@@ -9,6 +9,9 @@ Post-v2.15.0
- Userspace datapath:
* Auto load balancing of PMDs now partially supports cross-NUMA polling
cases, e.g if all PMD threads are running on the same NUMA node.
+ * Add all-zero IP SNAT handling to conntrack. In case of collision,
+ using ct(src=0.0.0.0), the source port will be replaced with another
+ non-colliding port in the ephemeral range (1024, 65535).
- ovs-ctl:
* New option '--no-record-hostname' to disable hostname configuration
in ovsdb on startup.
@@ -148,6 +148,39 @@ enum ct_update_res {
CT_TIMEOUT(ICMP_FIRST) \
CT_TIMEOUT(ICMP_REPLY)
+#define NAT_ACTION_SNAT_ALL (NAT_ACTION_SRC | NAT_ACTION_SRC_PORT)
+#define NAT_ACTION_DNAT_ALL (NAT_ACTION_DST | NAT_ACTION_DST_PORT)
+
+enum ct_ephemeral_range {
+ MIN_NAT_EPHEMERAL_PORT = 1024,
+ MAX_NAT_EPHEMERAL_PORT = 65535
+};
+
+#define IN_RANGE(curr, min, max) \
+ (curr >= min && curr <= max)
+
+#define NEXT_PORT_IN_RANGE(curr, min, max) \
+ (curr = (!IN_RANGE(curr, min, max) || curr == max) ? min : curr + 1)
+
+/* if the current port is out of range increase the attempts by
+ * one so that in the worst case scenario the current out of
+ * range port plus all the in-range ports get tested.
+ * Note that curr can be an out of range port only in case of
+ * source port (SNAT with port range unspecified or DNAT),
+ * furthermore the source port in the packet has to be less than
+ * MIN_NAT_EPHEMERAL_PORT. */
+#define N_PORT_ATTEMPTS(curr, min, max) \
+ ((!IN_RANGE(curr, min, max)) ? (max - min) + 2 : (max - min) + 1)
+
+/* loose in-range check, the first curr port can be any port out of
+ * the range. */
+#define FOR_EACH_PORT_IN_RANGE__(curr, min, max, INAME) \
+ for (uint16_t INAME = N_PORT_ATTEMPTS(curr, min, max); \
+ INAME > 0; INAME--, NEXT_PORT_IN_RANGE(curr, min, max))
+
+#define FOR_EACH_PORT_IN_RANGE(curr, min, max) \
+ FOR_EACH_PORT_IN_RANGE__(curr, min, max, OVS_JOIN(idx, __COUNTER__))
+
enum ct_timeout {
#define CT_TIMEOUT(NAME) CT_TM_##NAME,
CT_TIMEOUTS
@@ -108,8 +108,8 @@ static void set_label(struct dp_packet *, struct conn *,
static void *clean_thread_main(void *f_);
static bool
-nat_select_range_tuple(struct conntrack *ct, const struct conn *conn,
- struct conn *nat_conn);
+nat_get_unique_tuple(struct conntrack *ct, const struct conn *conn,
+ struct conn *nat_conn);
static uint8_t
reverse_icmp_type(uint8_t type);
@@ -728,11 +728,11 @@ pat_packet(struct dp_packet *pkt, const struct conn *conn)
}
} else if (conn->nat_info->nat_action & NAT_ACTION_DST) {
if (conn->key.nw_proto == IPPROTO_TCP) {
- struct tcp_header *th = dp_packet_l4(pkt);
- packet_set_tcp_port(pkt, th->tcp_src, conn->rev_key.src.port);
+ packet_set_tcp_port(pkt, conn->rev_key.dst.port,
+ conn->rev_key.src.port);
} else if (conn->key.nw_proto == IPPROTO_UDP) {
- struct udp_header *uh = dp_packet_l4(pkt);
- packet_set_udp_port(pkt, uh->udp_src, conn->rev_key.src.port);
+ packet_set_udp_port(pkt, conn->rev_key.dst.port,
+ conn->rev_key.src.port);
}
}
}
@@ -786,11 +786,9 @@ un_pat_packet(struct dp_packet *pkt, const struct conn *conn)
}
} else if (conn->nat_info->nat_action & NAT_ACTION_DST) {
if (conn->key.nw_proto == IPPROTO_TCP) {
- struct tcp_header *th = dp_packet_l4(pkt);
- packet_set_tcp_port(pkt, conn->key.dst.port, th->tcp_dst);
+ packet_set_tcp_port(pkt, conn->key.dst.port, conn->key.src.port);
} else if (conn->key.nw_proto == IPPROTO_UDP) {
- struct udp_header *uh = dp_packet_l4(pkt);
- packet_set_udp_port(pkt, conn->key.dst.port, uh->udp_dst);
+ packet_set_udp_port(pkt, conn->key.dst.port, conn->key.src.port);
}
}
}
@@ -810,12 +808,10 @@ reverse_pat_packet(struct dp_packet *pkt, const struct conn *conn)
}
} else if (conn->nat_info->nat_action & NAT_ACTION_DST) {
if (conn->key.nw_proto == IPPROTO_TCP) {
- struct tcp_header *th_in = dp_packet_l4(pkt);
- packet_set_tcp_port(pkt, th_in->tcp_src,
+ packet_set_tcp_port(pkt, conn->key.src.port,
conn->key.dst.port);
} else if (conn->key.nw_proto == IPPROTO_UDP) {
- struct udp_header *uh_in = dp_packet_l4(pkt);
- packet_set_udp_port(pkt, uh_in->udp_src,
+ packet_set_udp_port(pkt, conn->key.src.port,
conn->key.dst.port);
}
}
@@ -1029,14 +1025,14 @@ conn_not_found(struct conntrack *ct, struct dp_packet *pkt,
}
} else {
memcpy(nat_conn, nc, sizeof *nat_conn);
- bool nat_res = nat_select_range_tuple(ct, nc, nat_conn);
+ bool nat_res = nat_get_unique_tuple(ct, nc, nat_conn);
if (!nat_res) {
goto nat_res_exhaustion;
}
/* Update nc with nat adjustments made to nat_conn by
- * nat_select_range_tuple(). */
+ * nat_get_unique_tuple(). */
memcpy(nc, nat_conn, sizeof *nc);
}
@@ -2238,130 +2234,221 @@ nat_range_hash(const struct conn *conn, uint32_t basis)
return hash_finish(hash, 0);
}
-static bool
-nat_select_range_tuple(struct conntrack *ct, const struct conn *conn,
- struct conn *nat_conn)
-{
- enum { MIN_NAT_EPHEMERAL_PORT = 1024,
- MAX_NAT_EPHEMERAL_PORT = 65535 };
-
- uint16_t min_port;
- uint16_t max_port;
- uint16_t first_port;
- uint32_t hash = nat_range_hash(conn, ct->hash_basis);
+/* Ports are stored in host byte order for convenience. */
+static void
+set_sport_range(struct nat_action_info_t *ni, const struct conn_key *k,
+ uint32_t hash, uint16_t *curr, uint16_t *min,
+ uint16_t *max)
+{
+ if (((ni->nat_action & NAT_ACTION_SNAT_ALL) == NAT_ACTION_SRC) ||
+ ((ni->nat_action & NAT_ACTION_DST))) {
+ *curr = ntohs(k->src.port);
+ *min = MIN_NAT_EPHEMERAL_PORT;
+ *max = MAX_NAT_EPHEMERAL_PORT;
+ } else {
+ *min = ni->min_port;
+ *max = ni->max_port;
+ *curr = *min + (hash % ((*max - *min) + 1));
+ }
+}
- if ((conn->nat_info->nat_action & NAT_ACTION_SRC) &&
- (!(conn->nat_info->nat_action & NAT_ACTION_SRC_PORT))) {
- min_port = ntohs(conn->key.src.port);
- max_port = ntohs(conn->key.src.port);
- first_port = min_port;
- } else if ((conn->nat_info->nat_action & NAT_ACTION_DST) &&
- (!(conn->nat_info->nat_action & NAT_ACTION_DST_PORT))) {
- min_port = ntohs(conn->key.dst.port);
- max_port = ntohs(conn->key.dst.port);
- first_port = min_port;
+static void
+set_dport_range(struct nat_action_info_t *ni, const struct conn_key *k,
+ uint32_t hash, uint16_t *curr, uint16_t *min,
+ uint16_t *max)
+{
+ if (ni->nat_action & NAT_ACTION_DST_PORT) {
+ *min = ni->min_port;
+ *max = ni->max_port;
+ *curr = *min + (hash % ((*max - *min) + 1));
} else {
- uint16_t deltap = conn->nat_info->max_port - conn->nat_info->min_port;
- uint32_t port_index = hash % (deltap + 1);
- first_port = conn->nat_info->min_port + port_index;
- min_port = conn->nat_info->min_port;
- max_port = conn->nat_info->max_port;
+ *curr = ntohs(k->dst.port);
+ *min = *max = *curr;
}
+}
- uint32_t deltaa = 0;
- uint32_t address_index;
- union ct_addr ct_addr;
- memset(&ct_addr, 0, sizeof ct_addr);
- union ct_addr max_ct_addr;
- memset(&max_ct_addr, 0, sizeof max_ct_addr);
- max_ct_addr = conn->nat_info->max_addr;
+/* Gets the initial in range address based on the hash.
+ * Addresses are kept in network order. */
+static void
+get_addr_in_range(union ct_addr *min, union ct_addr *max,
+ union ct_addr *curr, uint32_t hash,
+ bool ipv4)
+{
+ uint32_t offt, range;
- if (conn->key.dl_type == htons(ETH_TYPE_IP)) {
- deltaa = ntohl(conn->nat_info->max_addr.ipv4) -
- ntohl(conn->nat_info->min_addr.ipv4);
- address_index = hash % (deltaa + 1);
- ct_addr.ipv4 = htonl(
- ntohl(conn->nat_info->min_addr.ipv4) + address_index);
+ if (ipv4) {
+ range = (ntohl(max->ipv4) - ntohl(min->ipv4)) + 1;
+ offt = hash % range;
+ curr->ipv4 = htonl(ntohl(min->ipv4) + offt);
} else {
- deltaa = nat_ipv6_addrs_delta(&conn->nat_info->min_addr.ipv6,
- &conn->nat_info->max_addr.ipv6);
- /* deltaa must be within 32 bits for full hash coverage. A 64 or
+ range = nat_ipv6_addrs_delta(&min->ipv6,
+ &max->ipv6) + 1;
+ /* range must be within 32 bits for full hash coverage. A 64 or
* 128 bit hash is unnecessary and hence not used here. Most code
* is kept common with V4; nat_ipv6_addrs_delta() will do the
* enforcement via max_ct_addr. */
- max_ct_addr = conn->nat_info->min_addr;
- nat_ipv6_addr_increment(&max_ct_addr.ipv6, deltaa);
- address_index = hash % (deltaa + 1);
- ct_addr.ipv6 = conn->nat_info->min_addr.ipv6;
- nat_ipv6_addr_increment(&ct_addr.ipv6, address_index);
- }
-
- uint16_t port = first_port;
- bool all_ports_tried = false;
- /* For DNAT or for specified port ranges, we don't use ephemeral ports. */
- bool ephemeral_ports_tried
- = conn->nat_info->nat_action & NAT_ACTION_DST ||
- conn->nat_info->nat_action & NAT_ACTION_SRC_PORT
- ? true : false;
- union ct_addr first_addr = ct_addr;
- bool pat_enabled = conn->key.nw_proto == IPPROTO_TCP ||
- conn->key.nw_proto == IPPROTO_UDP;
-
- while (true) {
+ offt = hash % range;
+ curr->ipv6 = min->ipv6;
+ nat_ipv6_addr_increment(&curr->ipv6, offt);
+ }
+}
+
+static void
+get_initial_addr(const struct conn *conn, union ct_addr *min,
+ union ct_addr *max, union ct_addr *curr,
+ uint32_t hash, bool ipv4)
+{
+ const union ct_addr zero_ip = {0};
+
+ /* all-zero CASE */
+ if (!memcmp(min, &zero_ip, sizeof(*min))) {
if (conn->nat_info->nat_action & NAT_ACTION_SRC) {
- nat_conn->rev_key.dst.addr = ct_addr;
- if (pat_enabled) {
- nat_conn->rev_key.dst.port = htons(port);
- }
- } else {
- nat_conn->rev_key.src.addr = ct_addr;
- if (pat_enabled) {
- nat_conn->rev_key.src.port = htons(port);
- }
+ *curr = conn->key.src.addr;
+ } else if (conn->nat_info->nat_action & NAT_ACTION_DST) {
+ *curr = conn->key.dst.addr;
+ }
+ } else {
+ get_addr_in_range(min, max, curr, hash, ipv4);
+ }
+}
+
+static void
+store_addr_to_key(union ct_addr *addr, struct conn_key *key,
+ uint16_t action)
+{
+ if (action & NAT_ACTION_SRC) {
+ key->dst.addr = *addr;
+ } else {
+ key->src.addr = *addr;
+ }
+}
+
+static void
+next_addr_in_range(union ct_addr *curr, union ct_addr *min,
+ union ct_addr *max, bool ipv4)
+{
+ if (ipv4) {
+ /* this check could be unified with IPv6, but let's avoid
+ * an unneeded memcmp() in case of IPv4. */
+ if (min->ipv4 == max->ipv4) {
+ return;
+ }
+
+ curr->ipv4 = (curr->ipv4 == max->ipv4) ?
+ min->ipv4 :
+ htonl(ntohl(curr->ipv4) + 1);
+ } else {
+ if (!memcmp(min, max, sizeof(*min))) {
+ return;
+ }
+
+ if (!memcmp(curr, max, sizeof(*curr))) {
+ *curr = *min;
+ return;
}
- bool found = conn_lookup(ct, &nat_conn->rev_key, time_msec(), NULL,
- NULL);
- if (!found) {
+ nat_ipv6_addr_increment(&curr->ipv6, 1);
+ }
+}
+
+static bool
+next_addr_in_range_guarded(union ct_addr *curr, union ct_addr *min,
+ union ct_addr *max, union ct_addr *guard,
+ bool ipv4)
+{
+ bool exhausted;
+
+ next_addr_in_range(curr, min, max, ipv4);
+
+ if (ipv4) {
+ exhausted = (curr->ipv4 == guard->ipv4);
+ } else {
+ exhausted = !memcmp(curr, guard, sizeof(*curr));
+ }
+
+ return exhausted;
+}
+
+/* This function tries to get a unique tuple.
+ * Every iteration checks that the reverse tuple doesn't
+ * collide with any existing one.
+ *
+ * in case of SNAT:
+ * - for each src IP address in the range (if any)
+ * - try to find a source port in range (if any)
+ * - if no port range exists, use the whole
+ * ephemeral range (after testing the port
+ * used by the sender), otherwise use the
+ * specified range
+ *
+ * in case of DNAT:
+ * - for each dst IP address in the range (if any)
+ * - for each dport in range (if any)
+ * - try to find a source port in the ephemeral range
+ * (after testing the port used by the sender)
+ *
+ * If none can be found, return exhaustion to the caller. */
+static bool
+nat_get_unique_tuple(struct conntrack *ct, const struct conn *conn,
+ struct conn *nat_conn)
+{
+ union ct_addr min_addr = {0}, max_addr = {0}, curr_addr = {0},
+ guard_addr = {0};
+ uint32_t hash = nat_range_hash(conn, ct->hash_basis);
+ bool pat_proto = conn->key.nw_proto == IPPROTO_TCP ||
+ conn->key.nw_proto == IPPROTO_UDP;
+ uint16_t min_dport, max_dport, curr_dport;
+ uint16_t min_sport, max_sport, curr_sport;
+
+ min_addr = conn->nat_info->min_addr;
+ max_addr = conn->nat_info->max_addr;
+
+ get_initial_addr(conn, &min_addr, &max_addr, &curr_addr, hash,
+ (conn->key.dl_type == htons(ETH_TYPE_IP)));
+
+ /* save the address we started from so that
+ * we can stop once we reach it. */
+ guard_addr = curr_addr;
+
+ set_sport_range(conn->nat_info, &conn->key, hash, &curr_sport,
+ &min_sport, &max_sport);
+ set_dport_range(conn->nat_info, &conn->key, hash, &curr_dport,
+ &min_dport, &max_dport);
+
+another_round:
+ store_addr_to_key(&curr_addr, &nat_conn->rev_key,
+ conn->nat_info->nat_action);
+
+ if (!pat_proto) {
+ if (!conn_lookup(ct, &nat_conn->rev_key,
+ time_msec(), NULL, NULL)) {
return true;
- } else if (pat_enabled && !all_ports_tried) {
- if (min_port == max_port) {
- all_ports_tried = true;
- } else if (port == max_port) {
- port = min_port;
- } else {
- port++;
- }
- if (port == first_port) {
- all_ports_tried = true;
- }
- } else {
- if (memcmp(&ct_addr, &max_ct_addr, sizeof ct_addr)) {
- if (conn->key.dl_type == htons(ETH_TYPE_IP)) {
- ct_addr.ipv4 = htonl(ntohl(ct_addr.ipv4) + 1);
- } else {
- nat_ipv6_addr_increment(&ct_addr.ipv6, 1);
- }
- } else {
- ct_addr = conn->nat_info->min_addr;
- }
- if (!memcmp(&ct_addr, &first_addr, sizeof ct_addr)) {
- if (pat_enabled && !ephemeral_ports_tried) {
- ephemeral_ports_tried = true;
- ct_addr = conn->nat_info->min_addr;
- first_addr = ct_addr;
- min_port = MIN_NAT_EPHEMERAL_PORT;
- max_port = MAX_NAT_EPHEMERAL_PORT;
- } else {
- break;
- }
+ }
+
+ goto next_addr;
+ }
+
+ FOR_EACH_PORT_IN_RANGE(curr_dport, min_dport, max_dport) {
+ nat_conn->rev_key.src.port = htons(curr_dport);
+ FOR_EACH_PORT_IN_RANGE(curr_sport, min_sport, max_sport) {
+ nat_conn->rev_key.dst.port = htons(curr_sport);
+ if (!conn_lookup(ct, &nat_conn->rev_key,
+ time_msec(), NULL, NULL)) {
+ return true;
}
- first_port = min_port;
- port = first_port;
- all_ports_tried = false;
}
}
- return false;
+
+ /* Check if next IP is in range and respin. Otherwise, notify
+ * exhaustion to the caller. */
+next_addr:
+ if (next_addr_in_range_guarded(&curr_addr, &min_addr,
+ &max_addr, &guard_addr,
+ conn->key.dl_type == htons(ETH_TYPE_IP))) {
+ return false;
+ }
+
+ goto another_round;
}
static enum ct_update_res
@@ -2138,8 +2138,7 @@ for <var>i</var> in [1,<var>n_members</var>]:
<code>nat(src=0.0.0.0)</code>. In this case, when a source port
collision is detected during the commit, the source port will be
translated to an ephemeral port. If there is no collision, no SNAT
- is performed. Note that this is currently only implemented in the
- Linux kernel datapath.
+ is performed.
</p>
<p>
@@ -99,12 +99,10 @@ m4_define([CHECK_CONNTRACK_NAT])
# CHECK_CONNTRACK_ZEROIP_SNAT()
#
# Perform requirements checks for running conntrack all-zero IP SNAT tests.
-# The userspace datapath does not support all-zero IP SNAT.
+# The userspace datapath always supports all-zero IP SNAT, so no check is
+# needed.
#
-m4_define([CHECK_CONNTRACK_ZEROIP_SNAT],
-[
- AT_SKIP_IF([:])
-])
+m4_define([CHECK_CONNTRACK_ZEROIP_SNAT])
# CHECK_CONNTRACK_TIMEOUT()
#