/*
- * Copyright (c) 2009, 2010, 2011, 2012, 2013 Nicira, Inc.
+ * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
*
* 'ofproto_mutex' must be held whenever an ofconn is created or destroyed or,
* more or less equivalently, whenever an ofconn is added to or removed from a
- * connmgr. 'ofproto_mutex' doesn't protect the data inside the ofconn. */
+ * connmgr. 'ofproto_mutex' doesn't protect the data inside the ofconn, except
+ * as specifically noted below. */
struct ofconn {
/* Configuration that persists from one connection to the next. */
uint32_t master_async_config[OAM_N_TYPES]; /* master, other */
uint32_t slave_async_config[OAM_N_TYPES]; /* slave */
- /* Flow monitors. */
- struct hmap monitors; /* Contains "struct ofmonitor"s. */
- struct list updates; /* List of "struct ofpbuf"s. */
- bool sent_abbrev_update; /* Does 'updates' contain NXFME_ABBREV? */
- struct rconn_packet_counter *monitor_counter;
- uint64_t monitor_paused;
+/* Flow monitors (e.g. NXST_FLOW_MONITOR). */
+
+ /* Configuration. Contains "struct ofmonitor"s. */
+ struct hmap monitors OVS_GUARDED_BY(ofproto_mutex);
+
+ /* Flow control.
+ *
+ * When too many flow monitor notifications back up in the transmit buffer,
+ * we pause the transmission of further notifications. These members track
+ * the flow control state.
+ *
+ * When notifications are flowing, 'monitor_paused' is 0. When
+ * notifications are paused, 'monitor_paused' is the value of
+ * 'monitor_seqno' at the point we paused.
+ *
+ * 'monitor_counter' counts the OpenFlow messages and bytes currently in
+ * flight. This value growing too large triggers pausing. */
+ uint64_t monitor_paused OVS_GUARDED_BY(ofproto_mutex);
+ struct rconn_packet_counter *monitor_counter OVS_GUARDED_BY(ofproto_mutex);
+
+ /* State of monitors for a single ongoing flow_mod.
+ *
+ * 'updates' is a list of "struct ofpbuf"s that contain
+ * NXST_FLOW_MONITOR_REPLY messages representing the changes made by the
+ * current flow_mod.
+ *
+ * When 'updates' is nonempty, 'sent_abbrev_update' is true if 'updates'
+ * contains an update event of type NXFME_ABBREV and false otherwise.. */
+ struct list updates OVS_GUARDED_BY(ofproto_mutex);
+ bool sent_abbrev_update OVS_GUARDED_BY(ofproto_mutex);
};
static struct ofconn *ofconn_create(struct connmgr *, struct rconn *,
static void ofconn_send(const struct ofconn *, struct ofpbuf *,
struct rconn_packet_counter *);
-static void do_send_packet_in(struct ofpbuf *, void *ofconn_);
+static void do_send_packet_ins(struct ofconn *, struct list *txq);
/* A listener for incoming OpenFlow "service" connections. */
struct ofservice {
OVS_REQUIRES(ofproto_mutex);
static struct ofconn *find_controller_by_target(struct connmgr *,
const char *target);
-static void update_fail_open(struct connmgr *);
+static void update_fail_open(struct connmgr *) OVS_EXCLUDED(ofproto_mutex);
static int set_pvconns(struct pvconn ***pvconnsp, size_t *n_pvconnsp,
const struct sset *);
shash_destroy(&new_controllers);
+ ovs_mutex_unlock(&ofproto_mutex);
+
update_in_band_remotes(mgr);
update_fail_open(mgr);
if (had_controllers != connmgr_has_controllers(mgr)) {
ofproto_flush_flows(mgr->ofproto);
}
- ovs_mutex_unlock(&ofproto_mutex);
}
/* Drops the connections between 'mgr' and all of its primary and secondary
/* Add all the remotes. */
HMAP_FOR_EACH (ofconn, hmap_node, &mgr->controllers) {
- struct sockaddr_in *sin = &addrs[n_addrs];
const char *target = rconn_get_target(ofconn->rconn);
+ struct sockaddr_storage ss;
- if (ofconn->band == OFPROTO_OUT_OF_BAND) {
- continue;
- }
-
- if (stream_parse_target_with_default_port(target,
- OFP_OLD_PORT,
- sin)) {
- n_addrs++;
+ if (ofconn->band == OFPROTO_IN_BAND
+ && stream_parse_target_with_default_port(target, OFP_OLD_PORT, &ss)
+ && ss.ss_family == AF_INET) {
+ addrs[n_addrs++] = *(struct sockaddr_in *) &ss;
}
}
for (i = 0; i < mgr->n_extra_remotes; i++) {
static void
update_fail_open(struct connmgr *mgr)
+ OVS_EXCLUDED(ofproto_mutex)
{
if (connmgr_has_controllers(mgr)
&& mgr->fail_mode == OFPROTO_FAIL_STANDALONE) {
return ofconn->role;
}
+void
+ofconn_send_role_status(struct ofconn *ofconn, uint32_t role, uint8_t reason)
+{
+ struct ofputil_role_status status;
+ struct ofpbuf *buf;
+
+ status.reason = reason;
+ status.role = role;
+ ofconn_get_master_election_id(ofconn, &status.generation_id);
+
+ buf = ofputil_encode_role_status(&status, ofconn_get_protocol(ofconn));
+
+ ofconn_send(ofconn, buf, NULL);
+}
+
/* Changes 'ofconn''s role to 'role'. If 'role' is OFPCR12_ROLE_MASTER then
* any existing master is demoted to a slave. */
void
ofconn_set_role(struct ofconn *ofconn, enum ofp12_controller_role role)
{
- if (role == OFPCR12_ROLE_MASTER) {
+ if (role != ofconn->role && role == OFPCR12_ROLE_MASTER) {
struct ofconn *other;
HMAP_FOR_EACH (other, hmap_node, &ofconn->connmgr->controllers) {
if (other->role == OFPCR12_ROLE_MASTER) {
other->role = OFPCR12_ROLE_SLAVE;
+ ofconn_send_role_status(other, OFPCR12_ROLE_SLAVE, OFPCRR_MASTER_REQUEST);
}
}
}
size_t i;
for (i = 0; i < N_SCHEDULERS; i++) {
- pinsched_run(ofconn->schedulers[i], do_send_packet_in, ofconn);
+ struct list txq;
+
+ pinsched_run(ofconn->schedulers[i], &txq);
+ do_send_packet_ins(ofconn, &txq);
}
rconn_run(ofconn->rconn);
return true;
}
+/* The default "table-miss" behaviour for OpenFlow1.3+ is to drop the
+ * packet rather than to send the packet to the controller.
+ *
+ * This function returns false to indicate the packet should be dropped if
+ * the controller action was the result of the default table-miss behaviour
+ * and the controller is using OpenFlow1.3+.
+ *
+ * Otherwise true is returned to indicate the packet should be forwarded to
+ * the controller */
+static bool
+ofconn_wants_packet_in_on_miss(struct ofconn *ofconn,
+ const struct ofproto_packet_in *pin)
+{
+ if (pin->miss_type == OFPROTO_PACKET_IN_MISS_WITHOUT_FLOW) {
+ enum ofputil_protocol protocol = ofconn_get_protocol(ofconn);
+
+ if (protocol != OFPUTIL_P_NONE
+ && ofputil_protocol_to_ofp_version(protocol) >= OFP13_VERSION) {
+ enum ofproto_table_config config;
+
+ config = ofproto_table_get_config(ofconn->connmgr->ofproto,
+ pin->up.table_id);
+ if (config == OFPROTO_TABLE_MISS_DEFAULT) {
+ return false;
+ }
+ }
+ }
+ return true;
+}
+
+/* The default "table-miss" behaviour for OpenFlow1.3+ is to drop the
+ * packet rather than to send the packet to the controller.
+ *
+ * This function returns false to indicate that a packet_in message
+ * for a "table-miss" should be sent to at least one controller.
+ * That is there is at least one controller with controller_id 0
+ * which connected using an OpenFlow version earlier than OpenFlow1.3.
+ *
+ * False otherwise.
+ *
+ * This logic assumes that "table-miss" packet_in messages
+ * are always sent to controller_id 0. */
+bool
+connmgr_wants_packet_in_on_miss(struct connmgr *mgr)
+{
+ struct ofconn *ofconn;
+
+ LIST_FOR_EACH (ofconn, node, &mgr->all_conns) {
+ enum ofputil_protocol protocol = ofconn_get_protocol(ofconn);
+
+ if (ofconn->controller_id == 0 &&
+ (protocol == OFPUTIL_P_NONE ||
+ ofputil_protocol_to_ofp_version(protocol) < OFP13_VERSION)) {
+ return true;
+ }
+ }
+ return false;
+}
+
/* Returns a human-readable name for an OpenFlow connection between 'mgr' and
* 'target', suitable for use in log messages for identifying the connection.
*
\f
/* Sending asynchronous messages. */
-static void schedule_packet_in(struct ofconn *, struct ofputil_packet_in);
+static void schedule_packet_in(struct ofconn *, struct ofproto_packet_in,
+ enum ofp_packet_in_reason wire_reason);
/* Sends an OFPT_PORT_STATUS message with 'opp' and 'reason' to appropriate
- * controllers managed by 'mgr'. */
+ * controllers managed by 'mgr'. For messages caused by a controller
+ * OFPT_PORT_MOD, specify 'source' as the controller connection that sent the
+ * request; otherwise, specify 'source' as NULL. */
void
-connmgr_send_port_status(struct connmgr *mgr,
+connmgr_send_port_status(struct connmgr *mgr, struct ofconn *source,
const struct ofputil_phy_port *pp, uint8_t reason)
{
/* XXX Should limit the number of queued port status change messages. */
if (ofconn_receives_async_msg(ofconn, OAM_PORT_STATUS, reason)) {
struct ofpbuf *msg;
+ /* Before 1.5, OpenFlow specified that OFPT_PORT_MOD should not
+ * generate OFPT_PORT_STATUS messages. That requirement was a
+ * relic of how OpenFlow originally supported a single controller,
+ * so that one could expect the controller to already know the
+ * changes it had made.
+ *
+ * EXT-338 changes OpenFlow 1.5 OFPT_PORT_MOD to send
+ * OFPT_PORT_STATUS messages to every controller. This is
+ * obviously more useful in the multi-controller case. We could
+ * always implement it that way in OVS, but that would risk
+ * confusing controllers that are intended for single-controller
+ * use only. (Imagine a controller that generates an OFPT_PORT_MOD
+ * in response to any OFPT_PORT_STATUS!)
+ *
+ * So this compromises: for OpenFlow 1.4 and earlier, it generates
+ * OFPT_PORT_STATUS for OFPT_PORT_MOD, but not back to the
+ * originating controller. In a single-controller environment, in
+ * particular, this means that it will never generate
+ * OFPT_PORT_STATUS for OFPT_PORT_MOD at all. */
+ if (ofconn == source
+ && rconn_get_version(ofconn->rconn) < OFP15_VERSION) {
+ continue;
+ }
+
msg = ofputil_encode_port_status(&ps, ofconn_get_protocol(ofconn));
ofconn_send(ofconn, msg, NULL);
}
}
}
+/* Normally a send-to-controller action uses reason OFPR_ACTION. However, in
+ * OpenFlow 1.3 and later, packet_ins generated by a send-to-controller action
+ * in a "table-miss" flow (one with priority 0 and completely wildcarded) are
+ * sent as OFPR_NO_MATCH. This function returns the reason that should
+ * actually be sent on 'ofconn' for 'pin'. */
+static enum ofp_packet_in_reason
+wire_reason(struct ofconn *ofconn, const struct ofproto_packet_in *pin)
+{
+ if (pin->miss_type == OFPROTO_PACKET_IN_MISS_FLOW
+ && pin->up.reason == OFPR_ACTION) {
+ enum ofputil_protocol protocol = ofconn_get_protocol(ofconn);
+
+ if (protocol != OFPUTIL_P_NONE
+ && ofputil_protocol_to_ofp_version(protocol) >= OFP13_VERSION) {
+ return OFPR_NO_MATCH;
+ }
+ }
+ return pin->up.reason;
+}
+
/* Given 'pin', sends an OFPT_PACKET_IN message to each OpenFlow controller as
* necessary according to their individual configurations.
*
* The caller doesn't need to fill in pin->buffer_id or pin->total_len. */
void
connmgr_send_packet_in(struct connmgr *mgr,
- const struct ofputil_packet_in *pin)
+ const struct ofproto_packet_in *pin)
{
struct ofconn *ofconn;
LIST_FOR_EACH (ofconn, node, &mgr->all_conns) {
- if (ofconn_receives_async_msg(ofconn, OAM_PACKET_IN, pin->reason)
+ enum ofp_packet_in_reason reason = wire_reason(ofconn, pin);
+
+ if (ofconn_wants_packet_in_on_miss(ofconn, pin)
+ && ofconn_receives_async_msg(ofconn, OAM_PACKET_IN, pin->up.reason)
&& ofconn->controller_id == pin->controller_id) {
- schedule_packet_in(ofconn, *pin);
+ schedule_packet_in(ofconn, *pin, reason);
}
}
}
-/* pinsched callback for sending 'ofp_packet_in' on 'ofconn'. */
static void
-do_send_packet_in(struct ofpbuf *ofp_packet_in, void *ofconn_)
+do_send_packet_ins(struct ofconn *ofconn, struct list *txq)
{
- struct ofconn *ofconn = ofconn_;
+ struct ofpbuf *pin, *next_pin;
+
+ LIST_FOR_EACH_SAFE (pin, next_pin, list_node, txq) {
+ list_remove(&pin->list_node);
- rconn_send_with_limit(ofconn->rconn, ofp_packet_in,
- ofconn->packet_in_counter, 100);
+ if (rconn_send_with_limit(ofconn->rconn, pin,
+ ofconn->packet_in_counter, 100) == EAGAIN) {
+ static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 5);
+
+ VLOG_INFO_RL(&rl, "%s: dropping packet-in due to queue overflow",
+ rconn_get_name(ofconn->rconn));
+ }
+ }
}
/* Takes 'pin', composes an OpenFlow packet-in message from it, and passes it
* to 'ofconn''s packet scheduler for sending. */
static void
-schedule_packet_in(struct ofconn *ofconn, struct ofputil_packet_in pin)
+schedule_packet_in(struct ofconn *ofconn, struct ofproto_packet_in pin,
+ enum ofp_packet_in_reason wire_reason)
{
struct connmgr *mgr = ofconn->connmgr;
+ uint16_t controller_max_len;
+ struct list txq;
- pin.total_len = pin.packet_len;
+ pin.up.total_len = pin.up.packet_len;
- /* Get OpenFlow buffer_id. */
- if (pin.reason == OFPR_ACTION) {
- pin.buffer_id = UINT32_MAX;
- } else if (mgr->fail_open && fail_open_is_active(mgr->fail_open)) {
- pin.buffer_id = pktbuf_get_null();
- } else if (!ofconn->pktbuf) {
- pin.buffer_id = UINT32_MAX;
+ pin.up.reason = wire_reason;
+ if (pin.up.reason == OFPR_ACTION) {
+ controller_max_len = pin.send_len; /* max_len */
} else {
- pin.buffer_id = pktbuf_save(ofconn->pktbuf, pin.packet, pin.packet_len,
- pin.fmd.in_port);
+ controller_max_len = ofconn->miss_send_len;
}
- /* Figure out how much of the packet to send. */
- if (pin.reason == OFPR_NO_MATCH) {
- pin.send_len = pin.packet_len;
+ /* Get OpenFlow buffer_id.
+ * For OpenFlow 1.2+, OFPCML_NO_BUFFER (== UINT16_MAX) specifies
+ * unbuffered. This behaviour doesn't violate prior versions, too. */
+ if (controller_max_len == UINT16_MAX) {
+ pin.up.buffer_id = UINT32_MAX;
+ } else if (mgr->fail_open && fail_open_is_active(mgr->fail_open)) {
+ pin.up.buffer_id = pktbuf_get_null();
+ } else if (!ofconn->pktbuf) {
+ pin.up.buffer_id = UINT32_MAX;
} else {
- /* Caller should have initialized 'send_len' to 'max_len' specified in
- * output action. */
+ pin.up.buffer_id = pktbuf_save(ofconn->pktbuf,
+ pin.up.packet, pin.up.packet_len,
+ pin.up.fmd.in_port);
}
- if (pin.buffer_id != UINT32_MAX) {
- pin.send_len = MIN(pin.send_len, ofconn->miss_send_len);
+
+ /* Figure out how much of the packet to send.
+ * If not buffered, send the entire packet. Otherwise, depending on
+ * the reason of packet-in, send what requested by the controller. */
+ if (pin.up.buffer_id != UINT32_MAX
+ && controller_max_len < pin.up.packet_len) {
+ pin.up.packet_len = controller_max_len;
}
- /* Make OFPT_PACKET_IN and hand over to packet scheduler. It might
- * immediately call into do_send_packet_in() or it might buffer it for a
- * while (until a later call to pinsched_run()). */
- pinsched_send(ofconn->schedulers[pin.reason == OFPR_NO_MATCH ? 0 : 1],
- pin.fmd.in_port,
- ofputil_encode_packet_in(&pin, ofconn_get_protocol(ofconn),
+ /* Make OFPT_PACKET_IN and hand over to packet scheduler. */
+ pinsched_send(ofconn->schedulers[pin.up.reason == OFPR_NO_MATCH ? 0 : 1],
+ pin.up.fmd.in_port,
+ ofputil_encode_packet_in(&pin.up,
+ ofconn_get_protocol(ofconn),
ofconn->packet_in_format),
- do_send_packet_in, ofconn);
+ &txq);
+ do_send_packet_ins(ofconn, &txq);
}
\f
/* Fail-open settings. */
ofpact_pad(&ofpacts);
match_init_catchall(&match);
- ofproto_add_flow(mgr->ofproto, &match, 0, ofpacts.data, ofpacts.size);
+ ofproto_add_flow(mgr->ofproto, &match, 0, ofpbuf_data(&ofpacts),
+ ofpbuf_size(&ofpacts));
ofpbuf_uninit(&ofpacts);
}
enum ofperr
ofmonitor_create(const struct ofputil_flow_monitor_request *request,
struct ofconn *ofconn, struct ofmonitor **monitorp)
+ OVS_REQUIRES(ofproto_mutex)
{
struct ofmonitor *m;
struct ofmonitor *
ofmonitor_lookup(struct ofconn *ofconn, uint32_t id)
+ OVS_REQUIRES(ofproto_mutex)
{
struct ofmonitor *m;
void
ofmonitor_destroy(struct ofmonitor *m)
+ OVS_REQUIRES(ofproto_mutex)
{
if (m) {
minimatch_destroy(&m->match);
default:
case NXFME_ABBREV:
- NOT_REACHED();
+ OVS_NOT_REACHED();
}
LIST_FOR_EACH (ofconn, node, &mgr->all_conns) {
ovs_mutex_unlock(&rule->mutex);
if (flags & NXFMF_ACTIONS) {
- fu.ofpacts = rule->actions->ofpacts;
- fu.ofpacts_len = rule->actions->ofpacts_len;
+ struct rule_actions *actions = rule_get_actions(rule);
+ fu.ofpacts = actions->ofpacts;
+ fu.ofpacts_len = actions->ofpacts_len;
} else {
fu.ofpacts = NULL;
fu.ofpacts_len = 0;
void
ofmonitor_flush(struct connmgr *mgr)
+ OVS_REQUIRES(ofproto_mutex)
{
struct ofconn *ofconn;
static void
ofmonitor_resume(struct ofconn *ofconn)
+ OVS_REQUIRES(ofproto_mutex)
{
struct rule_collection rules;
struct ofpbuf *resumed;
ofconn->monitor_paused = 0;
}
+static bool
+ofmonitor_may_resume(const struct ofconn *ofconn)
+ OVS_REQUIRES(ofproto_mutex)
+{
+ return (ofconn->monitor_paused != 0
+ && !rconn_packet_counter_n_packets(ofconn->monitor_counter));
+}
+
static void
ofmonitor_run(struct connmgr *mgr)
{
struct ofconn *ofconn;
+ ovs_mutex_lock(&ofproto_mutex);
LIST_FOR_EACH (ofconn, node, &mgr->all_conns) {
- if (ofconn->monitor_paused
- && !rconn_packet_counter_n_packets(ofconn->monitor_counter)) {
+ if (ofmonitor_may_resume(ofconn)) {
COVERAGE_INC(ofmonitor_resume);
ofmonitor_resume(ofconn);
}
}
+ ovs_mutex_unlock(&ofproto_mutex);
}
static void
{
struct ofconn *ofconn;
+ ovs_mutex_lock(&ofproto_mutex);
LIST_FOR_EACH (ofconn, node, &mgr->all_conns) {
- if (ofconn->monitor_paused
- && !rconn_packet_counter_n_packets(ofconn->monitor_counter)) {
+ if (ofmonitor_may_resume(ofconn)) {
poll_immediate_wake();
}
}
+ ovs_mutex_unlock(&ofproto_mutex);
}