#include "ofproto/ofproto-dpif-xlate.h"
+#include <errno.h>
+
#include "bfd.h"
#include "bitmap.h"
#include "bond.h"
* flow translation. */
#define MAX_RESUBMIT_RECURSION 64
+struct ovs_rwlock xlate_rwlock = OVS_RWLOCK_INITIALIZER;
+
struct xbridge {
struct hmap_node hmap_node; /* Node in global 'xbridges' map. */
struct ofproto_dpif *ofproto; /* Key in global 'xbridges' map. */
struct dpif_ipfix *ipfix; /* Ipfix handle, or null. */
struct stp *stp; /* STP or null if disabled. */
+ /* Special rules installed by ofproto-dpif. */
+ struct rule_dpif *miss_rule;
+ struct rule_dpif *no_packet_in_rule;
+
enum ofp_config_flags frag; /* Fragmentation handling. */
bool has_netflow; /* Bridge runs netflow? */
bool has_in_band; /* Bridge has in band control? */
uint16_t vlan);
static void compose_output_action(struct xlate_ctx *, ofp_port_t ofp_port);
-static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
-
static struct xbridge *xbridge_lookup(const struct ofproto_dpif *);
static struct xbundle *xbundle_lookup(const struct ofbundle *);
-static struct xport *xport_lookup(struct ofport_dpif *);
+static struct xport *xport_lookup(const struct ofport_dpif *);
static struct xport *get_ofp_port(const struct xbridge *, ofp_port_t ofp_port);
static struct skb_priority_to_dscp *get_skb_priority(const struct xport *,
uint32_t skb_priority);
void
xlate_ofproto_set(struct ofproto_dpif *ofproto, const char *name,
- struct dpif *dpif, const struct mac_learning *ml,
- struct stp *stp, const struct mbridge *mbridge,
+ struct dpif *dpif, struct rule_dpif *miss_rule,
+ struct rule_dpif *no_packet_in_rule,
+ const struct mac_learning *ml, struct stp *stp,
+ const struct mbridge *mbridge,
const struct dpif_sflow *sflow,
const struct dpif_ipfix *ipfix, enum ofp_config_flags frag,
bool forward_bpdu, bool has_in_band, bool has_netflow)
xbridge->has_in_band = has_in_band;
xbridge->has_netflow = has_netflow;
xbridge->frag = frag;
+ xbridge->miss_rule = miss_rule;
+ xbridge->no_packet_in_rule = no_packet_in_rule;
}
void
if (xport->peer) {
xport->peer->peer = NULL;
}
- xport->peer = peer ? xport_lookup(peer) : NULL;
+ xport->peer = xport_lookup(peer);
if (xport->peer) {
xport->peer->peer = xport;
}
if (xport->xbundle) {
list_remove(&xport->bundle_node);
}
- xport->xbundle = ofbundle ? xbundle_lookup(ofbundle) : NULL;
+ xport->xbundle = xbundle_lookup(ofbundle);
if (xport->xbundle) {
list_insert(&xport->xbundle->xports, &xport->bundle_node);
}
free(xport);
}
+/* Given a datpath, packet, and flow metadata ('backer', 'packet', and 'key'
+ * respectively), populates 'flow' with the result of odp_flow_key_to_flow().
+ * Optionally, if nonnull, populates 'fitnessp' with the fitness of 'flow' as
+ * returned by odp_flow_key_to_flow(). Also, optionally populates 'ofproto'
+ * with the ofproto_dpif, and 'odp_in_port' with the datapath in_port, that
+ * 'packet' ingressed.
+ *
+ * If 'ofproto' is nonnull, requires 'flow''s in_port to exist. Otherwise sets
+ * 'flow''s in_port to OFPP_NONE.
+ *
+ * This function does post-processing on data returned from
+ * odp_flow_key_to_flow() to help make VLAN splinters transparent to the rest
+ * of the upcall processing logic. In particular, if the extracted in_port is
+ * a VLAN splinter port, it replaces flow->in_port by the "real" port, sets
+ * flow->vlan_tci correctly for the VLAN of the VLAN splinter port, and pushes
+ * a VLAN header onto 'packet' (if it is nonnull).
+ *
+ * Similarly, this function also includes some logic to help with tunnels. It
+ * may modify 'flow' as necessary to make the tunneling implementation
+ * transparent to the upcall processing logic.
+ *
+ * Returns 0 if successful, ENODEV if the parsed flow has no associated ofport,
+ * or some other positive errno if there are other problems. */
+int
+xlate_receive(const struct dpif_backer *backer, struct ofpbuf *packet,
+ const struct nlattr *key, size_t key_len,
+ struct flow *flow, enum odp_key_fitness *fitnessp,
+ struct ofproto_dpif **ofproto, odp_port_t *odp_in_port)
+{
+ enum odp_key_fitness fitness;
+ const struct xport *xport;
+ int error = ENODEV;
+
+ ovs_rwlock_rdlock(&xlate_rwlock);
+ fitness = odp_flow_key_to_flow(key, key_len, flow);
+ if (fitness == ODP_FIT_ERROR) {
+ error = EINVAL;
+ goto exit;
+ }
+
+ if (odp_in_port) {
+ *odp_in_port = flow->in_port.odp_port;
+ }
+
+ xport = xport_lookup(tnl_port_should_receive(flow)
+ ? tnl_port_receive(flow)
+ : odp_port_to_ofport(backer, flow->in_port.odp_port));
+
+ flow->in_port.ofp_port = xport ? xport->ofp_port : OFPP_NONE;
+ if (!xport) {
+ goto exit;
+ }
+
+ if (vsp_adjust_flow(xport->xbridge->ofproto, flow)) {
+ if (packet) {
+ /* Make the packet resemble the flow, so that it gets sent to
+ * an OpenFlow controller properly, so that it looks correct
+ * for sFlow, and so that flow_extract() will get the correct
+ * vlan_tci if it is called on 'packet'.
+ *
+ * The allocated space inside 'packet' probably also contains
+ * 'key', that is, both 'packet' and 'key' are probably part of
+ * a struct dpif_upcall (see the large comment on that
+ * structure definition), so pushing data on 'packet' is in
+ * general not a good idea since it could overwrite 'key' or
+ * free it as a side effect. However, it's OK in this special
+ * case because we know that 'packet' is inside a Netlink
+ * attribute: pushing 4 bytes will just overwrite the 4-byte
+ * "struct nlattr", which is fine since we don't need that
+ * header anymore. */
+ eth_push_vlan(packet, flow->vlan_tci);
+ }
+ /* We can't reproduce 'key' from 'flow'. */
+ fitness = fitness == ODP_FIT_PERFECT ? ODP_FIT_TOO_MUCH : fitness;
+ }
+ error = 0;
+
+ if (ofproto) {
+ *ofproto = xport->xbridge->ofproto;
+ }
+
+exit:
+ if (fitnessp) {
+ *fitnessp = fitness;
+ }
+ ovs_rwlock_unlock(&xlate_rwlock);
+ return error;
+}
+
static struct xbridge *
xbridge_lookup(const struct ofproto_dpif *ofproto)
{
struct xbridge *xbridge;
+ if (!ofproto) {
+ return NULL;
+ }
+
HMAP_FOR_EACH_IN_BUCKET (xbridge, hmap_node, hash_pointer(ofproto, 0),
&xbridges) {
if (xbridge->ofproto == ofproto) {
{
struct xbundle *xbundle;
+ if (!ofbundle) {
+ return NULL;
+ }
+
HMAP_FOR_EACH_IN_BUCKET (xbundle, hmap_node, hash_pointer(ofbundle, 0),
&xbundles) {
if (xbundle->ofbundle == ofbundle) {
}
static struct xport *
-xport_lookup(struct ofport_dpif *ofport)
+xport_lookup(const struct ofport_dpif *ofport)
{
struct xport *xport;
+ if (!ofport) {
+ return NULL;
+ }
+
HMAP_FOR_EACH_IN_BUCKET (xport, hmap_node, hash_pointer(ofport, 0),
&xports) {
if (xport->ofport == ofport) {
ofport = bond_choose_output_slave(out_xbundle->bond, &ctx->xin->flow,
&ctx->xout->wc, vid);
- xport = ofport ? xport_lookup(ofport) : NULL;
+ xport = xport_lookup(ofport);
if (!xport) {
/* No slaves enabled, so drop packet. */
}
}
-static void
-update_learning_table(const struct xbridge *xbridge,
- const struct flow *flow, struct flow_wildcards *wc,
- int vlan, struct xbundle *in_xbundle)
+/* Checks whether a MAC learning update is necessary for MAC learning table
+ * 'ml' given that a packet matching 'flow' was received on 'in_xbundle' in
+ * 'vlan'.
+ *
+ * Most packets processed through the MAC learning table do not actually
+ * change it in any way. This function requires only a read lock on the MAC
+ * learning table, so it is much cheaper in this common case.
+ *
+ * Keep the code here synchronized with that in update_learning_table__()
+ * below. */
+static bool
+is_mac_learning_update_needed(const struct mac_learning *ml,
+ const struct flow *flow,
+ struct flow_wildcards *wc,
+ int vlan, struct xbundle *in_xbundle)
+ OVS_REQ_RDLOCK(ml->rwlock)
{
struct mac_entry *mac;
- /* Don't learn the OFPP_NONE port. */
- if (in_xbundle == &ofpp_none_bundle) {
- return;
+ if (!mac_learning_may_learn(ml, flow->dl_src, vlan)) {
+ return false;
}
- ovs_rwlock_wrlock(&xbridge->ml->rwlock);
+ mac = mac_learning_lookup(ml, flow->dl_src, vlan);
+ if (!mac || mac_entry_age(ml, mac)) {
+ return true;
+ }
+
+ if (is_gratuitous_arp(flow, wc)) {
+ /* We don't want to learn from gratuitous ARP packets that are
+ * reflected back over bond slaves so we lock the learning table. */
+ if (!in_xbundle->bond) {
+ return true;
+ } else if (mac_entry_is_grat_arp_locked(mac)) {
+ return false;
+ }
+ }
+
+ return mac->port.p != in_xbundle->ofbundle;
+}
+
+
+/* Updates MAC learning table 'ml' given that a packet matching 'flow' was
+ * received on 'in_xbundle' in 'vlan'.
+ *
+ * This code repeats all the checks in is_mac_learning_update_needed() because
+ * the lock was released between there and here and thus the MAC learning state
+ * could have changed.
+ *
+ * Keep the code here synchronized with that in is_mac_learning_update_needed()
+ * above. */
+static void
+update_learning_table__(const struct xbridge *xbridge,
+ const struct flow *flow, struct flow_wildcards *wc,
+ int vlan, struct xbundle *in_xbundle)
+ OVS_REQ_WRLOCK(xbridge->ml->rwlock)
+{
+ struct mac_entry *mac;
+
if (!mac_learning_may_learn(xbridge->ml, flow->dl_src, vlan)) {
- goto out;
+ return;
}
mac = mac_learning_insert(xbridge->ml, flow->dl_src, vlan);
if (!in_xbundle->bond) {
mac_entry_set_grat_arp_lock(mac);
} else if (mac_entry_is_grat_arp_locked(mac)) {
- goto out;
+ return;
}
}
/* The log messages here could actually be useful in debugging,
* so keep the rate limit relatively high. */
static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(30, 300);
+
VLOG_DBG_RL(&rl, "bridge %s: learned that "ETH_ADDR_FMT" is "
"on port %s in VLAN %d",
xbridge->name, ETH_ADDR_ARGS(flow->dl_src),
mac->port.p = in_xbundle->ofbundle;
mac_learning_changed(xbridge->ml);
}
-out:
+}
+
+static void
+update_learning_table(const struct xbridge *xbridge,
+ const struct flow *flow, struct flow_wildcards *wc,
+ int vlan, struct xbundle *in_xbundle)
+{
+ bool need_update;
+
+ /* Don't learn the OFPP_NONE port. */
+ if (in_xbundle == &ofpp_none_bundle) {
+ return;
+ }
+
+ /* First try the common case: no change to MAC learning table. */
+ ovs_rwlock_rdlock(&xbridge->ml->rwlock);
+ need_update = is_mac_learning_update_needed(xbridge->ml, flow, wc, vlan,
+ in_xbundle);
ovs_rwlock_unlock(&xbridge->ml->rwlock);
+
+ if (need_update) {
+ /* Slow path: MAC learning table might need an update. */
+ ovs_rwlock_wrlock(&xbridge->ml->rwlock);
+ update_learning_table__(xbridge, flow, wc, vlan, in_xbundle);
+ ovs_rwlock_unlock(&xbridge->ml->rwlock);
+ }
}
/* Determines whether packets in 'flow' within 'xbridge' should be forwarded or
struct flow_wildcards *wc = &ctx->xout->wc;
struct flow *flow = &ctx->xin->flow;
ovs_be16 flow_vlan_tci;
- uint32_t flow_skb_mark;
+ uint32_t flow_pkt_mark;
uint8_t flow_nw_tos;
odp_port_t out_port, odp_port;
uint8_t dscp;
}
ctx->xin->flow = old_flow;
- ctx->xbridge = xport->xbundle->xbridge;
+ ctx->xbridge = xport->xbridge;
if (ctx->xin->resubmit_stats) {
netdev_vport_inc_tx(xport->netdev, ctx->xin->resubmit_stats);
}
flow_vlan_tci = flow->vlan_tci;
- flow_skb_mark = flow->skb_mark;
+ flow_pkt_mark = flow->pkt_mark;
flow_nw_tos = flow->nw_tos;
if (dscp_from_skb_priority(xport, flow->skb_priority, &dscp)) {
out_port = ofp_port_to_odp_port(ctx->xbridge, vlandev_port);
flow->vlan_tci = htons(0);
}
- flow->skb_mark &= ~IPSEC_MARK;
}
if (out_port != ODPP_NONE) {
out:
/* Restore flow */
flow->vlan_tci = flow_vlan_tci;
- flow->skb_mark = flow_skb_mark;
+ flow->pkt_mark = flow_pkt_mark;
flow->nw_tos = flow_nw_tos;
}
compose_output_action__(ctx, ofp_port, true);
}
-/* Common rule processing in one place to avoid duplicating code. */
-static struct rule_dpif *
-ctx_rule_hooks(struct xlate_ctx *ctx, struct rule_dpif *rule,
- bool may_packet_in)
-{
- if (ctx->xin->resubmit_hook) {
- ctx->xin->resubmit_hook(ctx->xin, rule, ctx->recurse);
- }
- if (rule == NULL && may_packet_in) {
- /* XXX
- * check if table configuration flags
- * OFPTC_TABLE_MISS_CONTROLLER, default.
- * OFPTC_TABLE_MISS_CONTINUE,
- * OFPTC_TABLE_MISS_DROP
- * When OF1.0, OFPTC_TABLE_MISS_CONTINUE is used. What to do?
- */
- rule = rule_dpif_miss_rule(ctx->xbridge->ofproto, &ctx->xin->flow);
- }
- if (rule && ctx->xin->resubmit_stats) {
- rule_credit_stats(rule, ctx->xin->resubmit_stats);
- }
- return rule;
-}
-
static void
xlate_table_action(struct xlate_ctx *ctx,
ofp_port_t in_port, uint8_t table_id, bool may_packet_in)
/* Look up a flow with 'in_port' as the input port. */
ctx->xin->flow.in_port.ofp_port = in_port;
- rule = rule_dpif_lookup_in_table(ctx->xbridge->ofproto,
- &ctx->xin->flow, &ctx->xout->wc,
- table_id);
+ rule_dpif_lookup_in_table(ctx->xbridge->ofproto, &ctx->xin->flow,
+ &ctx->xout->wc, table_id, &rule);
/* Restore the original input port. Otherwise OFPP_NORMAL and
* OFPP_IN_PORT will have surprising behavior. */
ctx->xin->flow.in_port.ofp_port = old_in_port;
- rule = ctx_rule_hooks(ctx, rule, may_packet_in);
+ if (ctx->xin->resubmit_hook) {
+ ctx->xin->resubmit_hook(ctx->xin, rule, ctx->recurse);
+ }
+
+ if (rule == NULL && may_packet_in) {
+ struct xport *xport;
+
+ /* Makes clang's thread safety analysis happy. */
+ rule_release(rule);
+
+ /* XXX
+ * check if table configuration flags
+ * OFPTC_TABLE_MISS_CONTROLLER, default.
+ * OFPTC_TABLE_MISS_CONTINUE,
+ * OFPTC_TABLE_MISS_DROP
+ * When OF1.0, OFPTC_TABLE_MISS_CONTINUE is used. What to do? */
+ xport = get_ofp_port(ctx->xbridge, ctx->xin->flow.in_port.ofp_port);
+ rule = choose_miss_rule(xport ? xport->config : 0,
+ ctx->xbridge->miss_rule,
+ ctx->xbridge->no_packet_in_rule);
+ ovs_rwlock_rdlock(&rule->up.evict);
+ }
+
+ if (rule && ctx->xin->resubmit_stats) {
+ rule_credit_stats(rule, ctx->xin->resubmit_stats);
+ }
if (rule) {
struct rule_dpif *old_rule = ctx->rule;
ctx->rule = old_rule;
ctx->recurse--;
}
+ rule_release(rule);
ctx->table_id = old_table_id;
} else {
enum ofp_packet_in_reason reason,
uint16_t controller_id)
{
- struct ofputil_packet_in pin;
+ struct ofputil_packet_in *pin;
struct ofpbuf *packet;
struct flow key;
packet = ofpbuf_clone(ctx->xin->packet);
key.skb_priority = 0;
- key.skb_mark = 0;
+ key.pkt_mark = 0;
memset(&key.tunnel, 0, sizeof key.tunnel);
commit_odp_actions(&ctx->xin->flow, &ctx->base_flow,
odp_execute_actions(NULL, packet, &key, ctx->xout->odp_actions.data,
ctx->xout->odp_actions.size, NULL, NULL);
- pin.packet = packet->data;
- pin.packet_len = packet->size;
- pin.reason = reason;
- pin.controller_id = controller_id;
- pin.table_id = ctx->table_id;
- pin.cookie = ctx->rule ? ctx->rule->up.flow_cookie : 0;
+ pin = xmalloc(sizeof *pin);
+ pin->packet_len = packet->size;
+ pin->packet = ofpbuf_steal_data(packet);
+ pin->reason = reason;
+ pin->controller_id = controller_id;
+ pin->table_id = ctx->table_id;
+ pin->cookie = ctx->rule ? ctx->rule->up.flow_cookie : 0;
- pin.send_len = len;
- flow_get_metadata(&ctx->xin->flow, &pin.fmd);
+ pin->send_len = len;
+ flow_get_metadata(&ctx->xin->flow, &pin->fmd);
- ofproto_dpif_send_packet_in(ctx->xbridge->ofproto, &pin);
+ ofproto_dpif_send_packet_in(ctx->xbridge->ofproto, pin);
ofpbuf_delete(packet);
}
if (ctx->xin->tcp_flags & (TCP_FIN | TCP_RST) && ctx->rule) {
struct rule_dpif *rule = ctx->rule;
- if (list_is_empty(&rule->up.expirable)) {
- list_insert(&rule->up.ofproto->expirable, &rule->up.expirable);
- }
+ ovs_mutex_lock(&rule->up.ofproto->expirable_mutex);
+ if (list_is_empty(&rule->up.expirable)) {
+ list_insert(&rule->up.ofproto->expirable, &rule->up.expirable);
+ }
+ ovs_mutex_unlock(&rule->up.ofproto->expirable_mutex);
+ ovs_mutex_lock(&rule->up.timeout_mutex);
reduce_timeout(oft->fin_idle_timeout, &rule->up.idle_timeout);
reduce_timeout(oft->fin_hard_timeout, &rule->up.hard_timeout);
+ ovs_mutex_unlock(&rule->up.timeout_mutex);
}
}
return true;
}
-static bool
-tunnel_ecn_ok(struct xlate_ctx *ctx)
-{
- if (is_ip_any(&ctx->base_flow)
- && (ctx->xin->flow.tunnel.ip_tos & IP_ECN_MASK) == IP_ECN_CE) {
- if ((ctx->base_flow.nw_tos & IP_ECN_MASK) == IP_ECN_NOT_ECT) {
- VLOG_WARN_RL(&rl, "dropping tunnel packet marked ECN CE"
- " but is not ECN capable");
- return false;
- } else {
- /* Set the ECN CE value in the tunneled packet. */
- ctx->xin->flow.nw_tos |= IP_ECN_CE;
- }
- }
-
- return true;
-}
-
static void
do_xlate_actions(const struct ofpact *ofpacts, size_t ofpacts_len,
struct xlate_ctx *ctx)
{
struct flow_wildcards *wc = &ctx->xout->wc;
struct flow *flow = &ctx->xin->flow;
- bool was_evictable = true;
const struct ofpact *a;
- if (ctx->rule) {
- /* Don't let the rule we're working on get evicted underneath us. */
- was_evictable = ctx->rule->up.evictable;
- ctx->rule->up.evictable = false;
- }
-
OFPACT_FOR_EACH (a, ofpacts, ofpacts_len) {
struct ofpact_controller *controller;
const struct ofpact_metadata *metadata;
case OFPACT_SET_MPLS_TTL:
if (compose_set_mpls_ttl_action(ctx,
ofpact_get_SET_MPLS_TTL(a)->ttl)) {
- goto out;
+ return;
}
break;
case OFPACT_DEC_MPLS_TTL:
if (compose_dec_mpls_ttl_action(ctx)) {
- goto out;
+ return;
}
break;
case OFPACT_DEC_TTL:
wc->masks.nw_ttl = 0xff;
if (compose_dec_ttl(ctx, ofpact_get_DEC_TTL(a))) {
- goto out;
+ return;
}
break;
break;
}
}
-
-out:
- if (ctx->rule) {
- ctx->rule->up.evictable = was_evictable;
- }
}
void
struct flow orig_flow;
struct xlate_ctx ctx;
size_t ofpacts_len;
+ bool tnl_may_send;
COVERAGE_INC(xlate_actions);
+ ovs_rwlock_rdlock(&xlate_rwlock);
+
/* Flow initialization rules:
* - 'base_flow' must match the kernel's view of the packet at the
* time that action processing starts. 'flow' represents any
ctx.xbridge = xbridge_lookup(xin->ofproto);
if (!ctx.xbridge) {
- return;
+ goto out;
}
ctx.rule = xin->rule;
memset(&wc->masks.dl_type, 0xff, sizeof wc->masks.dl_type);
wc->masks.nw_frag |= FLOW_NW_FRAG_MASK;
- if (tnl_port_should_receive(&ctx.xin->flow)) {
- memset(&wc->masks.tunnel, 0xff, sizeof wc->masks.tunnel);
- /* skb_mark is currently used only by tunnels but that will likely
- * change in the future. */
- memset(&wc->masks.skb_mark, 0xff, sizeof wc->masks.skb_mark);
- }
+ tnl_may_send = tnl_xlate_init(&ctx.base_flow, flow, wc);
if (ctx.xbridge->has_netflow) {
netflow_mask_wc(flow, wc);
}
break;
case OFPC_FRAG_DROP:
- return;
+ goto out;
case OFPC_FRAG_REASM:
NOT_REACHED();
add_ipfix_action(&ctx);
sample_actions_len = ctx.xout->odp_actions.size;
- if (tunnel_ecn_ok(&ctx) && (!in_port || may_receive(in_port, &ctx))) {
+ if (tnl_may_send && (!in_port || may_receive(in_port, &ctx))) {
do_xlate_actions(ofpacts, ofpacts_len, &ctx);
/* We've let OFPP_NORMAL and the learning action look at the
* use non-header fields as part of the cache. */
memset(&wc->masks.metadata, 0, sizeof wc->masks.metadata);
memset(&wc->masks.regs, 0, sizeof wc->masks.regs);
+
+out:
+ ovs_rwlock_unlock(&xlate_rwlock);
}