#include "openflow/openflow.h"
#include "openvswitch/vlog.h"
#include "ovn-controller.h"
-#include "vswitch-idl.h"
+#include "physical.h"
#include "rconn.h"
#include "socket-util.h"
+#include "vswitch-idl.h"
VLOG_DEFINE_THIS_MODULE(ofctrl);
static void ovn_flow_log(const struct ovn_flow *, const char *action);
static void ovn_flow_destroy(struct ovn_flow *);
+static ovs_be32 queue_msg(struct ofpbuf *);
+static void queue_flow_mod(struct ofputil_flow_mod *);
+
/* OpenFlow connection to the switch. */
static struct rconn *swconn;
* rconn_get_connection_seqno(rconn), 'swconn' has reconnected. */
static unsigned int seqno;
+/* Connection state machine. */
+#define STATES \
+ STATE(S_NEW) \
+ STATE(S_GENEVE_TABLE_REQUESTED) \
+ STATE(S_GENEVE_TABLE_MOD_SENT) \
+ STATE(S_CLEAR_FLOWS) \
+ STATE(S_UPDATE_FLOWS)
+enum ofctrl_state {
+#define STATE(NAME) NAME,
+ STATES
+#undef STATE
+};
+
+/* Current state. */
+static enum ofctrl_state state;
+
+/* Transaction IDs for messages in flight to the switch. */
+static ovs_be32 xid, xid2;
+
/* Counter for in-flight OpenFlow messages on 'swconn'. We only send a new
* round of flow table modifications to the switch when the counter falls to
* zero, to avoid unbounded buffering. */
* installed in the switch. */
static struct hmap installed_flows;
+/* MFF_* field ID for our Geneve option. In S_GENEVE_TABLE_MOD_SENT, this is
+ * the option we requested (we don't know whether we obtained it yet). In
+ * S_CLEAR_FLOWS or S_UPDATE_FLOWS, this is really the option we have. */
+static enum mf_field_id mff_ovn_geneve;
+
static void ovn_flow_table_clear(struct hmap *flow_table);
static void ovn_flow_table_destroy(struct hmap *flow_table);
-static void ofctrl_update_flows(struct hmap *desired_flows);
-static void ofctrl_recv(const struct ofpbuf *msg);
+static void ofctrl_recv(const struct ofp_header *, enum ofptype);
void
ofctrl_init(void)
tx_counter = rconn_packet_counter_create();
hmap_init(&installed_flows);
}
+\f
+/* S_NEW, for a new connection.
+ *
+ * Sends NXT_GENEVE_TABLE_REQUEST and transitions to
+ * S_GENEVE_TABLE_REQUESTED. */
-/* Attempts to update the OpenFlow flows in bridge 'br_int' to those in
- * 'flow_table'. Removes all of the flows from 'flow_table' and frees them.
+static void
+run_S_NEW(void)
+{
+ struct ofpbuf *buf = ofpraw_alloc(OFPRAW_NXT_GENEVE_TABLE_REQUEST,
+ rconn_get_version(swconn), 0);
+ xid = queue_msg(buf);
+ state = S_GENEVE_TABLE_REQUESTED;
+}
+
+static void
+recv_S_NEW(const struct ofp_header *oh OVS_UNUSED,
+ enum ofptype type OVS_UNUSED)
+{
+ OVS_NOT_REACHED();
+}
+\f
+/* S_GENEVE_TABLE_REQUESTED, when NXT_GENEVE_TABLE_REQUEST has been sent
+ * and we're waiting for a reply.
*
- * The flow table will only be updated if we've got an OpenFlow connection to
- * 'br_int' and it's not backlogged. Otherwise, it'll have to wait until the
- * next iteration. */
-void
-ofctrl_run(const struct ovsrec_bridge *br_int, struct hmap *flow_table)
+ * If we receive an NXT_GENEVE_TABLE_REPLY:
+ *
+ * - If it contains our tunnel metadata option, assign its field ID to
+ * mff_ovn_geneve and transition to S_CLEAR_FLOWS.
+ *
+ * - Otherwise, if there is an unused tunnel metadata field ID, send
+ * NXT_GENEVE_TABLE_MOD and OFPT_BARRIER_REQUEST, and transition to
+ * S_GENEVE_TABLE_MOD_SENT.
+ *
+ * - Otherwise, log an error, disable Geneve, and transition to
+ * S_CLEAR_FLOWS.
+ *
+ * If we receive an OFPT_ERROR:
+ *
+ * - Log an error, disable Geneve, and transition to S_CLEAR_FLOWS. */
+
+static void
+run_S_GENEVE_TABLE_REQUESTED(void)
+{
+}
+
+static void
+recv_S_GENEVE_TABLE_REQUESTED(const struct ofp_header *oh, enum ofptype type)
+{
+ if (oh->xid != xid) {
+ ofctrl_recv(oh, type);
+ } else if (type == OFPTYPE_NXT_GENEVE_TABLE_REPLY) {
+ struct ofputil_geneve_table_reply reply;
+ enum ofperr error = ofputil_decode_geneve_table_reply(oh, &reply);
+ if (error) {
+ VLOG_ERR("failed to decode Geneve table request (%s)",
+ ofperr_to_string(error));
+ goto error;
+ }
+
+ const struct ofputil_geneve_map *map;
+ uint64_t md_free = UINT64_MAX;
+ BUILD_ASSERT(TUN_METADATA_NUM_OPTS == 64);
+
+ LIST_FOR_EACH (map, list_node, &reply.mappings) {
+ if (map->option_class == OVN_GENEVE_CLASS
+ && map->option_type == OVN_GENEVE_TYPE
+ && map->option_len == OVN_GENEVE_LEN) {
+ if (map->index >= TUN_METADATA_NUM_OPTS) {
+ VLOG_ERR("desired Geneve tunnel option 0x%"PRIx16","
+ "%"PRIu8",%"PRIu8" already in use with "
+ "unsupported index %"PRIu16,
+ map->option_class, map->option_type,
+ map->option_len, map->index);
+ goto error;
+ } else {
+ mff_ovn_geneve = MFF_TUN_METADATA0 + map->index;
+ state = S_CLEAR_FLOWS;
+ return;
+ }
+ }
+
+ if (map->index < TUN_METADATA_NUM_OPTS) {
+ md_free &= ~(UINT64_C(1) << map->index);
+ }
+ }
+
+ VLOG_DBG("OVN Geneve option not found");
+ if (!md_free) {
+ VLOG_ERR("no Geneve options free for use by OVN");
+ goto error;
+ }
+
+ unsigned int index = rightmost_1bit_idx(md_free);
+ mff_ovn_geneve = MFF_TUN_METADATA0 + index;
+ struct ofputil_geneve_map gm;
+ gm.option_class = OVN_GENEVE_CLASS;
+ gm.option_type = OVN_GENEVE_TYPE;
+ gm.option_len = OVN_GENEVE_LEN;
+ gm.index = index;
+
+ struct ofputil_geneve_table_mod gtm;
+ gtm.command = NXGTMC_ADD;
+ list_init(>m.mappings);
+ list_push_back(>m.mappings, &gm.list_node);
+
+ xid = queue_msg(ofputil_encode_geneve_table_mod(OFP13_VERSION, >m));
+ xid2 = queue_msg(ofputil_encode_barrier_request(OFP13_VERSION));
+ state = S_GENEVE_TABLE_MOD_SENT;
+ } else if (type == OFPTYPE_ERROR) {
+ VLOG_ERR("switch refused to allocate Geneve option (%s)",
+ ofperr_to_string(ofperr_decode_msg(oh, NULL)));
+ goto error;
+ } else {
+ char *s = ofp_to_string(oh, ntohs(oh->length), 1);
+ VLOG_ERR("unexpected reply to Geneve table request (%s)",
+ s);
+ free(s);
+ goto error;
+ }
+ return;
+
+error:
+ mff_ovn_geneve = 0;
+ state = S_CLEAR_FLOWS;
+}
+\f
+/* S_GENEVE_TABLE_MOD_SENT, when NXT_GENEVE_TABLE_MOD and OFPT_BARRIER_REQUEST
+ * have been sent and we're waiting for a reply to one or the other.
+ *
+ * If we receive an OFPT_ERROR:
+ *
+ * - If the error is NXGTMFC_ALREADY_MAPPED or NXGTMFC_DUP_ENTRY, we
+ * raced with some other controller. Transition to S_NEW.
+ *
+ * - Otherwise, log an error, disable Geneve, and transition to
+ * S_CLEAR_FLOWS.
+ *
+ * If we receive OFPT_BARRIER_REPLY:
+ *
+ * - Set the tunnel metadata field ID to the one that we requested.
+ * Transition to S_CLEAR_FLOWS.
+ */
+
+static void
+run_S_GENEVE_TABLE_MOD_SENT(void)
+{
+}
+
+static void
+recv_S_GENEVE_TABLE_MOD_SENT(const struct ofp_header *oh, enum ofptype type)
+{
+ if (oh->xid != xid && oh->xid != xid2) {
+ ofctrl_recv(oh, type);
+ } else if (oh->xid == xid2 && type == OFPTYPE_BARRIER_REPLY) {
+ state = S_CLEAR_FLOWS;
+ } else if (oh->xid == xid && type == OFPTYPE_ERROR) {
+ enum ofperr error = ofperr_decode_msg(oh, NULL);
+ if (error == OFPERR_NXGTMFC_ALREADY_MAPPED ||
+ error == OFPERR_NXGTMFC_DUP_ENTRY) {
+ VLOG_INFO("raced with another controller adding "
+ "Geneve option (%s); trying again",
+ ofperr_to_string(error));
+ state = S_NEW;
+ } else {
+ VLOG_ERR("error adding Geneve option (%s)",
+ ofperr_to_string(error));
+ goto error;
+ }
+ } else {
+ char *s = ofp_to_string(oh, ntohs(oh->length), 1);
+ VLOG_ERR("unexpected reply to Geneve option allocation request (%s)",
+ s);
+ free(s);
+ goto error;
+ }
+ return;
+
+error:
+ state = S_CLEAR_FLOWS;
+}
+\f
+/* S_CLEAR_FLOWS, after we've established a Geneve metadata field ID and it's
+ * time to set up some flows.
+ *
+ * Sends an OFPT_TABLE_MOD to clear all flows, then transitions to
+ * S_UPDATE_FLOWS. */
+
+static void
+run_S_CLEAR_FLOWS(void)
+{
+ /* Send a flow_mod to delete all flows. */
+ struct ofputil_flow_mod fm = {
+ .match = MATCH_CATCHALL_INITIALIZER,
+ .table_id = OFPTT_ALL,
+ .command = OFPFC_DELETE,
+ };
+ queue_flow_mod(&fm);
+ VLOG_DBG("clearing all flows");
+
+ /* Clear installed_flows, to match the state of the switch. */
+ ovn_flow_table_clear(&installed_flows);
+
+ state = S_UPDATE_FLOWS;
+}
+
+static void
+recv_S_CLEAR_FLOWS(const struct ofp_header *oh, enum ofptype type)
+{
+ ofctrl_recv(oh, type);
+}
+\f
+/* S_UPDATE_FLOWS, for maintaining the flow table over time.
+ *
+ * Compare the installed flows to the ones we want. Send OFPT_FLOW_MOD as
+ * necessary.
+ *
+ * This is a terminal state. We only transition out of it if the connection
+ * drops. */
+
+static void
+run_S_UPDATE_FLOWS(void)
+{
+ /* Nothing to do here.
+ *
+ * Being in this state enables ofctrl_put() to work, however. */
+}
+
+static void
+recv_S_UPDATE_FLOWS(const struct ofp_header *oh, enum ofptype type)
+{
+ ofctrl_recv(oh, type);
+}
+\f
+/* Runs the OpenFlow state machine against 'br_int', which is local to the
+ * hypervisor on which we are running. Attempts to negotiate a Geneve option
+ * field for class OVN_GENEVE_CLASS, type OVN_GENEVE_TYPE. If successful,
+ * returns the MFF_* field ID for the option, otherwise returns 0. */
+enum mf_field_id
+ofctrl_run(const struct ovsrec_bridge *br_int)
{
if (br_int) {
char *target;
rconn_run(swconn);
if (!rconn_is_connected(swconn)) {
- goto exit;
+ return 0;
}
- if (!rconn_packet_counter_n_packets(tx_counter)) {
- ofctrl_update_flows(flow_table);
+ if (seqno != rconn_get_connection_seqno(swconn)) {
+ seqno = rconn_get_connection_seqno(swconn);
+ state = S_NEW;
}
- for (int i = 0; i < 50; i++) {
+ enum ofctrl_state old_state;
+ do {
+ old_state = state;
+ switch (state) {
+#define STATE(NAME) case NAME: run_##NAME(); break;
+ STATES
+#undef STATE
+ default:
+ OVS_NOT_REACHED();
+ }
+ } while (state != old_state);
+
+ for (int i = 0; state == old_state && i < 50; i++) {
struct ofpbuf *msg = rconn_recv(swconn);
if (!msg) {
break;
}
- ofctrl_recv(msg);
+ const struct ofp_header *oh = msg->data;
+ enum ofptype type;
+ enum ofperr error;
+
+ error = ofptype_decode(&type, oh);
+ if (!error) {
+ switch (state) {
+#define STATE(NAME) case NAME: recv_##NAME(oh, type); break;
+ STATES
+#undef STATE
+ default:
+ OVS_NOT_REACHED();
+ }
+ } else {
+ char *s = ofp_to_string(oh, ntohs(oh->length), 1);
+ VLOG_WARN("could not decode OpenFlow message (%s): %s",
+ ofperr_to_string(error), s);
+ free(s);
+ }
+
ofpbuf_delete(msg);
}
-exit:
- ovn_flow_table_clear(flow_table);
+ return (state == S_CLEAR_FLOWS || state == S_UPDATE_FLOWS
+ ? mff_ovn_geneve : 0);
}
void
rconn_packet_counter_destroy(tx_counter);
}
\f
-static void
+static ovs_be32
queue_msg(struct ofpbuf *msg)
{
+ const struct ofp_header *oh = msg->data;
+ ovs_be32 xid = oh->xid;
rconn_send(swconn, msg, tx_counter);
+ return xid;
}
static void
-ofctrl_recv(const struct ofpbuf *msg)
+ofctrl_recv(const struct ofp_header *oh, enum ofptype type)
{
- enum ofptype type;
- struct ofpbuf b;
-
- b = *msg;
- if (ofptype_pull(&type, &b)) {
- return;
- }
-
- switch (type) {
- case OFPTYPE_ECHO_REQUEST:
- queue_msg(make_echo_reply(msg->data));
- break;
-
- case OFPTYPE_ECHO_REPLY:
- case OFPTYPE_PACKET_IN:
- case OFPTYPE_PORT_STATUS:
- case OFPTYPE_FLOW_REMOVED:
- /* Nothing to do. */
- break;
-
- case OFPTYPE_HELLO:
- case OFPTYPE_ERROR:
- case OFPTYPE_FEATURES_REQUEST:
- case OFPTYPE_FEATURES_REPLY:
- case OFPTYPE_GET_CONFIG_REQUEST:
- case OFPTYPE_GET_CONFIG_REPLY:
- case OFPTYPE_SET_CONFIG:
- case OFPTYPE_PACKET_OUT:
- case OFPTYPE_FLOW_MOD:
- case OFPTYPE_GROUP_MOD:
- case OFPTYPE_PORT_MOD:
- case OFPTYPE_TABLE_MOD:
- case OFPTYPE_BARRIER_REQUEST:
- case OFPTYPE_BARRIER_REPLY:
- case OFPTYPE_QUEUE_GET_CONFIG_REQUEST:
- case OFPTYPE_QUEUE_GET_CONFIG_REPLY:
- case OFPTYPE_DESC_STATS_REQUEST:
- case OFPTYPE_DESC_STATS_REPLY:
- case OFPTYPE_FLOW_STATS_REQUEST:
- case OFPTYPE_FLOW_STATS_REPLY:
- case OFPTYPE_AGGREGATE_STATS_REQUEST:
- case OFPTYPE_AGGREGATE_STATS_REPLY:
- case OFPTYPE_TABLE_STATS_REQUEST:
- case OFPTYPE_TABLE_STATS_REPLY:
- case OFPTYPE_PORT_STATS_REQUEST:
- case OFPTYPE_PORT_STATS_REPLY:
- case OFPTYPE_QUEUE_STATS_REQUEST:
- case OFPTYPE_QUEUE_STATS_REPLY:
- case OFPTYPE_PORT_DESC_STATS_REQUEST:
- case OFPTYPE_PORT_DESC_STATS_REPLY:
- case OFPTYPE_ROLE_REQUEST:
- case OFPTYPE_ROLE_REPLY:
- case OFPTYPE_ROLE_STATUS:
- case OFPTYPE_SET_FLOW_FORMAT:
- case OFPTYPE_FLOW_MOD_TABLE_ID:
- case OFPTYPE_SET_PACKET_IN_FORMAT:
- case OFPTYPE_FLOW_AGE:
- case OFPTYPE_SET_CONTROLLER_ID:
- case OFPTYPE_FLOW_MONITOR_STATS_REQUEST:
- case OFPTYPE_FLOW_MONITOR_STATS_REPLY:
- case OFPTYPE_FLOW_MONITOR_CANCEL:
- case OFPTYPE_FLOW_MONITOR_PAUSED:
- case OFPTYPE_FLOW_MONITOR_RESUMED:
- case OFPTYPE_GET_ASYNC_REQUEST:
- case OFPTYPE_GET_ASYNC_REPLY:
- case OFPTYPE_SET_ASYNC_CONFIG:
- case OFPTYPE_METER_MOD:
- case OFPTYPE_GROUP_STATS_REQUEST:
- case OFPTYPE_GROUP_STATS_REPLY:
- case OFPTYPE_GROUP_DESC_STATS_REQUEST:
- case OFPTYPE_GROUP_DESC_STATS_REPLY:
- case OFPTYPE_GROUP_FEATURES_STATS_REQUEST:
- case OFPTYPE_GROUP_FEATURES_STATS_REPLY:
- case OFPTYPE_METER_STATS_REQUEST:
- case OFPTYPE_METER_STATS_REPLY:
- case OFPTYPE_METER_CONFIG_STATS_REQUEST:
- case OFPTYPE_METER_CONFIG_STATS_REPLY:
- case OFPTYPE_METER_FEATURES_STATS_REQUEST:
- case OFPTYPE_METER_FEATURES_STATS_REPLY:
- case OFPTYPE_TABLE_FEATURES_STATS_REQUEST:
- case OFPTYPE_TABLE_FEATURES_STATS_REPLY:
- case OFPTYPE_TABLE_DESC_REQUEST:
- case OFPTYPE_TABLE_DESC_REPLY:
- case OFPTYPE_BUNDLE_CONTROL:
- case OFPTYPE_BUNDLE_ADD_MESSAGE:
- case OFPTYPE_NXT_GENEVE_TABLE_MOD:
- case OFPTYPE_NXT_GENEVE_TABLE_REQUEST:
- case OFPTYPE_NXT_GENEVE_TABLE_REPLY:
- default:
- /* Messages that are generally unexpected. */
+ if (type == OFPTYPE_ECHO_REQUEST) {
+ queue_msg(make_echo_reply(oh));
+ } else if (type != OFPTYPE_ECHO_REPLY &&
+ type != OFPTYPE_BARRIER_REPLY &&
+ type != OFPTYPE_PACKET_IN &&
+ type != OFPTYPE_PORT_STATUS &&
+ type != OFPTYPE_FLOW_REMOVED) {
if (VLOG_IS_DBG_ENABLED()) {
static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(30, 300);
- char *s = ofp_to_string(msg->data, msg->size, 2);
+ char *s = ofp_to_string(oh, ntohs(oh->length), 2);
VLOG_DBG_RL(&rl, "OpenFlow packet ignored: %s", s);
free(s);
}
ovn_flow_destroy(f);
}
}
+
static void
ovn_flow_table_destroy(struct hmap *flow_table)
{
queue_msg(ofputil_encode_flow_mod(fm, OFPUTIL_P_OF13_OXM));
}
-static void
-ofctrl_update_flows(struct hmap *desired_flows)
+/* Replaces the flow table on the switch, if possible, by the flows in
+ * 'flow_table', which should have been added with ofctrl_add_flow().
+ * Regardless of whether the flow table is updated, this deletes all of the
+ * flows from 'flow_table' and frees them. (The hmap itself isn't
+ * destroyed.)
+ *
+ * This called be called be ofctrl_run() within the main loop. */
+void
+ofctrl_put(struct hmap *flow_table)
{
- /* If we've (re)connected, don't make any assumptions about the flows in
- * the switch: delete all of them. (We'll immediately repopulate it
- * below.) */
- if (seqno != rconn_get_connection_seqno(swconn)) {
- seqno = rconn_get_connection_seqno(swconn);
-
- /* Send a flow_mod to delete all flows. */
- struct ofputil_flow_mod fm = {
- .match = MATCH_CATCHALL_INITIALIZER,
- .table_id = OFPTT_ALL,
- .command = OFPFC_DELETE,
- };
- queue_flow_mod(&fm);
- VLOG_DBG("clearing all flows");
-
- /* Clear installed_flows, to match the state of the switch. */
- ovn_flow_table_clear(&installed_flows);
+ /* The flow table can be updated if the connection to the switch is up and
+ * in the correct state and not backlogged with existing flow_mods. (Our
+ * criteria for being backlogged appear very conservative, but the socket
+ * between ovn-controller and OVS provides some buffering.) Otherwise,
+ * discard the flows. A solution to either of those problems will cause us
+ * to wake up and retry. */
+ if (state != S_UPDATE_FLOWS
+ || rconn_packet_counter_n_packets(tx_counter)) {
+ ovn_flow_table_clear(flow_table);
+ return;
}
/* Iterate through all of the installed flows. If any of them are no
* actions, update them. */
struct ovn_flow *i, *next;
HMAP_FOR_EACH_SAFE (i, next, hmap_node, &installed_flows) {
- struct ovn_flow *d = ovn_flow_lookup(desired_flows, i);
+ struct ovn_flow *d = ovn_flow_lookup(flow_table, i);
if (!d) {
/* Installed flow is no longer desirable. Delete it from the
* switch and from installed_flows. */
d->ofpacts_len = 0;
}
- hmap_remove(desired_flows, &d->hmap_node);
+ hmap_remove(flow_table, &d->hmap_node);
ovn_flow_destroy(d);
}
}
- /* The previous loop removed from 'desired_flows' all of the flows that are
- * already installed. Thus, any flows remaining in 'desired_flows' need to
+ /* The previous loop removed from 'flow_table' all of the flows that are
+ * already installed. Thus, any flows remaining in 'flow_table' need to
* be added to the flow table. */
struct ovn_flow *d;
- HMAP_FOR_EACH_SAFE (d, next, hmap_node, desired_flows) {
+ HMAP_FOR_EACH_SAFE (d, next, hmap_node, flow_table) {
/* Send flow_mod to add flow. */
struct ofputil_flow_mod fm = {
.match = d->match,
queue_flow_mod(&fm);
ovn_flow_log(d, "adding");
- /* Move 'd' from 'desired_flows' to installed_flows. */
- hmap_remove(desired_flows, &d->hmap_node);
+ /* Move 'd' from 'flow_table' to installed_flows. */
+ hmap_remove(flow_table, &d->hmap_node);
hmap_insert(&installed_flows, &d->hmap_node, d->hmap_node.hash);
}
}
projects / cascardo / ovs.git / commitdiff