->timeout_event_source);
+        if (state != DNS_QUERY_SENT) {
+                dns_query_transaction_stop(t);
+                dns_query_finish(t->query);
+        }
+}
+
+static int on_tcp_ready(sd_event_source *s, int fd, uint32_t revents, void *userdata) {
+        DnsQueryTransaction *t = userdata;
+        int r;
+
+        assert(t);
+
+        if (revents & EPOLLOUT) {
+                struct iovec iov[2];
+                be16_t sz;
+                ssize_t ss;
+
+                sz = htobe16(t->sent->size);
+
+                iov[0].iov_base = &sz;
+                iov[0].iov_len = sizeof(sz);
+                iov[1].iov_base = DNS_PACKET_DATA(t->sent);
+                iov[1].iov_len = t->sent->size;
+
+                IOVEC_INCREMENT(iov, 2, t->tcp_written);
+
+                ss = writev(fd, iov, 2);
+                if (ss < 0) {
+                        if (errno != EINTR && errno != EAGAIN) {
+                                dns_query_transaction_set_state(t, DNS_QUERY_RESOURCES);
+                                return -errno;
+                        }
+                } else
+                        t->tcp_written += ss;
+
+                /* Are we done? If so, disable the event source for EPOLLOUT */
+                if (t->tcp_written >= sizeof(sz) + t->sent->size) {
+                        r = sd_event_source_set_io_events(s, EPOLLIN);
+                        if (r < 0) {
+                                dns_query_transaction_set_state(t, DNS_QUERY_RESOURCES);
+                                return r;
+                        }
+                }
+        }
+
+        if (revents & (EPOLLIN|EPOLLHUP|EPOLLRDHUP)) {
+
+                if (t->tcp_read < sizeof(t->tcp_read_size)) {
+                        ssize_t ss;
+
+                        ss = read(fd, (uint8_t*) &t->tcp_read_size + t->tcp_read, sizeof(t->tcp_read_size) - t->tcp_read);
+                        if (ss < 0) {
+                                if (errno != EINTR && errno != EAGAIN) {
+                                        dns_query_transaction_set_state(t, DNS_QUERY_RESOURCES);
+                                        return -errno;
+                                }
+                        } else if (ss == 0) {
+                                dns_query_transaction_set_state(t, DNS_QUERY_RESOURCES);
+                                return -EIO;
+                        } else
+                                t->tcp_read += ss;
+                }
+
+                if (t->tcp_read >= sizeof(t->tcp_read_size)) {
+
+                        if (be16toh(t->tcp_read_size) < DNS_PACKET_HEADER_SIZE) {
+                                dns_query_transaction_set_state(t, DNS_QUERY_INVALID_REPLY);
+                                return -EBADMSG;
+                        }
+
+                        if (t->tcp_read < sizeof(t->tcp_read_size) + be16toh(t->tcp_read_size)) {
+                                ssize_t ss;
+
+                                if (!t->received) {
+                                        r = dns_packet_new(&t->received, be16toh(t->tcp_read_size));
+                                        if (r < 0) {
+                                                dns_query_transaction_set_state(t, DNS_QUERY_RESOURCES);
+                                                return r;
+                                        }
+                                }
+
+                                ss = read(fd,
+                                          (uint8_t*) DNS_PACKET_DATA(t->received) + t->tcp_read - sizeof(t->tcp_read_size),
+                                          sizeof(t->tcp_read_size) + be16toh(t->tcp_read_size) - t->tcp_read);
+                                if (ss < 0) {
+                                        if (errno != EINTR && errno != EAGAIN) {
+                                                dns_query_transaction_set_state(t, DNS_QUERY_RESOURCES);
+                                                return -errno;
+                                        }
+                                } else if (ss == 0) {
+                                        dns_query_transaction_set_state(t, DNS_QUERY_RESOURCES);
+                                        return -EIO;
+                                }  else
+                                        t->tcp_read += ss;
+                        }
+
+                        if (t->tcp_read >= sizeof(t->tcp_read_size) + be16toh(t->tcp_read_size)) {
+                                t->received->size = be16toh(t->tcp_read_size);
+                                dns_query_transaction_reply(t, t->received);
+                                return 0;
+                        }
+                }
+        }
+
+        return 0;
+}
+
+static int dns_query_transaction_start_tcp(DnsQueryTransaction *t) {
+        int r;
+
+        assert(t);
+
+        if (t->tcp_fd >= 0)
+                return 0;
+
+        t->tcp_written = 0;
+        t->tcp_read = 0;
+        t->received = dns_packet_unref(t->received);
+
+        t->tcp_fd = dns_scope_tcp_socket(t->scope);
+        if (t->tcp_fd < 0)
+                return t->tcp_fd;
+
+        r = sd_event_add_io(t->query->manager->event, &t->tcp_event_source, t->tcp_fd, EPOLLIN|EPOLLOUT, on_tcp_ready, t);
+        if (r < 0) {
+                t->tcp_fd = safe_close(t->tcp_fd);
+                return r;
+        }
 
-        dns_query_finish(t->query);
+        return 0;
 }
 
-int dns_query_transaction_reply(DnsQueryTransaction *t, DnsPacket *p) {
+void dns_query_transaction_reply(DnsQueryTransaction *t, DnsPacket *p) {
+        int r;
+
         assert(t);
         assert(p);
 
-        t->packet = dns_packet_ref(p);
+        if (t->state != DNS_QUERY_SENT)
+                return;
+
+        if (t->received != p) {
+                dns_packet_unref(t->received);
+                t->received = dns_packet_ref(p);
+        }
+
+        if (t->tcp_fd >= 0) {
+                if (DNS_PACKET_TC(p)) {
+                        /* Truncated via TCP? Somebody must be fucking with us */
+                        dns_query_transaction_set_state(t, DNS_QUERY_INVALID_REPLY);
+                        return;
+                }
+
+                if (DNS_PACKET_ID(p) != t->id) {
+                        /* Not the reply to our query? Somebody must be fucking with us */
+                        dns_query_transaction_set_state(t, DNS_QUERY_INVALID_REPLY);
+                        return;
+                }
+        }
+
+        if (DNS_PACKET_TC(p)) {
+                /* Response was truncated, let's try again with good old TCP */
+                r = dns_query_transaction_start_tcp(t);
+                if (r < 0) {
+                        dns_query_transaction_set_state(t, DNS_QUERY_RESOURCES);
+                        return;
+                }
+        }
 
         if (DNS_PACKET_RCODE(p) == DNS_RCODE_SUCCESS)
                 dns_query_transaction_set_state(t, DNS_QUERY_SUCCESS);
         else
                 dns_query_transaction_set_state(t, DNS_QUERY_FAILURE);
-
-        return 0;
 }
 
 static int on_transaction_timeout(sd_event_source *s, usec_t usec, void *userdata) {
@@ -127,26 +297,20 @@ static int on_transaction_timeout(sd_event_source *s, usec_t usec, void *userdat
 
         r = dns_query_transaction_start(t);
         if (r < 0)
-                dns_query_transaction_set_state(t, DNS_QUERY_FAILURE);
+                dns_query_transaction_set_state(t, DNS_QUERY_RESOURCES);
 
         return 0;
 }
 
-int dns_query_transaction_start(DnsQueryTransaction *t) {
+static int dns_query_make_packet(DnsQueryTransaction *t) {
         _cleanup_(dns_packet_unrefp) DnsPacket *p = NULL;
         unsigned n;
         int r;
 
         assert(t);
 
-        t->timeout_event_source = sd_event_source_unref(t->timeout_event_source);
-
-        if (t->n_attempts >= ATTEMPTS_MAX) {
-                dns_query_transaction_set_state(t, DNS_QUERY_ATTEMPTS_MAX);
+        if (t->sent)
                 return 0;
-        }
-
-        t->n_attempts++;
 
         r = dns_packet_new_query(&p, 0);
         if (r < 0)
@@ -161,7 +325,38 @@ int dns_query_transaction_start(DnsQueryTransaction *t) {
         DNS_PACKET_HEADER(p)->qdcount = htobe16(t->query->n_keys);
         DNS_PACKET_HEADER(p)->id = t->id;
 
-        r = dns_scope_send(t->scope, p);
+        t->sent = p;
+        p = NULL;
+
+        return 0;
+}
+
+int dns_query_transaction_start(DnsQueryTransaction *t) {
+        int r;
+
+        assert(t);
+
+        dns_query_transaction_stop(t);
+
+        if (t->n_attempts >= ATTEMPTS_MAX) {
+                dns_query_transaction_set_state(t, DNS_QUERY_ATTEMPTS_MAX);
+                return 0;
+        }
+        t->n_attempts++;
+
+        r = dns_query_make_packet(t);
+        if (r < 0)
+                return r;
+
+        /* Try via UDP, and if that fails due to large size try via TCP */
+        r = dns_scope_send(t->scope, t->sent);
+        if (r == -EMSGSIZE)
+                r = dns_query_transaction_start_tcp(t);
+
+        if (r == -ESRCH) {
+                dns_query_transaction_set_state(t, DNS_QUERY_NO_SERVERS);
+                return 0;
+        }
         if (r < 0) {
                 /* Couldn't send? Try immediately again, with a new server */
                 dns_scope_next_dns_server(t->scope);
@@ -169,18 +364,12 @@ int dns_query_transaction_start(DnsQueryTransaction *t) {
                 return dns_query_transaction_start(t);
         }
 
-        if (r > 0) {
-                int q;
-
-                q = sd_event_add_time(t->query->manager->event, &t->timeout_event_source, CLOCK_MONOTONIC, now(CLOCK_MONOTONIC) + TRANSACTION_TIMEOUT_USEC, 0, on_transaction_timeout, t);
-                if (q < 0)
-                        return q;
-
-                dns_query_transaction_set_state(t, DNS_QUERY_SENT);
-        } else
-                dns_query_transaction_set_state(t, DNS_QUERY_SKIPPED);
+        r = sd_event_add_time(t->query->manager->event, &t->timeout_event_source, CLOCK_MONOTONIC, now(CLOCK_MONOTONIC) + TRANSACTION_TIMEOUT_USEC, 0, on_transaction_timeout, t);
+        if (r < 0)
+                return r;
 
-        return r;
+        dns_query_transaction_set_state(t, DNS_QUERY_SENT);
+        return 1;
 }
 
 DnsQuery *dns_query_free(DnsQuery *q) {
@@ -190,7 +379,7 @@ DnsQuery *dns_query_free(DnsQuery *q) {
                 return NULL;
 
         sd_bus_message_unref(q->request);
-        dns_packet_unref(q->packet);
+        dns_packet_unref(q->received);
         sd_event_source_unref(q->timeout_event_source);
 
         while (q->transactions)
@@ -367,8 +556,8 @@ fail:
 
 void dns_query_finish(DnsQuery *q) {
         DnsQueryTransaction *t;
-        DnsQueryState state = DNS_QUERY_SKIPPED;
-        uint16_t rcode = 0;
+        DnsQueryState state = DNS_QUERY_NO_SERVERS;
+        DnsPacket *received = NULL;
 
         assert(q);
 
@@ -381,25 +570,28 @@ void dns_query_finish(DnsQuery *q) {
                 if (t->state == DNS_QUERY_SENT || t->state == DNS_QUERY_NULL)
                         return;
 
-                /* One of the transactions is sucecssful, let's use it */
+                /* One of the transactions is successful, let's use it */
                 if (t->state == DNS_QUERY_SUCCESS) {
-                        q->packet = dns_packet_ref(t->packet);
+                        q->received = dns_packet_ref(t->received);
                         dns_query_set_state(q, DNS_QUERY_SUCCESS);
                         return;
                 }
 
+                /* One of the transactions has failed, let's see
+                 * whether we find anything better, but if not, return
+                 * its response packet */
                 if (t->state == DNS_QUERY_FAILURE) {
+                        received = t->received;
                         state = DNS_QUERY_FAILURE;
+                        continue;
+                }
 
-                        if (rcode == 0 && t->packet)
-                                rcode = DNS_PACKET_RCODE(t->packet);
-
-                } else if (state == DNS_QUERY_SKIPPED && t->state != DNS_QUERY_SKIPPED)
+                if (state == DNS_QUERY_NO_SERVERS && t->state != DNS_QUERY_NO_SERVERS)
                         state = t->state;
         }
 
         if (state == DNS_QUERY_FAILURE)
-                q->rcode = rcode;
+                q->received = dns_packet_ref(received);
 
         dns_query_set_state(q, state);
 }
diff --git a/src/resolve/resolved-dns-query.h b/src/resolve/resolved-dns-query.h
index a07f174706..d5bc08dc30 100644
--- a/src/resolve/resolved-dns-query.h
+++ b/src/resolve/resolved-dns-query.h
@@ -39,10 +39,11 @@ typedef enum DnsQueryState {
         DNS_QUERY_SENT,
         DNS_QUERY_FAILURE,
         DNS_QUERY_SUCCESS,
-        DNS_QUERY_SKIPPED,
+        DNS_QUERY_NO_SERVERS,
         DNS_QUERY_TIMEOUT,
         DNS_QUERY_ATTEMPTS_MAX,
         DNS_QUERY_INVALID_REPLY,
+        DNS_QUERY_RESOURCES,
 } DnsQueryState;
 
 struct DnsQueryTransaction {
@@ -55,7 +56,13 @@ struct DnsQueryTransaction {
         sd_event_source *timeout_event_source;
         unsigned n_attempts;
 
-        DnsPacket *packet;
+        DnsPacket *sent, *received;
+
+        /* TCP connection logic */
+        int tcp_fd;
+        sd_event_source *tcp_event_source;
+        size_t tcp_written, tcp_read;
+        be16_t tcp_read_size;
 
         LIST_FIELDS(DnsQueryTransaction, transactions_by_query);
         LIST_FIELDS(DnsQueryTransaction, transactions_by_scope);
@@ -71,8 +78,7 @@ struct DnsQuery {
 
         sd_event_source *timeout_event_source;
 
-        uint16_t rcode;
-        DnsPacket *packet;
+        DnsPacket *received;
 
         sd_bus_message *request;
         unsigned char request_family;
@@ -92,6 +98,6 @@ void dns_query_finish(DnsQuery *q);
 
 DnsQueryTransaction* dns_query_transaction_free(DnsQueryTransaction *t);
 int dns_query_transaction_start(DnsQueryTransaction *t);
-int dns_query_transaction_reply(DnsQueryTransaction *t, DnsPacket *p);
+void dns_query_transaction_reply(DnsQueryTransaction *t, DnsPacket *p);
 
 DEFINE_TRIVIAL_CLEANUP_FUNC(DnsQuery*, dns_query_free);
diff --git a/src/resolve/resolved-dns-scope.c b/src/resolve/resolved-dns-scope.c
index 4e0a742768..0a70cb1f0f 100644
--- a/src/resolve/resolved-dns-scope.c
+++ b/src/resolve/resolved-dns-scope.c
@@ -19,7 +19,10 @@
   along with systemd; If not, see <http://www.gnu.org/licenses/>.
 ***/
 
+#include <netinet/tcp.h>
+
 #include "strv.h"
+#include "socket-util.h"
 #include "resolved-dns-domain.h"
 #include "resolved-dns-scope.h"
 
@@ -92,10 +95,14 @@ int dns_scope_send(DnsScope *s, DnsPacket *p) {
 
         srv = dns_scope_get_server(s);
         if (!srv)
-                return 0;
+                return -ESRCH;
+
+        if (s->link) {
+                if (p->size > s->link->mtu)
+                        return -EMSGSIZE;
 
-        if (s->link)
                 ifindex = s->link->ifindex;
+        }
 
         if (srv->family == AF_INET)
                 r = manager_dns_ipv4_send(s->manager, srv, ifindex, p);
@@ -110,6 +117,52 @@ int dns_scope_send(DnsScope *s, DnsPacket *p) {
         return 1;
 }
 
+int dns_scope_tcp_socket(DnsScope *s) {
+        _cleanup_close_ int fd = -1;
+        union sockaddr_union sa = {};
+        socklen_t salen;
+        int one, ifindex, ret;
+        DnsServer *srv;
+        int r;
+
+        assert(s);
+
+        srv = dns_scope_get_server(s);
+        if (!srv)
+                return -ESRCH;
+
+        if (s->link)
+                ifindex = s->link->ifindex;
+
+        sa.sa.sa_family = srv->family;
+        if (srv->family == AF_INET) {
+                sa.in.sin_port = htobe16(53);
+                sa.in.sin_addr = srv->address.in;
+                salen = sizeof(sa.in);
+        } else if (srv->family == AF_INET6) {
+                sa.in6.sin6_port = htobe16(53);
+                sa.in6.sin6_addr = srv->address.in6;
+                sa.in6.sin6_scope_id = ifindex;
+                salen = sizeof(sa.in6);
+        } else
+                return -EAFNOSUPPORT;
+
+        fd = socket(srv->family, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0);
+        if (fd < 0)
+                return -errno;
+
+        one = 1;
+        setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &one, sizeof(one));
+
+        r = connect(fd, &sa.sa, salen);
+        if (r < 0 && errno != EINPROGRESS)
+                return -errno;
+
+        ret = fd;
+        fd = -1;
+        return ret;
+}
+
 DnsScopeMatch dns_scope_test(DnsScope *s, const char *domain) {
         char **i;
 
diff --git a/src/resolve/resolved-dns-scope.h b/src/resolve/resolved-dns-scope.h
index 50f1106bf1..97544f9e48 100644
--- a/src/resolve/resolved-dns-scope.h
+++ b/src/resolve/resolved-dns-scope.h
@@ -63,6 +63,8 @@ int dns_scope_new(Manager *m, DnsScope **ret, DnsScopeType t);
 DnsScope* dns_scope_free(DnsScope *s);
 
 int dns_scope_send(DnsScope *s, DnsPacket *p);
+int dns_scope_tcp_socket(DnsScope *s);
+
 DnsScopeMatch dns_scope_test(DnsScope *s, const char *domain);
 
 DnsServer *dns_scope_get_server(DnsScope *s);
diff --git a/src/resolve/resolved-manager.c b/src/resolve/resolved-manager.c
index 09b758f3f4..3d2979dbbe 100644
--- a/src/resolve/resolved-manager.c
+++ b/src/resolve/resolved-manager.c
@@ -577,7 +577,7 @@ int manager_dns_ipv4_recv(Manager *m, DnsPacket **ret) {
 
         l = recvmsg(fd, &mh, 0);
         if (l < 0) {
-                if (errno == EAGAIN)
+                if (errno == EAGAIN || errno == EINTR)
                         return 0;
 
                 return -errno;
@@ -626,7 +626,7 @@ int manager_dns_ipv6_recv(Manager *m, DnsPacket **ret) {
 
         l = recvmsg(fd, &mh, 0);
         if (l < 0) {
-                if (errno == EAGAIN)
+                if (errno == EAGAIN || errno == EINTR)
                         return 0;
 
                 return -errno;
@@ -657,7 +657,8 @@ static int on_dns_ipv4_packet(sd_event_source *s, int fd, uint32_t revents, void
         if (!t)
                 return 0;
 
-        return dns_query_transaction_reply(t, p);
+        dns_query_transaction_reply(t, p);
+        return 0;
 }
 
 static int on_dns_ipv6_packet(sd_event_source *s, int fd, uint32_t revents, void *userdata) {
@@ -674,7 +675,8 @@ static int on_dns_ipv6_packet(sd_event_source *s, int fd, uint32_t revents, void
         if (!t)
                 return 0;
 
-        return dns_query_transaction_reply(t, p);
+        dns_query_transaction_reply(t, p);
+        return 0;
 }
 
 int manager_dns_ipv4_fd(Manager *m) {