#include "ofproto-dpif-ipfix.h"
#include "ofproto-dpif-mirror.h"
#include "ofproto-dpif-sflow.h"
+#include "ofproto-dpif-upcall.h"
#include "ofproto-dpif-xlate.h"
#include "poll-loop.h"
#include "simap.h"
COVERAGE_DEFINE(facet_unexpected);
COVERAGE_DEFINE(facet_suppress);
COVERAGE_DEFINE(subfacet_install_fail);
+COVERAGE_DEFINE(packet_in_overflow);
COVERAGE_DEFINE(flow_mod_overflow);
+#define N_THREADS 16
+
/* Number of implemented OpenFlow tables. */
enum { N_TABLES = 255 };
enum { TBL_INTERNAL = N_TABLES - 1 }; /* Used for internal hidden rules. */
struct flow_miss;
struct facet;
-static struct rule_dpif *rule_dpif_lookup(struct ofproto_dpif *,
- const struct flow *,
- struct flow_wildcards *wc);
-
static void rule_get_stats(struct rule *, uint64_t *packets, uint64_t *bytes);
struct ofbundle {
#define SUBFACET_DESTROY_MAX_BATCH 50
-static struct subfacet *subfacet_create(struct facet *, struct flow_miss *miss,
- long long int now);
+static struct subfacet *subfacet_create(struct facet *, struct flow_miss *);
static struct subfacet *subfacet_find(struct dpif_backer *,
const struct nlattr *key, size_t key_len,
uint32_t key_hash);
uint8_t tcp_flags; /* TCP flags seen for this 'rule'. */
struct xlate_out xout;
- bool fail_open; /* Facet matched the fail open rule. */
/* Storage for a single subfacet, to reduce malloc() time and space
* overhead. (A facet always has at least one subfacet and in the common
long long int learn_rl; /* Rate limiter for facet_learn(). */
};
-static struct facet *facet_create(const struct flow_miss *, struct rule_dpif *,
- struct xlate_out *,
- struct dpif_flow_stats *);
+static struct facet *facet_create(const struct flow_miss *);
static void facet_remove(struct facet *);
static void facet_free(struct facet *);
static void facet_flush_stats(struct facet *);
static void facet_reset_counters(struct facet *);
+static void flow_push_stats(struct ofproto_dpif *, struct flow *,
+ struct dpif_flow_stats *, bool may_learn);
static void facet_push_stats(struct facet *, bool may_learn);
static void facet_learn(struct facet *);
static void facet_account(struct facet *);
int vid;
};
-static bool vsp_adjust_flow(const struct ofproto_dpif *, struct flow *);
static void vsp_remove(struct ofport_dpif *);
static void vsp_add(struct ofport_dpif *, ofp_port_t realdev_ofp_port, int vid);
COVERAGE_DEFINE(rev_mac_learning);
COVERAGE_DEFINE(rev_inconsistency);
-/* Drop keys are odp flow keys which have drop flows installed in the kernel.
- * These are datapath flows which have no associated ofproto, if they did we
- * would use facets. */
-struct drop_key {
- struct hmap_node hmap_node;
- struct nlattr *key;
- size_t key_len;
-};
-
struct avg_subfacet_rates {
double add_rate; /* Moving average of new flows created per minute. */
double del_rate; /* Moving average of flows deleted per minute. */
char *type;
int refcount;
struct dpif *dpif;
+ struct udpif *udpif;
struct timer next_expiration;
- struct hmap odp_to_ofport_map; /* ODP port to ofport mapping. */
+
+ struct ovs_rwlock odp_to_ofport_lock;
+ struct hmap odp_to_ofport_map OVS_GUARDED; /* ODP port to ofport map. */
struct simap tnl_backers; /* Set of dpif ports backing tunnels. */
static struct shash all_dpif_backers = SHASH_INITIALIZER(&all_dpif_backers);
static void drop_key_clear(struct dpif_backer *);
-static struct ofport_dpif *
-odp_port_to_ofport(const struct dpif_backer *, odp_port_t odp_port);
static void update_moving_averages(struct dpif_backer *backer);
struct ofproto_dpif {
long long int stp_last_tick;
/* VLAN splinters. */
- struct hmap realdev_vid_map; /* (realdev,vid) -> vlandev. */
- struct hmap vlandev_map; /* vlandev -> (realdev,vid). */
+ struct ovs_mutex vsp_mutex;
+ struct hmap realdev_vid_map OVS_GUARDED; /* (realdev,vid) -> vlandev. */
+ struct hmap vlandev_map OVS_GUARDED; /* vlandev -> (realdev,vid). */
/* Ports. */
struct sset ports; /* Set of standard port names. */
struct ovs_mutex flow_mod_mutex;
struct list flow_mods OVS_GUARDED;
size_t n_flow_mods OVS_GUARDED;
+
+ struct ovs_mutex pin_mutex;
+ struct list pins OVS_GUARDED;
+ size_t n_pins OVS_GUARDED;
};
/* Defer flow mod completion until "ovs-appctl ofproto/unclog"? (Useful only
const struct ofpbuf *packet, struct ds *);
/* Upcalls. */
-#define FLOW_MISS_MAX_BATCH 50
-static int handle_upcalls(struct dpif_backer *, unsigned int max_batch);
+static void handle_upcalls(struct dpif_backer *);
/* Flow expiration. */
static int expire(struct dpif_backer *);
ofproto_dpif_send_packet_in(struct ofproto_dpif *ofproto,
struct ofputil_packet_in *pin)
{
- connmgr_send_packet_in(ofproto->up.connmgr, pin);
+ ovs_mutex_lock(&ofproto->pin_mutex);
+ if (ofproto->n_pins > 1024) {
+ ovs_mutex_unlock(&ofproto->pin_mutex);
+ COVERAGE_INC(packet_in_overflow);
+ free(CONST_CAST(void *, pin->packet));
+ free(pin);
+ return;
+ }
+
+ list_push_back(&ofproto->pins, &pin->list_node);
+ ofproto->n_pins++;
+ ovs_mutex_unlock(&ofproto->pin_mutex);
}
\f
/* Factory functions. */
error = dpif_recv_set(backer->dpif, backer->recv_set_enable);
if (error) {
+ udpif_recv_set(backer->udpif, 0, false);
VLOG_ERR("Failed to enable receiving packets in dpif.");
return error;
}
+ udpif_recv_set(backer->udpif, N_THREADS, backer->recv_set_enable);
dpif_flow_flush(backer->dpif);
backer->need_revalidate = REV_RECONFIGURE;
}
continue;
}
+ ovs_rwlock_wrlock(&xlate_rwlock);
xlate_ofproto_set(ofproto, ofproto->up.name,
- ofproto->backer->dpif, ofproto->ml,
+ ofproto->backer->dpif, ofproto->miss_rule,
+ ofproto->no_packet_in_rule, ofproto->ml,
ofproto->stp, ofproto->mbridge,
ofproto->sflow, ofproto->ipfix,
ofproto->up.frag_handling,
ofport->up.pp.config, ofport->is_tunnel,
ofport->may_enable);
}
+ ovs_rwlock_unlock(&xlate_rwlock);
+ /* Only ofproto-dpif cares about the facet classifier so we just
+ * lock cls_cursor_init() to appease the thread safety analysis. */
+ ovs_rwlock_rdlock(&ofproto->facets.rwlock);
cls_cursor_init(&cursor, &ofproto->facets, NULL);
+ ovs_rwlock_unlock(&ofproto->facets.rwlock);
CLS_CURSOR_FOR_EACH_SAFE (facet, next, cr, &cursor) {
facet_revalidate(facet);
run_fast_rl();
}
}
+
+ udpif_revalidate(backer->udpif);
}
if (!backer->recv_set_enable) {
/* 'ofport''s datapath port number has changed from
* 'ofport->odp_port' to 'port.port_no'. Update our internal data
* structures to match. */
+ ovs_rwlock_wrlock(&backer->odp_to_ofport_lock);
hmap_remove(&backer->odp_to_ofport_map, &ofport->odp_port_node);
ofport->odp_port = port.port_no;
hmap_insert(&backer->odp_to_ofport_map, &ofport->odp_port_node,
hash_odp_port(port.port_no));
+ ovs_rwlock_unlock(&backer->odp_to_ofport_lock);
backer->need_revalidate = REV_RECONFIGURE;
}
}
}
static int
-dpif_backer_run_fast(struct dpif_backer *backer, int max_batch)
+dpif_backer_run_fast(struct dpif_backer *backer)
{
- unsigned int work;
-
- /* If recv_set_enable is false, we should not handle upcalls. */
- if (!backer->recv_set_enable) {
- return 0;
- }
-
- /* Handle one or more batches of upcalls, until there's nothing left to do
- * or until we do a fixed total amount of work.
- *
- * We do work in batches because it can be much cheaper to set up a number
- * of flows and fire off their patches all at once. We do multiple batches
- * because in some cases handling a packet can cause another packet to be
- * queued almost immediately as part of the return flow. Both
- * optimizations can make major improvements on some benchmarks and
- * presumably for real traffic as well. */
- work = 0;
- while (work < max_batch) {
- int retval = handle_upcalls(backer, max_batch - work);
- if (retval <= 0) {
- return -retval;
- }
- work += retval;
- }
+ udpif_run(backer->udpif);
+ handle_upcalls(backer);
return 0;
}
return 0;
}
- return dpif_backer_run_fast(backer, FLOW_MISS_MAX_BATCH);
+ return dpif_backer_run_fast(backer);
}
static void
run_fast_rl(void)
{
static long long int port_rl = LLONG_MIN;
- static unsigned int backer_rl = 0;
if (time_msec() >= port_rl) {
struct ofproto_dpif *ofproto;
}
port_rl = time_msec() + 200;
}
-
- /* XXX: We have to be careful not to do too much work in this function. If
- * we call dpif_backer_run_fast() too often, or with too large a batch,
- * performance improves signifcantly, but at a cost. It's possible for the
- * number of flows in the datapath to increase without bound, and for poll
- * loops to take 10s of seconds. The correct solution to this problem,
- * long term, is to separate flow miss handling into it's own thread so it
- * isn't affected by revalidations, and expirations. Until then, this is
- * the best we can do. */
- if (++backer_rl >= 10) {
- struct shash_node *node;
-
- backer_rl = 0;
- SHASH_FOR_EACH (node, &all_dpif_backers) {
- dpif_backer_run_fast(node->data, 1);
- }
- }
}
static void
hmap_destroy(&backer->drop_keys);
simap_destroy(&backer->tnl_backers);
+ ovs_rwlock_destroy(&backer->odp_to_ofport_lock);
hmap_destroy(&backer->odp_to_ofport_map);
node = shash_find(&all_dpif_backers, backer->type);
free(backer->type);
shash_delete(&all_dpif_backers, node);
+ udpif_destroy(backer->udpif);
dpif_close(backer->dpif);
ovs_assert(hmap_is_empty(&backer->subfacets));
free(backer);
return error;
}
+ backer->udpif = udpif_create(backer, backer->dpif);
backer->type = xstrdup(type);
backer->governor = NULL;
backer->refcount = 1;
hmap_init(&backer->odp_to_ofport_map);
+ ovs_rwlock_init(&backer->odp_to_ofport_lock);
hmap_init(&backer->drop_keys);
hmap_init(&backer->subfacets);
timer_set_duration(&backer->next_expiration, 1000);
close_dpif_backer(backer);
return error;
}
+ udpif_recv_set(backer->udpif, N_THREADS, backer->recv_set_enable);
backer->max_n_subfacet = 0;
backer->created = time_msec();
ofproto->ml = mac_learning_create(MAC_ENTRY_DEFAULT_IDLE_TIME);
ofproto->mbridge = mbridge_create();
ofproto->has_bonded_bundles = false;
+ ovs_mutex_init(&ofproto->vsp_mutex, PTHREAD_MUTEX_NORMAL);
classifier_init(&ofproto->facets);
ofproto->consistency_rl = LLONG_MIN;
ofproto->n_flow_mods = 0;
ovs_mutex_unlock(&ofproto->flow_mod_mutex);
+ ovs_mutex_init(&ofproto->pin_mutex, PTHREAD_MUTEX_NORMAL);
+ ovs_mutex_lock(&ofproto->pin_mutex);
+ list_init(&ofproto->pins);
+ ofproto->n_pins = 0;
+ ovs_mutex_unlock(&ofproto->pin_mutex);
+
ofproto_dpif_unixctl_init();
hmap_init(&ofproto->vlandev_map);
return error;
}
- *rulep = rule_dpif_lookup_in_table(ofproto, &fm.match.flow, NULL,
- TBL_INTERNAL);
- ovs_assert(*rulep != NULL);
+ if (rule_dpif_lookup_in_table(ofproto, &fm.match.flow, NULL, TBL_INTERNAL,
+ rulep)) {
+ ovs_rwlock_unlock(&(*rulep)->up.evict);
+ } else {
+ NOT_REACHED();
+ }
return 0;
}
{
struct ofproto_dpif *ofproto = ofproto_dpif_cast(ofproto_);
struct rule_dpif *rule, *next_rule;
+ struct ofputil_flow_mod *pin, *next_pin;
struct ofputil_flow_mod *fm, *next_fm;
struct oftable *table;
ofproto->backer->need_revalidate = REV_RECONFIGURE;
+ ovs_rwlock_wrlock(&xlate_rwlock);
xlate_remove_ofproto(ofproto);
+ ovs_rwlock_unlock(&xlate_rwlock);
hmap_remove(&all_ofproto_dpifs, &ofproto->all_ofproto_dpifs_node);
complete_operations(ofproto);
OFPROTO_FOR_EACH_TABLE (table, &ofproto->up) {
struct cls_cursor cursor;
+ ovs_rwlock_wrlock(&table->cls.rwlock);
cls_cursor_init(&cursor, &table->cls, NULL);
CLS_CURSOR_FOR_EACH_SAFE (rule, next_rule, up.cr, &cursor) {
- ofproto_rule_destroy(&rule->up);
+ ofproto_rule_destroy(&ofproto->up, &table->cls, &rule->up);
}
+ ovs_rwlock_unlock(&table->cls.rwlock);
}
ovs_mutex_lock(&ofproto->flow_mod_mutex);
ovs_mutex_unlock(&ofproto->flow_mod_mutex);
ovs_mutex_destroy(&ofproto->flow_mod_mutex);
+ ovs_mutex_lock(&ofproto->pin_mutex);
+ LIST_FOR_EACH_SAFE (pin, next_pin, list_node, &ofproto->pins) {
+ list_remove(&pin->list_node);
+ ofproto->n_pins--;
+ free(pin->ofpacts);
+ free(pin);
+ }
+ ovs_mutex_unlock(&ofproto->pin_mutex);
+ ovs_mutex_destroy(&ofproto->pin_mutex);
+
mbridge_unref(ofproto->mbridge);
netflow_destroy(ofproto->netflow);
sset_destroy(&ofproto->ghost_ports);
sset_destroy(&ofproto->port_poll_set);
+ ovs_mutex_destroy(&ofproto->vsp_mutex);
+
close_dpif_backer(ofproto->backer);
}
run_fast(struct ofproto *ofproto_)
{
struct ofproto_dpif *ofproto = ofproto_dpif_cast(ofproto_);
- struct ofputil_flow_mod *fm, *next;
+ struct ofputil_packet_in *pin, *next_pin;
+ struct ofputil_flow_mod *fm, *next_fm;
+ struct list flow_mods, pins;
struct ofport_dpif *ofport;
- struct list flow_mods;
/* Do not perform any periodic activity required by 'ofproto' while
* waiting for flow restore to complete. */
}
ovs_mutex_unlock(&ofproto->flow_mod_mutex);
- LIST_FOR_EACH_SAFE (fm, next, list_node, &flow_mods) {
+ LIST_FOR_EACH_SAFE (fm, next_fm, list_node, &flow_mods) {
int error = ofproto_flow_mod(&ofproto->up, fm);
if (error && !VLOG_DROP_WARN(&rl)) {
VLOG_WARN("learning action failed to modify flow table (%s)",
free(fm);
}
+ ovs_mutex_lock(&ofproto->pin_mutex);
+ if (ofproto->n_pins) {
+ pins = ofproto->pins;
+ list_moved(&pins);
+ list_init(&ofproto->pins);
+ ofproto->n_pins = 0;
+ } else {
+ list_init(&pins);
+ }
+ ovs_mutex_unlock(&ofproto->pin_mutex);
+
+ LIST_FOR_EACH_SAFE (pin, next_pin, list_node, &pins) {
+ connmgr_send_packet_in(ofproto->up.connmgr, pin);
+ list_remove(&pin->list_node);
+ free(CONST_CAST(void *, pin->packet));
+ free(pin);
+ }
+
HMAP_FOR_EACH (ofport, up.hmap_node, &ofproto->up.ports) {
port_run_fast(ofport);
}
ovs_rwlock_unlock(&ofproto->ml->rwlock);
/* Check the consistency of a random facet, to aid debugging. */
+ ovs_rwlock_rdlock(&ofproto->facets.rwlock);
if (time_msec() >= ofproto->consistency_rl
&& !classifier_is_empty(&ofproto->facets)
&& !ofproto->backer->need_revalidate) {
ofproto->backer->need_revalidate = REV_INCONSISTENCY;
}
}
+ ovs_rwlock_unlock(&ofproto->facets.rwlock);
return 0;
}
}
dpif_wait(ofproto->backer->dpif);
- dpif_recv_wait(ofproto->backer->dpif);
+ udpif_wait(ofproto->backer->udpif);
if (ofproto->sflow) {
dpif_sflow_wait(ofproto->sflow);
}
size_t n_subfacets = 0;
struct facet *facet;
+ ovs_rwlock_rdlock(&ofproto->facets.rwlock);
simap_increase(usage, "facets", classifier_count(&ofproto->facets));
+ ovs_rwlock_unlock(&ofproto->facets.rwlock);
+ ovs_rwlock_rdlock(&ofproto->facets.rwlock);
cls_cursor_init(&cursor, &ofproto->facets, NULL);
CLS_CURSOR_FOR_EACH (facet, cr, &cursor) {
n_subfacets += list_size(&facet->subfacets);
}
+ ovs_rwlock_unlock(&ofproto->facets.rwlock);
simap_increase(usage, "subfacets", n_subfacets);
}
return EBUSY;
}
+ ovs_rwlock_wrlock(&ofproto->backer->odp_to_ofport_lock);
hmap_insert(&ofproto->backer->odp_to_ofport_map, &port->odp_port_node,
hash_odp_port(port->odp_port));
+ ovs_rwlock_unlock(&ofproto->backer->odp_to_ofport_lock);
}
dpif_port_destroy(&dpif_port);
const char *dp_port_name;
ofproto->backer->need_revalidate = REV_RECONFIGURE;
+ ovs_rwlock_wrlock(&xlate_rwlock);
xlate_ofport_remove(port);
+ ovs_rwlock_unlock(&xlate_rwlock);
dp_port_name = netdev_vport_get_dpif_port(port->up.netdev, namebuf,
sizeof namebuf);
}
if (port->odp_port != ODPP_NONE && !port->is_tunnel) {
+ ovs_rwlock_wrlock(&ofproto->backer->odp_to_ofport_lock);
hmap_remove(&ofproto->backer->odp_to_ofport_map, &port->odp_port_node);
+ ovs_rwlock_unlock(&ofproto->backer->odp_to_ofport_lock);
}
tnl_port_del(port);
ofproto = bundle->ofproto;
mbridge_unregister_bundle(ofproto->mbridge, bundle->aux);
+ ovs_rwlock_wrlock(&xlate_rwlock);
xlate_bundle_remove(bundle);
+ ovs_rwlock_unlock(&xlate_rwlock);
LIST_FOR_EACH_SAFE (port, next_port, bundle_node, &bundle->ports) {
bundle_del_port(port);
{
const struct ofproto_dpif *ofproto;
struct dpif_backer *backer;
- const char *peer_name;
+ char *peer_name;
if (!netdev_vport_is_patch(ofport->up.netdev)) {
return;
HMAP_FOR_EACH (ofproto, all_ofproto_dpifs_node, &all_ofproto_dpifs) {
struct ofport *peer_ofport;
struct ofport_dpif *peer;
- const char *peer_peer;
+ char *peer_peer;
if (ofproto->backer != backer) {
continue;
ofport->peer = peer;
ofport->peer->peer = ofport;
}
+ free(peer_peer);
- return;
+ break;
}
+ free(peer_name);
}
static void
\f
/* Upcall handling. */
-/* Flow miss batching.
- *
- * Some dpifs implement operations faster when you hand them off in a batch.
- * To allow batching, "struct flow_miss" queues the dpif-related work needed
- * for a given flow. Each "struct flow_miss" corresponds to sending one or
- * more packets, plus possibly installing the flow in the dpif.
- *
- * So far we only batch the operations that affect flow setup time the most.
- * It's possible to batch more than that, but the benefit might be minimal. */
-struct flow_miss {
- struct hmap_node hmap_node;
- struct ofproto_dpif *ofproto;
- struct flow flow;
- enum odp_key_fitness key_fitness;
- const struct nlattr *key;
- size_t key_len;
- struct list packets;
- enum dpif_upcall_type upcall_type;
-};
-
struct flow_miss_op {
struct dpif_op dpif_op;
struct subfacet *subfacet;
};
-/* Sends an OFPT_PACKET_IN message for 'packet' of type OFPR_NO_MATCH to each
- * OpenFlow controller as necessary according to their individual
- * configurations. */
-static void
-send_packet_in_miss(struct ofproto_dpif *ofproto, const struct ofpbuf *packet,
- const struct flow *flow)
-{
- struct ofputil_packet_in pin;
-
- pin.packet = packet->data;
- pin.packet_len = packet->size;
- pin.reason = OFPR_NO_MATCH;
- pin.controller_id = 0;
-
- pin.table_id = 0;
- pin.cookie = 0;
-
- pin.send_len = 0; /* not used for flow table misses */
-
- flow_get_metadata(flow, &pin.fmd);
-
- connmgr_send_packet_in(ofproto->up.connmgr, &pin);
-}
-
-static struct flow_miss *
-flow_miss_find(struct hmap *todo, const struct ofproto_dpif *ofproto,
- const struct flow *flow, uint32_t hash)
-{
- struct flow_miss *miss;
-
- HMAP_FOR_EACH_WITH_HASH (miss, hmap_node, hash, todo) {
- if (miss->ofproto == ofproto && flow_equal(&miss->flow, flow)) {
- return miss;
- }
- }
-
- return NULL;
-}
-
-/* Partially Initializes 'op' as an "execute" operation for 'miss' and
- * 'packet'. The caller must initialize op->actions and op->actions_len. If
- * 'miss' is associated with a subfacet the caller must also initialize the
- * returned op->subfacet, and if anything needs to be freed after processing
- * the op, the caller must initialize op->garbage also. */
-static void
-init_flow_miss_execute_op(struct flow_miss *miss, struct ofpbuf *packet,
- struct flow_miss_op *op)
-{
- if (miss->flow.in_port.ofp_port
- != vsp_realdev_to_vlandev(miss->ofproto, miss->flow.in_port.ofp_port,
- miss->flow.vlan_tci)) {
- /* This packet was received on a VLAN splinter port. We
- * added a VLAN to the packet to make the packet resemble
- * the flow, but the actions were composed assuming that
- * the packet contained no VLAN. So, we must remove the
- * VLAN header from the packet before trying to execute the
- * actions. */
- eth_pop_vlan(packet);
- }
-
- op->subfacet = NULL;
- op->xout_garbage = false;
- op->dpif_op.type = DPIF_OP_EXECUTE;
- op->dpif_op.u.execute.key = miss->key;
- op->dpif_op.u.execute.key_len = miss->key_len;
- op->dpif_op.u.execute.packet = packet;
- ofpbuf_use_stack(&op->mask, &op->maskbuf, sizeof op->maskbuf);
-}
-
-/* Helper for handle_flow_miss_without_facet() and
- * handle_flow_miss_with_facet(). */
-static void
-handle_flow_miss_common(struct ofproto_dpif *ofproto, struct ofpbuf *packet,
- const struct flow *flow, bool fail_open)
-{
- if (fail_open) {
- /*
- * Extra-special case for fail-open mode.
- *
- * We are in fail-open mode and the packet matched the fail-open
- * rule, but we are connected to a controller too. We should send
- * the packet up to the controller in the hope that it will try to
- * set up a flow and thereby allow us to exit fail-open.
- *
- * See the top-level comment in fail-open.c for more information.
- */
- send_packet_in_miss(ofproto, packet, flow);
- }
-}
-
/* Figures out whether a flow that missed in 'ofproto', whose details are in
* 'miss' masked by 'wc', is likely to be worth tracking in detail in userspace
* and (usually) installing a datapath flow. The answer is usually "yes" (a
* flows we impose some heuristics to decide which flows are likely to be worth
* tracking. */
static bool
-flow_miss_should_make_facet(struct flow_miss *miss, struct flow_wildcards *wc)
+flow_miss_should_make_facet(struct flow_miss *miss)
{
struct dpif_backer *backer = miss->ofproto->backer;
uint32_t hash;
backer->governor = governor_create();
}
- hash = flow_hash_in_wildcards(&miss->flow, wc, 0);
+ hash = flow_hash_in_wildcards(&miss->flow, &miss->xout.wc, 0);
return governor_should_install_flow(backer->governor, hash,
list_size(&miss->packets));
}
-/* Handles 'miss' without creating a facet or subfacet or creating any datapath
- * flow. 'miss->flow' must have matched 'rule' and been xlated into 'xout'.
- * May add an "execute" operation to 'ops' and increment '*n_ops'. */
-static void
-handle_flow_miss_without_facet(struct rule_dpif *rule, struct xlate_out *xout,
- struct flow_miss *miss,
- struct flow_miss_op *ops, size_t *n_ops)
-{
- struct ofpbuf *packet;
-
- LIST_FOR_EACH (packet, list_node, &miss->packets) {
-
- COVERAGE_INC(facet_suppress);
-
- handle_flow_miss_common(miss->ofproto, packet, &miss->flow,
- rule->up.cr.priority == FAIL_OPEN_PRIORITY);
-
- if (xout->slow) {
- struct xlate_in xin;
-
- xlate_in_init(&xin, miss->ofproto, &miss->flow, rule, 0, packet);
- xlate_actions_for_side_effects(&xin);
- }
-
- if (xout->odp_actions.size) {
- struct flow_miss_op *op = &ops[*n_ops];
- struct dpif_execute *execute = &op->dpif_op.u.execute;
-
- init_flow_miss_execute_op(miss, packet, op);
- xlate_out_copy(&op->xout, xout);
- execute->actions = op->xout.odp_actions.data;
- execute->actions_len = op->xout.odp_actions.size;
- op->xout_garbage = true;
-
- (*n_ops)++;
- }
- }
-}
-
/* Handles 'miss', which matches 'facet'. May add any required datapath
* operations to 'ops', incrementing '*n_ops' for each new op.
*
- * All of the packets in 'miss' are considered to have arrived at time 'now'.
- * This is really important only for new facets: if we just called time_msec()
- * here, then the new subfacet or its packets could look (occasionally) as
- * though it was used some time after the facet was used. That can make a
- * one-packet flow look like it has a nonzero duration, which looks odd in
- * e.g. NetFlow statistics.
- *
- * If non-null, 'stats' will be folded into 'facet'. */
+ * All of the packets in 'miss' are considered to have arrived at time
+ * 'miss->stats.used'. This is really important only for new facets: if we
+ * just called time_msec() here, then the new subfacet or its packets could
+ * look (occasionally) as though it was used some time after the facet was
+ * used. That can make a one-packet flow look like it has a nonzero duration,
+ * which looks odd in e.g. NetFlow statistics. */
static void
handle_flow_miss_with_facet(struct flow_miss *miss, struct facet *facet,
- long long int now, struct dpif_flow_stats *stats,
struct flow_miss_op *ops, size_t *n_ops)
{
enum subfacet_path want_path;
struct subfacet *subfacet;
- struct ofpbuf *packet;
-
- want_path = facet->xout.slow ? SF_SLOW_PATH : SF_FAST_PATH;
-
- LIST_FOR_EACH (packet, list_node, &miss->packets) {
- struct flow_miss_op *op = &ops[*n_ops];
-
- handle_flow_miss_common(miss->ofproto, packet, &miss->flow,
- facet->fail_open);
- if (want_path != SF_FAST_PATH) {
- struct rule_dpif *rule;
- struct xlate_in xin;
-
- rule = rule_dpif_lookup(facet->ofproto, &facet->flow, NULL);
- xlate_in_init(&xin, facet->ofproto, &miss->flow, rule, 0, packet);
- xlate_actions_for_side_effects(&xin);
- }
+ facet->packet_count += miss->stats.n_packets;
+ facet->prev_packet_count += miss->stats.n_packets;
+ facet->byte_count += miss->stats.n_bytes;
+ facet->prev_byte_count += miss->stats.n_bytes;
- if (facet->xout.odp_actions.size) {
- struct dpif_execute *execute = &op->dpif_op.u.execute;
-
- init_flow_miss_execute_op(miss, packet, op);
- execute->actions = facet->xout.odp_actions.data,
- execute->actions_len = facet->xout.odp_actions.size;
- (*n_ops)++;
- }
- }
+ subfacet = subfacet_create(facet, miss);
+ want_path = facet->xout.slow ? SF_SLOW_PATH : SF_FAST_PATH;
/* Don't install the flow if it's the result of the "userspace"
* action for an already installed facet. This can occur when a
* be rejected as overlapping by the datapath. */
if (miss->upcall_type == DPIF_UC_ACTION
&& !list_is_empty(&facet->subfacets)) {
- if (stats) {
- facet->used = MAX(facet->used, stats->used);
- facet->packet_count += stats->n_packets;
- facet->byte_count += stats->n_bytes;
- facet->tcp_flags |= stats->tcp_flags;
- }
return;
}
- subfacet = subfacet_create(facet, miss, now);
- if (stats) {
- subfacet_update_stats(subfacet, stats);
- }
-
+ subfacet = subfacet_create(facet, miss);
if (subfacet->path != want_path) {
struct flow_miss_op *op = &ops[(*n_ops)++];
struct dpif_flow_put *put = &op->dpif_op.u.flow_put;
handle_flow_miss(struct flow_miss *miss, struct flow_miss_op *ops,
size_t *n_ops)
{
- struct ofproto_dpif *ofproto = miss->ofproto;
- struct dpif_flow_stats stats__;
- struct dpif_flow_stats *stats = &stats__;
- struct ofpbuf *packet;
struct facet *facet;
- long long int now;
- now = time_msec();
- memset(stats, 0, sizeof *stats);
- stats->used = now;
- LIST_FOR_EACH (packet, list_node, &miss->packets) {
- stats->tcp_flags |= packet_get_tcp_flags(packet, &miss->flow);
- stats->n_bytes += packet->size;
- stats->n_packets++;
- }
+ miss->ofproto->n_missed += list_size(&miss->packets);
- facet = facet_lookup_valid(ofproto, &miss->flow);
+ facet = facet_lookup_valid(miss->ofproto, &miss->flow);
if (!facet) {
- struct flow_wildcards wc;
- struct rule_dpif *rule;
- struct xlate_out xout;
- struct xlate_in xin;
-
- flow_wildcards_init_catchall(&wc);
- rule = rule_dpif_lookup(ofproto, &miss->flow, &wc);
- rule_credit_stats(rule, stats);
-
- xlate_in_init(&xin, ofproto, &miss->flow, rule, stats->tcp_flags,
- NULL);
- xin.resubmit_stats = stats;
- xin.may_learn = true;
- xlate_actions(&xin, &xout);
- flow_wildcards_or(&xout.wc, &xout.wc, &wc);
-
/* There does not exist a bijection between 'struct flow' and datapath
* flow keys with fitness ODP_FIT_TO_LITTLE. This breaks a fundamental
* assumption used throughout the facet and subfacet handling code.
* Since we have to handle these misses in userspace anyway, we simply
* skip facet creation, avoiding the problem altogether. */
if (miss->key_fitness == ODP_FIT_TOO_LITTLE
- || !flow_miss_should_make_facet(miss, &xout.wc)) {
- handle_flow_miss_without_facet(rule, &xout, miss, ops, n_ops);
+ || !flow_miss_should_make_facet(miss)) {
return;
}
- facet = facet_create(miss, rule, &xout, stats);
- stats = NULL;
+ facet = facet_create(miss);
}
- handle_flow_miss_with_facet(miss, facet, now, stats, ops, n_ops);
+ handle_flow_miss_with_facet(miss, facet, ops, n_ops);
}
static struct drop_key *
}
hmap_remove(&backer->drop_keys, &drop_key->hmap_node);
- free(drop_key->key);
- free(drop_key);
- }
-}
-
-/* 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. */
-static int
-ofproto_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)
-{
- const struct ofport_dpif *port;
- enum odp_key_fitness fitness;
- int error = ENODEV;
-
- 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;
+ drop_key_destroy(drop_key);
}
- port = (tnl_port_should_receive(flow)
- ? tnl_port_receive(flow)
- : odp_port_to_ofport(backer, flow->in_port.odp_port));
- flow->in_port.ofp_port = port ? port->up.ofp_port : OFPP_NONE;
- if (!port) {
- goto exit;
- }
-
- /* XXX: Since the tunnel module is not scoped per backer, for a tunnel port
- * it's theoretically possible that we'll receive an ofport belonging to an
- * entirely different datapath. In practice, this can't happen because no
- * platforms has two separate datapaths which each support tunneling. */
- ovs_assert(ofproto_dpif_cast(port->up.ofproto)->backer == backer);
-
- if (vsp_adjust_flow(ofproto_dpif_cast(port->up.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 = ofproto_dpif_cast(port->up.ofproto);
- }
-
-exit:
- if (fitnessp) {
- *fitnessp = fitness;
- }
- return error;
+ udpif_drop_key_clear(backer->udpif);
}
static void
-handle_miss_upcalls(struct dpif_backer *backer, struct dpif_upcall *upcalls,
- size_t n_upcalls)
+handle_flow_misses(struct dpif_backer *backer, struct flow_miss_batch *fmb)
{
- struct dpif_upcall *upcall;
+ struct flow_miss_op flow_miss_ops[FLOW_MISS_MAX_BATCH];
+ struct dpif_op *dpif_ops[FLOW_MISS_MAX_BATCH];
struct flow_miss *miss;
- struct flow_miss misses[FLOW_MISS_MAX_BATCH];
- struct flow_miss_op flow_miss_ops[FLOW_MISS_MAX_BATCH * 2];
- struct dpif_op *dpif_ops[FLOW_MISS_MAX_BATCH * 2];
- struct hmap todo;
- int n_misses;
- size_t n_ops;
- size_t i;
-
- if (!n_upcalls) {
- return;
- }
-
- /* Construct the to-do list.
- *
- * This just amounts to extracting the flow from each packet and sticking
- * the packets that have the same flow in the same "flow_miss" structure so
- * that we can process them together. */
- hmap_init(&todo);
- n_misses = 0;
- for (upcall = upcalls; upcall < &upcalls[n_upcalls]; upcall++) {
- struct flow_miss *miss = &misses[n_misses];
- struct flow_miss *existing_miss;
- struct ofproto_dpif *ofproto;
- odp_port_t odp_in_port;
- struct flow flow;
- uint32_t hash;
- int error;
-
- error = ofproto_receive(backer, upcall->packet, upcall->key,
- upcall->key_len, &flow, &miss->key_fitness,
- &ofproto, &odp_in_port);
- if (error == ENODEV) {
- struct drop_key *drop_key;
-
- /* Received packet on datapath port for which we couldn't
- * associate an ofproto. This can happen if a port is removed
- * while traffic is being received. Print a rate-limited message
- * in case it happens frequently. Install a drop flow so
- * that future packets of the flow are inexpensively dropped
- * in the kernel. */
- VLOG_INFO_RL(&rl, "received packet on unassociated datapath port "
- "%"PRIu32, odp_in_port);
-
- drop_key = drop_key_lookup(backer, upcall->key, upcall->key_len);
- if (!drop_key) {
- int ret;
- ret = dpif_flow_put(backer->dpif,
- DPIF_FP_CREATE | DPIF_FP_MODIFY,
- upcall->key, upcall->key_len,
- NULL, 0, NULL, 0, NULL);
-
- if (!ret) {
- drop_key = xmalloc(sizeof *drop_key);
- drop_key->key = xmemdup(upcall->key, upcall->key_len);
- drop_key->key_len = upcall->key_len;
-
- hmap_insert(&backer->drop_keys, &drop_key->hmap_node,
- hash_bytes(drop_key->key, drop_key->key_len, 0));
- }
- }
- continue;
- }
- if (error) {
- continue;
- }
-
- ofproto->n_missed++;
- flow_extract(upcall->packet, flow.skb_priority, flow.skb_mark,
- &flow.tunnel, &flow.in_port, &miss->flow);
-
- /* Add other packets to a to-do list. */
- hash = flow_hash(&miss->flow, 0);
- existing_miss = flow_miss_find(&todo, ofproto, &miss->flow, hash);
- if (!existing_miss) {
- hmap_insert(&todo, &miss->hmap_node, hash);
- miss->ofproto = ofproto;
- miss->key = upcall->key;
- miss->key_len = upcall->key_len;
- miss->upcall_type = upcall->type;
- list_init(&miss->packets);
-
- n_misses++;
- } else {
- miss = existing_miss;
- }
- list_push_back(&miss->packets, &upcall->packet->list_node);
- }
+ size_t n_ops, i;
/* Process each element in the to-do list, constructing the set of
* operations to batch. */
n_ops = 0;
- HMAP_FOR_EACH (miss, hmap_node, &todo) {
+ HMAP_FOR_EACH (miss, hmap_node, &fmb->misses) {
handle_flow_miss(miss, flow_miss_ops, &n_ops);
}
ovs_assert(n_ops <= ARRAY_SIZE(flow_miss_ops));
subfacet->path = SF_NOT_INSTALLED;
}
-
- /* Free memory. */
- if (flow_miss_ops[i].xout_garbage) {
- xlate_out_uninit(&flow_miss_ops[i].xout);
- }
- }
- hmap_destroy(&todo);
-}
-
-static enum { SFLOW_UPCALL, MISS_UPCALL, BAD_UPCALL, FLOW_SAMPLE_UPCALL,
- IPFIX_UPCALL }
-classify_upcall(const struct dpif_upcall *upcall)
-{
- size_t userdata_len;
- union user_action_cookie cookie;
-
- /* First look at the upcall type. */
- switch (upcall->type) {
- case DPIF_UC_ACTION:
- break;
-
- case DPIF_UC_MISS:
- return MISS_UPCALL;
-
- case DPIF_N_UC_TYPES:
- default:
- VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32, upcall->type);
- return BAD_UPCALL;
- }
-
- /* "action" upcalls need a closer look. */
- if (!upcall->userdata) {
- VLOG_WARN_RL(&rl, "action upcall missing cookie");
- return BAD_UPCALL;
- }
- userdata_len = nl_attr_get_size(upcall->userdata);
- if (userdata_len < sizeof cookie.type
- || userdata_len > sizeof cookie) {
- VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %zu",
- userdata_len);
- return BAD_UPCALL;
- }
- memset(&cookie, 0, sizeof cookie);
- memcpy(&cookie, nl_attr_get(upcall->userdata), userdata_len);
- if (userdata_len == sizeof cookie.sflow
- && cookie.type == USER_ACTION_COOKIE_SFLOW) {
- return SFLOW_UPCALL;
- } else if (userdata_len == sizeof cookie.slow_path
- && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) {
- return MISS_UPCALL;
- } else if (userdata_len == sizeof cookie.flow_sample
- && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) {
- return FLOW_SAMPLE_UPCALL;
- } else if (userdata_len == sizeof cookie.ipfix
- && cookie.type == USER_ACTION_COOKIE_IPFIX) {
- return IPFIX_UPCALL;
- } else {
- VLOG_WARN_RL(&rl, "invalid user cookie of type %"PRIu16
- " and size %zu", cookie.type, userdata_len);
- return BAD_UPCALL;
}
}
struct flow flow;
odp_port_t odp_in_port;
- if (ofproto_receive(backer, upcall->packet, upcall->key, upcall->key_len,
- &flow, NULL, &ofproto, &odp_in_port)
+ if (xlate_receive(backer, upcall->packet, upcall->key, upcall->key_len,
+ &flow, NULL, &ofproto, &odp_in_port)
|| !ofproto->sflow) {
return;
}
union user_action_cookie cookie;
struct flow flow;
- if (ofproto_receive(backer, upcall->packet, upcall->key, upcall->key_len,
- &flow, NULL, &ofproto, NULL)
+ if (xlate_receive(backer, upcall->packet, upcall->key, upcall->key_len,
+ &flow, NULL, &ofproto, NULL)
|| !ofproto->ipfix) {
return;
}
struct ofproto_dpif *ofproto;
struct flow flow;
- if (ofproto_receive(backer, upcall->packet, upcall->key, upcall->key_len,
- &flow, NULL, &ofproto, NULL)
+ if (xlate_receive(backer, upcall->packet, upcall->key, upcall->key_len,
+ &flow, NULL, &ofproto, NULL)
|| !ofproto->ipfix) {
return;
}
dpif_ipfix_bridge_sample(ofproto->ipfix, upcall->packet, &flow);
}
-static int
-handle_upcalls(struct dpif_backer *backer, unsigned int max_batch)
+static void
+handle_upcalls(struct dpif_backer *backer)
{
- struct dpif_upcall misses[FLOW_MISS_MAX_BATCH];
- struct ofpbuf miss_bufs[FLOW_MISS_MAX_BATCH];
- uint64_t miss_buf_stubs[FLOW_MISS_MAX_BATCH][4096 / 8];
+ struct flow_miss_batch *fmb;
int n_processed;
- int n_misses;
- int i;
- ovs_assert(max_batch <= FLOW_MISS_MAX_BATCH);
-
- n_misses = 0;
- for (n_processed = 0; n_processed < max_batch; n_processed++) {
- struct dpif_upcall *upcall = &misses[n_misses];
- struct ofpbuf *buf = &miss_bufs[n_misses];
- int error;
+ for (n_processed = 0; n_processed < FLOW_MISS_MAX_BATCH; n_processed++) {
+ struct upcall *upcall = upcall_next(backer->udpif);
- ofpbuf_use_stub(buf, miss_buf_stubs[n_misses],
- sizeof miss_buf_stubs[n_misses]);
- error = dpif_recv(backer->dpif, upcall, buf);
- if (error) {
- ofpbuf_uninit(buf);
+ if (!upcall) {
break;
}
- switch (classify_upcall(upcall)) {
- case MISS_UPCALL:
- /* Handle it later. */
- n_misses++;
- break;
-
+ switch (upcall->type) {
case SFLOW_UPCALL:
- handle_sflow_upcall(backer, upcall);
- ofpbuf_uninit(buf);
+ handle_sflow_upcall(backer, &upcall->dpif_upcall);
break;
case FLOW_SAMPLE_UPCALL:
- handle_flow_sample_upcall(backer, upcall);
- ofpbuf_uninit(buf);
+ handle_flow_sample_upcall(backer, &upcall->dpif_upcall);
break;
case IPFIX_UPCALL:
- handle_ipfix_upcall(backer, upcall);
- ofpbuf_uninit(buf);
+ handle_ipfix_upcall(backer, &upcall->dpif_upcall);
break;
case BAD_UPCALL:
- ofpbuf_uninit(buf);
break;
+
+ case MISS_UPCALL:
+ NOT_REACHED();
}
+
+ upcall_destroy(upcall);
}
- /* Handle deferred MISS_UPCALL processing. */
- handle_miss_upcalls(backer, misses, n_misses);
- for (i = 0; i < n_misses; i++) {
- ofpbuf_uninit(&miss_bufs[i]);
+ for (n_processed = 0; n_processed < FLOW_MISS_MAX_BATCH; n_processed++) {
+ struct drop_key *drop_key = drop_key_next(backer->udpif);
+ if (!drop_key) {
+ break;
+ }
+
+ if (!drop_key_lookup(backer, drop_key->key, drop_key->key_len)) {
+ hmap_insert(&backer->drop_keys, &drop_key->hmap_node,
+ hash_bytes(drop_key->key, drop_key->key_len, 0));
+ dpif_flow_put(backer->dpif, DPIF_FP_CREATE | DPIF_FP_MODIFY,
+ drop_key->key, drop_key->key_len,
+ NULL, 0, NULL, 0, NULL);
+ } else {
+ drop_key_destroy(drop_key);
+ }
}
- return n_processed;
+ fmb = flow_miss_batch_next(backer->udpif);
+ if (fmb) {
+ handle_flow_misses(backer, fmb);
+ flow_miss_batch_destroy(fmb);
+ }
}
\f
/* Flow expiration. */
/* Expire OpenFlow flows whose idle_timeout or hard_timeout
* has passed. */
+ ovs_mutex_lock(&ofproto->up.expirable_mutex);
LIST_FOR_EACH_SAFE (rule, next_rule, expirable,
&ofproto->up.expirable) {
rule_expire(rule_dpif_cast(rule));
}
+ ovs_mutex_unlock(&ofproto->up.expirable_mutex);
/* All outstanding data in existing flows has been accounted, so it's a
* good time to do bond rebalancing. */
static void
rule_expire(struct rule_dpif *rule)
{
+ uint16_t idle_timeout, hard_timeout;
long long int now;
uint8_t reason;
return;
}
+ ovs_mutex_lock(&rule->up.timeout_mutex);
+ hard_timeout = rule->up.hard_timeout;
+ idle_timeout = rule->up.idle_timeout;
+ ovs_mutex_unlock(&rule->up.timeout_mutex);
+
/* Has 'rule' expired? */
now = time_msec();
- if (rule->up.hard_timeout
- && now > rule->up.modified + rule->up.hard_timeout * 1000) {
+ if (hard_timeout && now > rule->up.modified + hard_timeout * 1000) {
reason = OFPRR_HARD_TIMEOUT;
- } else if (rule->up.idle_timeout
- && now > rule->up.used + rule->up.idle_timeout * 1000) {
+ } else if (idle_timeout && now > rule->up.used + idle_timeout * 1000) {
reason = OFPRR_IDLE_TIMEOUT;
} else {
return;
}
- COVERAGE_INC(ofproto_dpif_expired);
+ if (!ovs_rwlock_trywrlock(&rule->up.evict)) {
+ COVERAGE_INC(ofproto_dpif_expired);
- /* Get rid of the rule. */
- ofproto_rule_expire(&rule->up, reason);
+ /* Get rid of the rule. */
+ ofproto_rule_expire(&rule->up, reason);
+ }
}
\f
/* Facets. */
* The facet will initially have no subfacets. The caller should create (at
* least) one subfacet with subfacet_create(). */
static struct facet *
-facet_create(const struct flow_miss *miss, struct rule_dpif *rule,
- struct xlate_out *xout, struct dpif_flow_stats *stats)
+facet_create(const struct flow_miss *miss)
{
struct ofproto_dpif *ofproto = miss->ofproto;
struct facet *facet;
facet = xzalloc(sizeof *facet);
facet->ofproto = miss->ofproto;
- facet->packet_count = facet->prev_packet_count = stats->n_packets;
- facet->byte_count = facet->prev_byte_count = stats->n_bytes;
- facet->tcp_flags = stats->tcp_flags;
- facet->used = stats->used;
+ facet->used = miss->stats.used;
facet->flow = miss->flow;
facet->learn_rl = time_msec() + 500;
netflow_flow_init(&facet->nf_flow);
netflow_flow_update_time(ofproto->netflow, &facet->nf_flow, facet->used);
- xlate_out_copy(&facet->xout, xout);
+ xlate_out_copy(&facet->xout, &miss->xout);
match_init(&match, &facet->flow, &facet->xout.wc);
cls_rule_init(&facet->cr, &match, OFP_DEFAULT_PRIORITY);
+ ovs_rwlock_wrlock(&ofproto->facets.rwlock);
classifier_insert(&ofproto->facets, &facet->cr);
+ ovs_rwlock_unlock(&ofproto->facets.rwlock);
facet->nf_flow.output_iface = facet->xout.nf_output_iface;
- facet->fail_open = rule->up.cr.priority == FAIL_OPEN_PRIORITY;
-
return facet;
}
&facet->subfacets) {
subfacet_destroy__(subfacet);
}
+ ovs_rwlock_wrlock(&facet->ofproto->facets.rwlock);
classifier_remove(&facet->ofproto->facets, &facet->cr);
+ ovs_rwlock_unlock(&facet->ofproto->facets.rwlock);
cls_rule_destroy(&facet->cr);
facet_free(facet);
}
{
if (facet) {
struct ofproto_dpif *ofproto = facet->ofproto;
- const struct rule_dpif *rule = rule_dpif_lookup(ofproto, &facet->flow,
- NULL);
- const struct ofpact *ofpacts = rule->up.ofpacts;
- size_t ofpacts_len = rule->up.ofpacts_len;
-
- if (ofpacts_len > 0 &&
- ofpacts->type == OFPACT_CONTROLLER &&
- ofpact_next(ofpacts) >= ofpact_end(ofpacts, ofpacts_len)) {
- return true;
- }
+ const struct ofpact *ofpacts;
+ struct rule_dpif *rule;
+ size_t ofpacts_len;
+ bool is_controller;
+
+ rule_dpif_lookup(ofproto, &facet->flow, NULL, &rule);
+ ofpacts_len = rule->up.ofpacts_len;
+ ofpacts = rule->up.ofpacts;
+ is_controller = ofpacts_len > 0
+ && ofpacts->type == OFPACT_CONTROLLER
+ && ofpact_next(ofpacts) >= ofpact_end(ofpacts, ofpacts_len);
+ rule_release(rule);
+ return is_controller;
}
return false;
}
static struct facet *
facet_find(struct ofproto_dpif *ofproto, const struct flow *flow)
{
- struct cls_rule *cr = classifier_lookup(&ofproto->facets, flow, NULL);
+ struct cls_rule *cr;
+
+ ovs_rwlock_rdlock(&ofproto->facets.rwlock);
+ cr = classifier_lookup(&ofproto->facets, flow, NULL);
+ ovs_rwlock_unlock(&ofproto->facets.rwlock);
return cr ? CONTAINER_OF(cr, struct facet, cr) : NULL;
}
struct xlate_in xin;
struct rule_dpif *rule;
- bool ok, fail_open;
+ bool ok;
/* Check the datapath actions for consistency. */
- rule = rule_dpif_lookup(facet->ofproto, &facet->flow, NULL);
+ rule_dpif_lookup(facet->ofproto, &facet->flow, NULL, &rule);
xlate_in_init(&xin, facet->ofproto, &facet->flow, rule, 0, NULL);
xlate_actions(&xin, &xout);
+ rule_release(rule);
- fail_open = rule->up.cr.priority == FAIL_OPEN_PRIORITY;
ok = ofpbuf_equal(&facet->xout.odp_actions, &xout.odp_actions)
- && facet->xout.slow == xout.slow
- && facet->fail_open == fail_open;
+ && facet->xout.slow == xout.slow;
if (!ok && !VLOG_DROP_WARN(&rl)) {
struct ds s = DS_EMPTY_INITIALIZER;
ds_put_format(&s, " slow path incorrect. should be %d", xout.slow);
}
- if (facet->fail_open != fail_open) {
- ds_put_format(&s, " fail open incorrect. should be %s",
- fail_open ? "true" : "false");
- }
ds_destroy(&s);
}
xlate_out_uninit(&xout);
* where it is and recompiles its actions anyway.
*
* - If any of 'facet''s subfacets correspond to a new flow according to
- * ofproto_receive(), 'facet' is removed.
+ * xlate_receive(), 'facet' is removed.
*
* Returns true if 'facet' is still valid. False if 'facet' was removed. */
static bool
struct flow recv_flow;
int error;
- error = ofproto_receive(ofproto->backer, NULL, subfacet->key,
- subfacet->key_len, &recv_flow, NULL,
- &recv_ofproto, NULL);
+ error = xlate_receive(ofproto->backer, NULL, subfacet->key,
+ subfacet->key_len, &recv_flow, NULL,
+ &recv_ofproto, NULL);
if (error
|| recv_ofproto != ofproto
|| facet != facet_find(ofproto, &recv_flow)) {
}
flow_wildcards_init_catchall(&wc);
- new_rule = rule_dpif_lookup(ofproto, &facet->flow, &wc);
+ rule_dpif_lookup(ofproto, &facet->flow, &wc, &new_rule);
/* Calculate new datapath actions.
*
|| memcmp(&facet->xout.wc, &xout.wc, sizeof xout.wc)) {
facet_remove(facet);
xlate_out_uninit(&xout);
+ rule_release(new_rule);
return false;
}
facet->xout.mirrors = xout.mirrors;
facet->nf_flow.output_iface = facet->xout.nf_output_iface;
facet->used = MAX(facet->used, new_rule->up.created);
- facet->fail_open = new_rule->up.cr.priority == FAIL_OPEN_PRIORITY;
xlate_out_uninit(&xout);
+ rule_release(new_rule);
return true;
}
facet->accounted_bytes = 0;
}
+static void
+flow_push_stats(struct ofproto_dpif *ofproto, struct flow *flow,
+ struct dpif_flow_stats *stats, bool may_learn)
+{
+ struct ofport_dpif *in_port;
+ struct rule_dpif *rule;
+ struct xlate_in xin;
+
+ in_port = get_ofp_port(ofproto, flow->in_port.ofp_port);
+ if (in_port && in_port->is_tunnel) {
+ netdev_vport_inc_rx(in_port->up.netdev, stats);
+ }
+
+ rule_dpif_lookup(ofproto, flow, NULL, &rule);
+ rule_credit_stats(rule, stats);
+ xlate_in_init(&xin, ofproto, flow, rule, stats->tcp_flags, NULL);
+ xin.resubmit_stats = stats;
+ xin.may_learn = may_learn;
+ xlate_actions_for_side_effects(&xin);
+ rule_release(rule);
+}
+
static void
facet_push_stats(struct facet *facet, bool may_learn)
{
stats.tcp_flags = facet->tcp_flags;
if (may_learn || stats.n_packets || facet->used > facet->prev_used) {
- struct ofproto_dpif *ofproto = facet->ofproto;
- struct ofport_dpif *in_port;
- struct rule_dpif *rule;
- struct xlate_in xin;
-
facet->prev_packet_count = facet->packet_count;
facet->prev_byte_count = facet->byte_count;
facet->prev_used = facet->used;
- in_port = get_ofp_port(ofproto, facet->flow.in_port.ofp_port);
- if (in_port && in_port->is_tunnel) {
- netdev_vport_inc_rx(in_port->up.netdev, &stats);
- }
-
- rule = rule_dpif_lookup(ofproto, &facet->flow, NULL);
- rule_credit_stats(rule, &stats);
- netflow_flow_update_time(ofproto->netflow, &facet->nf_flow,
+ netflow_flow_update_time(facet->ofproto->netflow, &facet->nf_flow,
facet->used);
netflow_flow_update_flags(&facet->nf_flow, facet->tcp_flags);
- mirror_update_stats(ofproto->mbridge, facet->xout.mirrors,
+ mirror_update_stats(facet->ofproto->mbridge, facet->xout.mirrors,
stats.n_packets, stats.n_bytes);
-
- xlate_in_init(&xin, ofproto, &facet->flow, rule, stats.tcp_flags,
- NULL);
- xin.resubmit_stats = &stats;
- xin.may_learn = may_learn;
- xlate_actions_for_side_effects(&xin);
+ flow_push_stats(facet->ofproto, &facet->flow, &stats, may_learn);
}
}
struct cls_cursor cursor;
struct facet *facet;
+ ovs_rwlock_rdlock(&ofproto->facets.rwlock);
cls_cursor_init(&cursor, &ofproto->facets, NULL);
CLS_CURSOR_FOR_EACH (facet, cr, &cursor) {
facet_push_stats(facet, false);
run_fast_rl();
}
}
+ ovs_rwlock_unlock(&ofproto->facets.rwlock);
}
rl = time_msec() + 100;
* existing subfacet if there is one, otherwise creates and returns a
* new subfacet. */
static struct subfacet *
-subfacet_create(struct facet *facet, struct flow_miss *miss,
- long long int now)
+subfacet_create(struct facet *facet, struct flow_miss *miss)
{
struct dpif_backer *backer = miss->ofproto->backer;
enum odp_key_fitness key_fitness = miss->key_fitness;
subfacet->key_fitness = key_fitness;
subfacet->key = xmemdup(key, key_len);
subfacet->key_len = key_len;
- subfacet->used = now;
- subfacet->created = now;
+ subfacet->used = miss->stats.used;
+ subfacet->created = subfacet->used;
subfacet->dp_packet_count = 0;
subfacet->dp_byte_count = 0;
subfacet->path = SF_NOT_INSTALLED;
/* Lookup 'flow' in 'ofproto''s classifier. If 'wc' is non-null, sets
* the fields that were relevant as part of the lookup. */
-static struct rule_dpif *
+void
rule_dpif_lookup(struct ofproto_dpif *ofproto, const struct flow *flow,
- struct flow_wildcards *wc)
+ struct flow_wildcards *wc, struct rule_dpif **rule)
{
- struct rule_dpif *rule;
+ struct ofport_dpif *port;
- rule = rule_dpif_lookup_in_table(ofproto, flow, wc, 0);
- if (rule) {
- return rule;
+ if (rule_dpif_lookup_in_table(ofproto, flow, wc, 0, rule)) {
+ return;
+ }
+ port = get_ofp_port(ofproto, flow->in_port.ofp_port);
+ if (!port) {
+ VLOG_WARN_RL(&rl, "packet-in on unknown OpenFlow port %"PRIu16,
+ flow->in_port.ofp_port);
}
- return rule_dpif_miss_rule(ofproto, flow);
+ *rule = choose_miss_rule(port ? port->up.pp.config : 0, ofproto->miss_rule,
+ ofproto->no_packet_in_rule);
+ ovs_rwlock_rdlock(&(*rule)->up.evict);
}
-struct rule_dpif *
+bool
rule_dpif_lookup_in_table(struct ofproto_dpif *ofproto,
const struct flow *flow, struct flow_wildcards *wc,
- uint8_t table_id)
+ uint8_t table_id, struct rule_dpif **rule)
+ OVS_ACQ_RDLOCK((*rule)->up.evict)
{
struct cls_rule *cls_rule;
struct classifier *cls;
bool frag;
+ *rule = NULL;
if (table_id >= N_TABLES) {
- return NULL;
+ return false;
}
if (wc) {
}
cls = &ofproto->up.tables[table_id].cls;
+ ovs_rwlock_rdlock(&cls->rwlock);
frag = (flow->nw_frag & FLOW_NW_FRAG_ANY) != 0;
if (frag && ofproto->up.frag_handling == OFPC_FRAG_NORMAL) {
/* We must pretend that transport ports are unavailable. */
} else {
cls_rule = classifier_lookup(cls, flow, wc);
}
- return rule_dpif_cast(rule_from_cls_rule(cls_rule));
+
+ *rule = rule_dpif_cast(rule_from_cls_rule(cls_rule));
+ if (*rule && ovs_rwlock_tryrdlock(&(*rule)->up.evict)) {
+ /* The rule is in the process of being removed. Best we can do is
+ * pretend it isn't there. */
+ *rule = NULL;
+ }
+ ovs_rwlock_unlock(&cls->rwlock);
+
+ return *rule != NULL;
}
+/* Given a port configuration (specified as zero if there's no port), chooses
+ * which of 'miss_rule' and 'no_packet_in_rule' should be used in case of a
+ * flow table miss. */
struct rule_dpif *
-rule_dpif_miss_rule(struct ofproto_dpif *ofproto, const struct flow *flow)
+choose_miss_rule(enum ofputil_port_config config, struct rule_dpif *miss_rule,
+ struct rule_dpif *no_packet_in_rule)
{
- struct ofport_dpif *port;
-
- port = get_ofp_port(ofproto, flow->in_port.ofp_port);
- if (!port) {
- VLOG_WARN_RL(&rl, "packet-in on unknown OpenFlow port %"PRIu16,
- flow->in_port.ofp_port);
- return ofproto->miss_rule;
- }
+ return config & OFPUTIL_PC_NO_PACKET_IN ? no_packet_in_rule : miss_rule;
+}
- if (port->up.pp.config & OFPUTIL_PC_NO_PACKET_IN) {
- return ofproto->no_packet_in_rule;
+void
+rule_release(struct rule_dpif *rule)
+{
+ if (rule) {
+ ovs_rwlock_unlock(&rule->up.evict);
}
- return ofproto->miss_rule;
}
static void
struct cls_cursor cursor;
struct facet *facet;
+ ovs_rwlock_rdlock(&ofproto->facets.rwlock);
cls_cursor_init(&cursor, &ofproto->facets, NULL);
CLS_CURSOR_FOR_EACH (facet, cr, &cursor) {
send_active_timeout(ofproto, facet);
}
+ ovs_rwlock_unlock(&ofproto->facets.rwlock);
}
\f
static struct ofproto_dpif *
backer = node->data;
}
- /* Extract the ofproto_dpif object from the ofproto_receive()
- * function. */
- if (ofproto_receive(backer, NULL, odp_key.data,
- odp_key.size, &flow, NULL, &ofproto, NULL)) {
+ if (xlate_receive(backer, NULL, odp_key.data, odp_key.size, &flow,
+ NULL, &ofproto, NULL)) {
unixctl_command_reply_error(conn, "Invalid datapath flow");
goto exit;
}
const struct ofpbuf *packet, struct ds *ds)
{
struct rule_dpif *rule;
+ struct flow_wildcards wc;
ds_put_cstr(ds, "Flow: ");
flow_format(ds, flow);
ds_put_char(ds, '\n');
- rule = rule_dpif_lookup(ofproto, flow, NULL);
+ flow_wildcards_init_catchall(&wc);
+ rule_dpif_lookup(ofproto, flow, &wc, &rule);
trace_format_rule(ds, 0, rule);
if (rule == ofproto->miss_rule) {
trace.xin.report_hook = trace_report;
xlate_actions(&trace.xin, &trace.xout);
+ flow_wildcards_or(&trace.xout.wc, &trace.xout.wc, &wc);
ds_put_char(ds, '\n');
trace_format_flow(ds, 0, "Final flow", &trace);
xlate_out_uninit(&trace.xout);
}
+
+ rule_release(rule);
}
static void
int errors;
errors = 0;
+ ovs_rwlock_rdlock(&ofproto->facets.rwlock);
cls_cursor_init(&cursor, &ofproto->facets, NULL);
CLS_CURSOR_FOR_EACH (facet, cr, &cursor) {
if (!facet_check_consistency(facet)) {
errors++;
}
}
+ ovs_rwlock_unlock(&ofproto->facets.rwlock);
if (errors) {
ofproto->backer->need_revalidate = REV_INCONSISTENCY;
}
return;
}
+ ovs_rwlock_rdlock(&ofproto->facets.rwlock);
cls_cursor_init(&cursor, &ofproto->facets, NULL);
CLS_CURSOR_FOR_EACH (facet, cr, &cursor) {
cls_rule_format(&facet->cr, &ds);
}
ds_put_cstr(&ds, "\n");
}
+ ovs_rwlock_unlock(&ofproto->facets.rwlock);
ds_chomp(&ds, '\n');
unixctl_command_reply(conn, ds_cstr(&ds));
bool
ofproto_has_vlan_splinters(const struct ofproto_dpif *ofproto)
+ OVS_EXCLUDED(ofproto->vsp_mutex)
{
- return !hmap_is_empty(&ofproto->realdev_vid_map);
+ bool ret;
+
+ ovs_mutex_lock(&ofproto->vsp_mutex);
+ ret = !hmap_is_empty(&ofproto->realdev_vid_map);
+ ovs_mutex_unlock(&ofproto->vsp_mutex);
+ return ret;
}
-/* Returns the OFP port number of the Linux VLAN device that corresponds to
- * 'vlan_tci' on the network device with port number 'realdev_ofp_port' in
- * 'struct ofport_dpif'. For example, given 'realdev_ofp_port' of eth0 and
- * 'vlan_tci' 9, it would return the port number of eth0.9.
- *
- * Unless VLAN splinters are enabled for port 'realdev_ofp_port', this
- * function just returns its 'realdev_ofp_port' argument. */
-ofp_port_t
-vsp_realdev_to_vlandev(const struct ofproto_dpif *ofproto,
- ofp_port_t realdev_ofp_port, ovs_be16 vlan_tci)
+static ofp_port_t
+vsp_realdev_to_vlandev__(const struct ofproto_dpif *ofproto,
+ ofp_port_t realdev_ofp_port, ovs_be16 vlan_tci)
+ OVS_REQUIRES(ofproto->vsp_mutex)
{
if (!hmap_is_empty(&ofproto->realdev_vid_map)) {
int vid = vlan_tci_to_vid(vlan_tci);
return realdev_ofp_port;
}
+/* Returns the OFP port number of the Linux VLAN device that corresponds to
+ * 'vlan_tci' on the network device with port number 'realdev_ofp_port' in
+ * 'struct ofport_dpif'. For example, given 'realdev_ofp_port' of eth0 and
+ * 'vlan_tci' 9, it would return the port number of eth0.9.
+ *
+ * Unless VLAN splinters are enabled for port 'realdev_ofp_port', this
+ * function just returns its 'realdev_ofp_port' argument. */
+ofp_port_t
+vsp_realdev_to_vlandev(const struct ofproto_dpif *ofproto,
+ ofp_port_t realdev_ofp_port, ovs_be16 vlan_tci)
+ OVS_EXCLUDED(ofproto->vsp_mutex)
+{
+ ofp_port_t ret;
+
+ ovs_mutex_lock(&ofproto->vsp_mutex);
+ ret = vsp_realdev_to_vlandev__(ofproto, realdev_ofp_port, vlan_tci);
+ ovs_mutex_unlock(&ofproto->vsp_mutex);
+ return ret;
+}
+
static struct vlan_splinter *
vlandev_find(const struct ofproto_dpif *ofproto, ofp_port_t vlandev_ofp_port)
{
static ofp_port_t
vsp_vlandev_to_realdev(const struct ofproto_dpif *ofproto,
ofp_port_t vlandev_ofp_port, int *vid)
+ OVS_REQUIRES(ofproto->vsp_mutex)
{
if (!hmap_is_empty(&ofproto->vlandev_map)) {
const struct vlan_splinter *vsp;
* 'flow->vlan_tci' to the VLAN VID, and returns true. Otherwise (which is
* always the case unless VLAN splinters are enabled), returns false without
* making any changes. */
-static bool
+bool
vsp_adjust_flow(const struct ofproto_dpif *ofproto, struct flow *flow)
+ OVS_EXCLUDED(ofproto->vsp_mutex)
{
ofp_port_t realdev;
int vid;
+ ovs_mutex_lock(&ofproto->vsp_mutex);
realdev = vsp_vlandev_to_realdev(ofproto, flow->in_port.ofp_port, &vid);
+ ovs_mutex_unlock(&ofproto->vsp_mutex);
if (!realdev) {
return false;
}
struct ofproto_dpif *ofproto = ofproto_dpif_cast(port->up.ofproto);
struct vlan_splinter *vsp;
+ ovs_mutex_lock(&ofproto->vsp_mutex);
vsp = vlandev_find(ofproto, port->up.ofp_port);
if (vsp) {
hmap_remove(&ofproto->vlandev_map, &vsp->vlandev_node);
} else {
VLOG_ERR("missing vlan device record");
}
+ ovs_mutex_unlock(&ofproto->vsp_mutex);
}
static void
{
struct ofproto_dpif *ofproto = ofproto_dpif_cast(port->up.ofproto);
+ ovs_mutex_lock(&ofproto->vsp_mutex);
if (!vsp_vlandev_to_realdev(ofproto, port->up.ofp_port, NULL)
- && (vsp_realdev_to_vlandev(ofproto, realdev_ofp_port, htons(vid))
+ && (vsp_realdev_to_vlandev__(ofproto, realdev_ofp_port, htons(vid))
== realdev_ofp_port)) {
struct vlan_splinter *vsp;
vsp = xmalloc(sizeof *vsp);
- hmap_insert(&ofproto->vlandev_map, &vsp->vlandev_node,
- hash_ofp_port(port->up.ofp_port));
- hmap_insert(&ofproto->realdev_vid_map, &vsp->realdev_vid_node,
- hash_realdev_vid(realdev_ofp_port, vid));
vsp->realdev_ofp_port = realdev_ofp_port;
vsp->vlandev_ofp_port = port->up.ofp_port;
vsp->vid = vid;
port->realdev_ofp_port = realdev_ofp_port;
+
+ hmap_insert(&ofproto->vlandev_map, &vsp->vlandev_node,
+ hash_ofp_port(port->up.ofp_port));
+ hmap_insert(&ofproto->realdev_vid_map, &vsp->realdev_vid_node,
+ hash_realdev_vid(realdev_ofp_port, vid));
} else {
VLOG_ERR("duplicate vlan device record");
}
+ ovs_mutex_unlock(&ofproto->vsp_mutex);
}
static odp_port_t
return ofport ? ofport->odp_port : ODPP_NONE;
}
-static struct ofport_dpif *
+struct ofport_dpif *
odp_port_to_ofport(const struct dpif_backer *backer, odp_port_t odp_port)
{
struct ofport_dpif *port;
+ ovs_rwlock_rdlock(&backer->odp_to_ofport_lock);
HMAP_FOR_EACH_IN_BUCKET (port, odp_port_node, hash_odp_port(odp_port),
&backer->odp_to_ofport_map) {
if (port->odp_port == odp_port) {
+ ovs_rwlock_unlock(&backer->odp_to_ofport_lock);
return port;
}
}
+ ovs_rwlock_unlock(&backer->odp_to_ofport_lock);
return NULL;
}
git://git.onelab.eu / sliver-openvswitch.git / blobdiff