X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=ofproto%2Fofproto-dpif.c;h=b52d4eeff48c8ca00f0da282d3849b76b3d09bf2;hb=31ef9f5178dee18a40ce609aad406e925411eb06;hp=6ef2eb834af77decde1059e0c698db4512510e32;hpb=ec89fc6ff8367c446da811bbf74626c5bf6c2b30;p=sliver-openvswitch.git diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c index 6ef2eb834..b52d4eeff 100644 --- a/ofproto/ofproto-dpif.c +++ b/ofproto/ofproto-dpif.c @@ -52,6 +52,7 @@ #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" @@ -82,11 +83,29 @@ BUILD_ASSERT_DECL(N_TABLES >= 2 && N_TABLES <= 255); struct flow_miss; struct facet; -static struct rule_dpif *rule_dpif_lookup(struct ofproto_dpif *, - const struct flow *, - struct flow_wildcards *wc); +struct rule_dpif { + struct rule up; + + /* These statistics: + * + * - Do include packets and bytes from facets that have been deleted or + * whose own statistics have been folded into the rule. + * + * - Do include packets and bytes sent "by hand" that were accounted to + * the rule without any facet being involved (this is a rare corner + * case in rule_execute()). + * + * - Do not include packet or bytes that can be obtained from any facet's + * packet_count or byte_count member or that can be obtained from the + * datapath by, e.g., dpif_flow_get() for any subfacet. + */ + struct ovs_mutex stats_mutex; + uint64_t packet_count OVS_GUARDED; /* Number of packets received. */ + uint64_t byte_count OVS_GUARDED; /* Number of bytes received. */ +}; static void rule_get_stats(struct rule *, uint64_t *packets, uint64_t *bytes); +static struct rule_dpif *rule_dpif_cast(const struct rule *); struct ofbundle { struct hmap_node hmap_node; /* In struct ofproto's "bundles" hmap. */ @@ -170,8 +189,7 @@ struct subfacet { #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); @@ -209,8 +227,7 @@ static void subfacet_uninstall(struct subfacet *); * Flow expiration works in terms of subfacets, so a facet must have at * least one subfacet or it will never expire, leaking memory. */ struct facet { - /* Owners. */ - struct hmap_node hmap_node; /* In owning ofproto's 'facets' hmap. */ + /* Owner. */ struct ofproto_dpif *ofproto; /* Owned data. */ @@ -247,7 +264,6 @@ struct facet { 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 @@ -259,9 +275,7 @@ struct facet { 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 *); @@ -274,6 +288,8 @@ static bool facet_check_consistency(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 *); @@ -382,15 +398,6 @@ COVERAGE_DEFINE(rev_flow_table); 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. */ @@ -401,6 +408,7 @@ struct dpif_backer { char *type; int refcount; struct dpif *dpif; + struct udpif *udpif; struct timer next_expiration; struct ovs_rwlock odp_to_ofport_lock; @@ -442,6 +450,9 @@ struct dpif_backer { /* Number of subfacets added or deleted from 'created' to 'last_minute.' */ unsigned long long int total_subfacet_add_count; unsigned long long int total_subfacet_del_count; + + /* Number of upcall handling threads. */ + unsigned int n_handler_threads; }; /* All existing ofproto_backer instances, indexed by ofproto->up.type. */ @@ -534,8 +545,7 @@ static void ofproto_trace(struct ofproto_dpif *, const struct flow *, 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 *); @@ -571,6 +581,8 @@ ofproto_dpif_flow_mod(struct ofproto_dpif *ofproto, ovs_mutex_unlock(&ofproto->flow_mod_mutex); } +/* Appends 'pin' to the queue of "packet ins" to be sent to the controller. + * Takes ownership of 'pin' and pin->packet. */ void ofproto_dpif_send_packet_in(struct ofproto_dpif *ofproto, struct ofputil_packet_in *pin) @@ -708,13 +720,24 @@ type_run(const char *type) 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_handler_threads, + backer->recv_set_enable); dpif_flow_flush(backer->dpif); backer->need_revalidate = REV_RECONFIGURE; } + /* If the n_handler_threads is reconfigured, call udpif_recv_set() + * to reset the handler threads. */ + if (backer->n_handler_threads != n_handler_threads) { + udpif_recv_set(backer->udpif, n_handler_threads, + backer->recv_set_enable); + backer->n_handler_threads = n_handler_threads; + } + if (backer->need_revalidate) { struct ofproto_dpif *ofproto; struct simap_node *node; @@ -798,6 +821,7 @@ type_run(const char *type) continue; } + ovs_rwlock_wrlock(&xlate_rwlock); xlate_ofproto_set(ofproto, ofproto->up.name, ofproto->backer->dpif, ofproto->miss_rule, ofproto->no_packet_in_rule, ofproto->ml, @@ -819,7 +843,7 @@ type_run(const char *type) HMAP_FOR_EACH (ofport, up.hmap_node, &ofproto->up.ports) { int stp_port = ofport->stp_port ? stp_port_no(ofport->stp_port) - : 0; + : -1; xlate_ofport_set(ofproto, ofport->bundle, ofport, ofport->up.ofp_port, ofport->odp_port, ofport->up.netdev, ofport->cfm, @@ -828,13 +852,20 @@ type_run(const char *type) 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) { @@ -998,32 +1029,10 @@ process_dpif_port_error(struct dpif_backer *backer, int error) } 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; } @@ -1040,14 +1049,13 @@ type_run_fast(const char *type) 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; @@ -1057,23 +1065,6 @@ run_fast_rl(void) } 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 @@ -1093,6 +1084,8 @@ type_wait(const char *type) } timer_wait(&backer->next_expiration); + dpif_wait(backer->dpif); + udpif_wait(backer->udpif); } /* Basic life-cycle. */ @@ -1133,6 +1126,7 @@ close_dpif_backer(struct dpif_backer *backer) 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)); @@ -1202,6 +1196,7 @@ open_dpif_backer(const char *type, struct dpif_backer **backerp) free(backer); return error; } + backer->udpif = udpif_create(backer, backer->dpif); backer->type = xstrdup(type); backer->governor = NULL; @@ -1249,6 +1244,9 @@ open_dpif_backer(const char *type, struct dpif_backer **backerp) close_dpif_backer(backer); return error; } + udpif_recv_set(backer->udpif, n_handler_threads, + backer->recv_set_enable); + backer->n_handler_threads = n_handler_threads; backer->max_n_subfacet = 0; backer->created = time_msec(); @@ -1271,7 +1269,7 @@ construct(struct ofproto *ofproto_) { struct ofproto_dpif *ofproto = ofproto_dpif_cast(ofproto_); struct shash_node *node, *next; - odp_port_t max_ports; + uint32_t max_ports; int error; error = open_dpif_backer(ofproto->up.type, &ofproto->backer); @@ -1280,8 +1278,7 @@ construct(struct ofproto *ofproto_) } max_ports = dpif_get_max_ports(ofproto->backer->dpif); - ofproto_init_max_ports(ofproto_, u16_to_ofp(MIN(odp_to_u32(max_ports), - ofp_to_u16(OFPP_MAX)))); + ofproto_init_max_ports(ofproto_, MIN(max_ports, ofp_to_u16(OFPP_MAX))); ofproto->netflow = NULL; ofproto->sflow = NULL; @@ -1291,20 +1288,20 @@ construct(struct ofproto *ofproto_) 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); + ovs_mutex_init(&ofproto->vsp_mutex); classifier_init(&ofproto->facets); ofproto->consistency_rl = LLONG_MIN; list_init(&ofproto->completions); - ovs_mutex_init(&ofproto->flow_mod_mutex, PTHREAD_MUTEX_NORMAL); + ovs_mutex_init(&ofproto->flow_mod_mutex); ovs_mutex_lock(&ofproto->flow_mod_mutex); list_init(&ofproto->flow_mods); ofproto->n_flow_mods = 0; ovs_mutex_unlock(&ofproto->flow_mod_mutex); - ovs_mutex_init(&ofproto->pin_mutex, PTHREAD_MUTEX_NORMAL); + ovs_mutex_init(&ofproto->pin_mutex); ovs_mutex_lock(&ofproto->pin_mutex); list_init(&ofproto->pins); ofproto->n_pins = 0; @@ -1381,9 +1378,12 @@ add_internal_flow(struct ofproto_dpif *ofproto, int id, 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)) { + rule_dpif_release(*rulep); + } else { + NOT_REACHED(); + } return 0; } @@ -1440,12 +1440,25 @@ destruct(struct ofproto *ofproto_) { struct ofproto_dpif *ofproto = ofproto_dpif_cast(ofproto_); struct rule_dpif *rule, *next_rule; - struct ofputil_flow_mod *pin, *next_pin; + struct ofputil_packet_in *pin, *next_pin; struct ofputil_flow_mod *fm, *next_fm; + struct facet *facet, *next_facet; + struct cls_cursor cursor; struct oftable *table; + 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_facet, cr, &cursor) { + facet_remove(facet); + } + ofproto->backer->need_revalidate = REV_RECONFIGURE; + ovs_rwlock_wrlock(&xlate_rwlock); xlate_remove_ofproto(ofproto); + ovs_rwlock_unlock(&xlate_rwlock); + + flow_miss_batch_ofproto_destroyed(ofproto->backer->udpif, ofproto); hmap_remove(&all_ofproto_dpifs, &ofproto->all_ofproto_dpifs_node); complete_operations(ofproto); @@ -1453,11 +1466,14 @@ destruct(struct ofproto *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_delete(&ofproto->up, &table->cls, &rule->up); } + ovs_rwlock_unlock(&table->cls.rwlock); } + complete_operations(ofproto); ovs_mutex_lock(&ofproto->flow_mod_mutex); LIST_FOR_EACH_SAFE (fm, next_fm, list_node, &ofproto->flow_mods) { @@ -1473,7 +1489,7 @@ destruct(struct ofproto *ofproto_) LIST_FOR_EACH_SAFE (pin, next_pin, list_node, &ofproto->pins) { list_remove(&pin->list_node); ofproto->n_pins--; - free(pin->ofpacts); + free(CONST_CAST(void *, pin->packet)); free(pin); } ovs_mutex_unlock(&ofproto->pin_mutex); @@ -1516,14 +1532,9 @@ run_fast(struct ofproto *ofproto_) } ovs_mutex_lock(&ofproto->flow_mod_mutex); - if (ofproto->n_flow_mods) { - flow_mods = ofproto->flow_mods; - list_moved(&flow_mods); - list_init(&ofproto->flow_mods); - ofproto->n_flow_mods = 0; - } else { - list_init(&flow_mods); - } + list_move(&flow_mods, &ofproto->flow_mods); + list_init(&ofproto->flow_mods); + ofproto->n_flow_mods = 0; ovs_mutex_unlock(&ofproto->flow_mod_mutex); LIST_FOR_EACH_SAFE (fm, next_fm, list_node, &flow_mods) { @@ -1539,14 +1550,9 @@ run_fast(struct ofproto *ofproto_) } 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); - } + list_move(&pins, &ofproto->pins); + list_init(&ofproto->pins); + ofproto->n_pins = 0; ovs_mutex_unlock(&ofproto->pin_mutex); LIST_FOR_EACH_SAFE (pin, next_pin, list_node, &pins) { @@ -1601,6 +1607,9 @@ run(struct ofproto *ofproto_) if (ofproto->sflow) { dpif_sflow_run(ofproto->sflow); } + if (ofproto->ipfix) { + dpif_ipfix_run(ofproto->ipfix); + } HMAP_FOR_EACH (ofport, up.hmap_node, &ofproto->up.ports) { port_run(ofport); @@ -1617,6 +1626,7 @@ run(struct ofproto *ofproto_) 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) { @@ -1636,6 +1646,7 @@ run(struct ofproto *ofproto_) ofproto->backer->need_revalidate = REV_INCONSISTENCY; } } + ovs_rwlock_unlock(&ofproto->facets.rwlock); return 0; } @@ -1655,11 +1666,12 @@ wait(struct ofproto *ofproto_) return; } - dpif_wait(ofproto->backer->dpif); - dpif_recv_wait(ofproto->backer->dpif); if (ofproto->sflow) { dpif_sflow_wait(ofproto->sflow); } + if (ofproto->ipfix) { + dpif_ipfix_wait(ofproto->ipfix); + } HMAP_FOR_EACH (ofport, up.hmap_node, &ofproto->up.ports) { port_wait(ofport); } @@ -1688,12 +1700,16 @@ get_memory_usage(const struct ofproto *ofproto_, struct simap *usage) 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); } @@ -1864,7 +1880,9 @@ port_destruct(struct ofport *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); @@ -1915,6 +1933,10 @@ port_modified(struct ofport *port_) cfm_set_netdev(port->cfm, port->up.netdev); } + if (port->bfd) { + bfd_set_netdev(port->bfd, port->up.netdev); + } + if (port->is_tunnel && tnl_port_reconfigure(port, port->up.netdev, port->odp_port)) { ofproto_dpif_cast(port->up.ofproto)->backer->need_revalidate = @@ -1979,20 +2001,25 @@ set_ipfix( { struct ofproto_dpif *ofproto = ofproto_dpif_cast(ofproto_); struct dpif_ipfix *di = ofproto->ipfix; + bool has_options = bridge_exporter_options || flow_exporters_options; - if (bridge_exporter_options || flow_exporters_options) { - if (!di) { - di = ofproto->ipfix = dpif_ipfix_create(); - } + if (has_options && !di) { + di = ofproto->ipfix = dpif_ipfix_create(); + } + + if (di) { + /* Call set_options in any case to cleanly flush the flow + * caches in the last exporters that are to be destroyed. */ dpif_ipfix_set_options( di, bridge_exporter_options, flow_exporters_options, n_flow_exporters_options); - } else { - if (di) { + + if (!has_options) { dpif_ipfix_unref(di); ofproto->ipfix = NULL; } } + return 0; } @@ -2049,7 +2076,8 @@ set_bfd(struct ofport *ofport_, const struct smap *cfg) struct bfd *old; old = ofport->bfd; - ofport->bfd = bfd_configure(old, netdev_get_name(ofport->up.netdev), cfg); + ofport->bfd = bfd_configure(old, netdev_get_name(ofport->up.netdev), + cfg, ofport->up.netdev); if (ofport->bfd != old) { ofproto->backer->need_revalidate = REV_RECONFIGURE; } @@ -2424,7 +2452,7 @@ bundle_add_port(struct ofbundle *bundle, ofp_port_t ofp_port, if (port->bundle != bundle) { bundle->ofproto->backer->need_revalidate = REV_RECONFIGURE; if (port->bundle) { - bundle_del_port(port); + bundle_remove(&port->up); } port->bundle = bundle; @@ -2455,7 +2483,9 @@ bundle_destroy(struct ofbundle *bundle) 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); @@ -2885,7 +2915,7 @@ ofport_update_peer(struct ofport_dpif *ofport) { 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; @@ -2907,7 +2937,7 @@ ofport_update_peer(struct ofport_dpif *ofport) 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; @@ -2925,9 +2955,11 @@ ofport_update_peer(struct ofport_dpif *ofport) ofport->peer = peer; ofport->peer->peer = ofport; } + free(peer_peer); - return; + break; } + free(peer_name); } static void @@ -2958,6 +2990,8 @@ port_run(struct ofport_dpif *ofport) long long int carrier_seq = netdev_get_carrier_resets(ofport->up.netdev); bool carrier_changed = carrier_seq != ofport->carrier_seq; bool enable = netdev_get_carrier(ofport->up.netdev); + bool cfm_enable = false; + bool bfd_enable = false; ofport->carrier_seq = carrier_seq; @@ -2967,16 +3001,20 @@ port_run(struct ofport_dpif *ofport) int cfm_opup = cfm_get_opup(ofport->cfm); cfm_run(ofport->cfm); - enable = enable && !cfm_get_fault(ofport->cfm); + cfm_enable = !cfm_get_fault(ofport->cfm); if (cfm_opup >= 0) { - enable = enable && cfm_opup; + cfm_enable = cfm_enable && cfm_opup; } } if (ofport->bfd) { bfd_run(ofport->bfd); - enable = enable && bfd_forwarding(ofport->bfd); + bfd_enable = bfd_forwarding(ofport->bfd); + } + + if (ofport->bfd || ofport->cfm) { + enable = enable && (cfm_enable || bfd_enable); } if (ofport->bundle) { @@ -3093,7 +3131,7 @@ port_del(struct ofproto *ofproto_, ofp_port_t ofp_port) sset_find_and_delete(&ofproto->ghost_ports, netdev_get_name(ofport->up.netdev)); ofproto->backer->need_revalidate = REV_RECONFIGURE; - if (!ofport->is_tunnel) { + if (!ofport->is_tunnel && !netdev_vport_is_patch(ofport->up.netdev)) { error = dpif_port_del(ofproto->backer->dpif, ofport->odp_port); if (!error) { /* The caller is going to close ofport->up.netdev. If this is a @@ -3250,26 +3288,6 @@ port_is_lacp_current(const struct ofport *ofport_) /* 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; @@ -3285,96 +3303,6 @@ struct flow_miss_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 @@ -3383,7 +3311,7 @@ handle_flow_miss_common(struct ofproto_dpif *ofproto, struct ofpbuf *packet, * 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; @@ -3408,96 +3336,33 @@ flow_miss_should_make_facet(struct flow_miss *miss, struct flow_wildcards *wc) 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); - } - if (facet->xout.odp_actions.size) { - struct dpif_execute *execute = &op->dpif_op.u.execute; + 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; - 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)++; - } - } + 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 @@ -3506,20 +3371,10 @@ handle_flow_miss_with_facet(struct flow_miss *miss, struct facet *facet, * 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; @@ -3559,55 +3414,25 @@ static void 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 * @@ -3646,109 +3471,24 @@ drop_key_clear(struct dpif_backer *backer) } hmap_remove(&backer->drop_keys, &drop_key->hmap_node); - free(drop_key->key); - free(drop_key); + drop_key_destroy(drop_key); } + + 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 = xlate_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)); @@ -3781,66 +3521,6 @@ handle_miss_upcalls(struct dpif_backer *backer, struct dpif_upcall *upcalls, 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; } } @@ -3909,66 +3589,64 @@ handle_ipfix_upcall(struct dpif_backer *backer, 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); + for (n_processed = 0; n_processed < FLOW_MISS_MAX_BATCH; n_processed++) { + struct upcall *upcall = upcall_next(backer->udpif); - 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; - - 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); + } } /* Flow expiration. */ @@ -4200,7 +3878,7 @@ subfacet_max_idle(const struct dpif_backer *backer) * pass made by update_stats(), because the former function never looks at * uninstallable subfacets. */ - enum { BUCKET_WIDTH = ROUND_UP(100, TIME_UPDATE_INTERVAL) }; + enum { BUCKET_WIDTH = 100 }; enum { N_BUCKETS = 5000 / BUCKET_WIDTH }; int buckets[N_BUCKETS] = { 0 }; int total, subtotal, bucket; @@ -4321,8 +3999,6 @@ rule_expire(struct rule_dpif *rule) } COVERAGE_INC(ofproto_dpif_expired); - - /* Get rid of the rule. */ ofproto_rule_expire(&rule->up, reason); } @@ -4340,8 +4016,7 @@ rule_expire(struct rule_dpif *rule) * 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; @@ -4349,10 +4024,7 @@ facet_create(const struct flow_miss *miss, struct rule_dpif *rule, 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; @@ -4360,15 +4032,15 @@ facet_create(const struct flow_miss *miss, struct rule_dpif *rule, 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; } @@ -4432,7 +4104,9 @@ facet_remove(struct facet *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); } @@ -4516,16 +4190,19 @@ facet_is_controller_flow(struct facet *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_dpif_release(rule); + return is_controller; } return false; } @@ -4576,7 +4253,11 @@ facet_flush_stats(struct facet *facet) 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; } @@ -4608,17 +4289,16 @@ facet_check_consistency(struct facet *facet) 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_dpif_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; @@ -4639,10 +4319,6 @@ facet_check_consistency(struct facet *facet) 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); @@ -4694,7 +4370,7 @@ facet_revalidate(struct facet *facet) } 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. * @@ -4717,6 +4393,7 @@ facet_revalidate(struct facet *facet) || memcmp(&facet->xout.wc, &xout.wc, sizeof xout.wc)) { facet_remove(facet); xlate_out_uninit(&xout); + rule_dpif_release(new_rule); return false; } @@ -4746,9 +4423,9 @@ facet_revalidate(struct facet *facet) 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_dpif_release(new_rule); return true; } @@ -4762,6 +4439,28 @@ facet_reset_counters(struct facet *facet) 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_dpif_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_dpif_release(rule); +} + static void facet_push_stats(struct facet *facet, bool may_learn) { @@ -4777,33 +4476,16 @@ 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); } } @@ -4821,6 +4503,7 @@ push_all_stats__(bool run_fast) 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); @@ -4828,6 +4511,7 @@ push_all_stats__(bool run_fast) run_fast_rl(); } } + ovs_rwlock_unlock(&ofproto->facets.rwlock); } rl = time_msec() + 100; @@ -4840,14 +4524,42 @@ push_all_stats(void) } void -rule_credit_stats(struct rule_dpif *rule, const struct dpif_flow_stats *stats) +rule_dpif_credit_stats(struct rule_dpif *rule, + const struct dpif_flow_stats *stats) { ovs_mutex_lock(&rule->stats_mutex); rule->packet_count += stats->n_packets; rule->byte_count += stats->n_bytes; - ofproto_rule_update_used(&rule->up, stats->used); + rule->up.used = MAX(rule->up.used, stats->used); ovs_mutex_unlock(&rule->stats_mutex); } + +bool +rule_dpif_fail_open(const struct rule_dpif *rule) +{ + return rule->up.cr.priority == FAIL_OPEN_PRIORITY; +} + +ovs_be64 +rule_dpif_get_flow_cookie(const struct rule_dpif *rule) +{ + return rule->up.flow_cookie; +} + +void +rule_dpif_reduce_timeouts(struct rule_dpif *rule, uint16_t idle_timeout, + uint16_t hard_timeout) +{ + ofproto_rule_reduce_timeouts(&rule->up, idle_timeout, hard_timeout); +} + +void +rule_dpif_get_actions(const struct rule_dpif *rule, + const struct ofpact **ofpacts, size_t *ofpacts_len) +{ + *ofpacts = rule->up.ofpacts; + *ofpacts_len = rule->up.ofpacts_len; +} /* Subfacets. */ @@ -4873,8 +4585,7 @@ subfacet_find(struct dpif_backer *backer, const struct nlattr *key, * 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; @@ -4908,8 +4619,8 @@ subfacet_create(struct facet *facet, struct flow_miss *miss, 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; @@ -5102,16 +4813,14 @@ subfacet_update_stats(struct subfacet *subfacet, /* 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 ofport_dpif *port; - struct rule_dpif *rule; - 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) { @@ -5119,21 +4828,23 @@ rule_dpif_lookup(struct ofproto_dpif *ofproto, const struct flow *flow, flow->in_port.ofp_port); } - return choose_miss_rule(port ? port->up.pp.config : 0, ofproto->miss_rule, - ofproto->no_packet_in_rule); + choose_miss_rule(port ? port->up.pp.config : 0, ofproto->miss_rule, + ofproto->no_packet_in_rule, rule); } -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_TRY_RDLOCK(true, (*rule)->up.rwlock) { struct cls_rule *cls_rule; struct classifier *cls; bool frag; + *rule = NULL; if (table_id >= N_TABLES) { - return NULL; + return false; } if (wc) { @@ -5142,6 +4853,7 @@ rule_dpif_lookup_in_table(struct ofproto_dpif *ofproto, } 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. */ @@ -5157,17 +4869,37 @@ rule_dpif_lookup_in_table(struct ofproto_dpif *ofproto, } 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.rwlock)) { + /* 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 * +void choose_miss_rule(enum ofputil_port_config config, struct rule_dpif *miss_rule, - struct rule_dpif *no_packet_in_rule) + struct rule_dpif *no_packet_in_rule, struct rule_dpif **rule) + OVS_NO_THREAD_SAFETY_ANALYSIS { - return config & OFPUTIL_PC_NO_PACKET_IN ? no_packet_in_rule : miss_rule; + *rule = config & OFPUTIL_PC_NO_PACKET_IN ? no_packet_in_rule : miss_rule; + ovs_rwlock_rdlock(&(*rule)->up.rwlock); +} + +void +rule_dpif_release(struct rule_dpif *rule) + OVS_NO_THREAD_SAFETY_ANALYSIS +{ + if (rule) { + ovs_rwlock_unlock(&rule->up.rwlock); + } } static void @@ -5185,6 +4917,11 @@ complete_operation(struct rule_dpif *rule) } } +static struct rule_dpif *rule_dpif_cast(const struct rule *rule) +{ + return rule ? CONTAINER_OF(rule, struct rule_dpif, up) : NULL; +} + static struct rule * rule_alloc(void) { @@ -5203,20 +4940,32 @@ static enum ofperr rule_construct(struct rule *rule_) { struct rule_dpif *rule = rule_dpif_cast(rule_); - ovs_mutex_init(&rule->stats_mutex, PTHREAD_MUTEX_NORMAL); + ovs_mutex_init(&rule->stats_mutex); ovs_mutex_lock(&rule->stats_mutex); rule->packet_count = 0; rule->byte_count = 0; ovs_mutex_unlock(&rule->stats_mutex); - complete_operation(rule); return 0; } static void -rule_destruct(struct rule *rule_) +rule_insert(struct rule *rule_) +{ + struct rule_dpif *rule = rule_dpif_cast(rule_); + complete_operation(rule); +} + +static void +rule_delete(struct rule *rule_) { struct rule_dpif *rule = rule_dpif_cast(rule_); complete_operation(rule); +} + +static void +rule_destruct(struct rule *rule_) +{ + struct rule_dpif *rule = rule_dpif_cast(rule_); ovs_mutex_destroy(&rule->stats_mutex); } @@ -5249,7 +4998,7 @@ rule_dpif_execute(struct rule_dpif *rule, const struct flow *flow, struct xlate_in xin; dpif_flow_stats_extract(flow, packet, time_msec(), &stats); - rule_credit_stats(rule, &stats); + rule_dpif_credit_stats(rule, &stats); xlate_in_init(&xin, ofproto, flow, rule, stats.tcp_flags, packet); xin.resubmit_stats = &stats; @@ -5271,10 +5020,17 @@ rule_execute(struct rule *rule, const struct flow *flow, } static void -rule_modify_actions(struct rule *rule_) +rule_modify_actions(struct rule *rule_, bool reset_counters) { struct rule_dpif *rule = rule_dpif_cast(rule_); + if (reset_counters) { + ovs_mutex_lock(&rule->stats_mutex); + rule->packet_count = 0; + rule->byte_count = 0; + ovs_mutex_unlock(&rule->stats_mutex); + } + complete_operation(rule); } @@ -5483,10 +5239,12 @@ send_netflow_active_timeouts(struct ofproto_dpif *ofproto) 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); } static struct ofproto_dpif * @@ -5762,7 +5520,7 @@ ofproto_unixctl_trace(struct unixctl_conn *conn, int argc, const char *argv[], /* Use the metadata from the flow and the packet argument * to reconstruct the flow. */ - flow_extract(packet, flow.skb_priority, flow.skb_mark, NULL, + flow_extract(packet, flow.skb_priority, flow.pkt_mark, NULL, &in_port_, &flow); } } @@ -5782,12 +5540,14 @@ ofproto_trace(struct ofproto_dpif *ofproto, const struct flow *flow, 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) { @@ -5817,6 +5577,7 @@ ofproto_trace(struct ofproto_dpif *ofproto, const struct flow *flow, 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); @@ -5857,6 +5618,8 @@ ofproto_trace(struct ofproto_dpif *ofproto, const struct flow *flow, xlate_out_uninit(&trace.xout); } + + rule_dpif_release(rule); } static void @@ -5885,12 +5648,14 @@ ofproto_dpif_self_check__(struct ofproto_dpif *ofproto, struct ds *reply) 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; } @@ -6111,6 +5876,7 @@ ofproto_unixctl_dpif_dump_megaflows(struct unixctl_conn *conn, 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); @@ -6133,6 +5899,7 @@ ofproto_unixctl_dpif_dump_megaflows(struct unixctl_conn *conn, } ds_put_cstr(&ds, "\n"); } + ovs_rwlock_unlock(&ofproto->facets.rwlock); ds_chomp(&ds, '\n'); unixctl_command_reply(conn, ds_cstr(&ds)); @@ -6435,7 +6202,7 @@ 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_REQ_WRLOCK(ofproto->vsp_mutex) + OVS_REQUIRES(ofproto->vsp_mutex) { if (!hmap_is_empty(&ofproto->vlandev_map)) { const struct vlan_splinter *vsp; @@ -6648,6 +6415,8 @@ const struct ofproto_class ofproto_dpif_class = { NULL, /* rule_choose_table */ rule_alloc, rule_construct, + rule_insert, + rule_delete, rule_destruct, rule_dealloc, rule_get_stats, @@ -6681,4 +6450,10 @@ const struct ofproto_class ofproto_dpif_class = { NULL, /* meter_set */ NULL, /* meter_get */ NULL, /* meter_del */ + NULL, /* group_alloc */ + NULL, /* group_construct */ + NULL, /* group_destruct */ + NULL, /* group_dealloc */ + NULL, /* group_modify */ + NULL, /* group_get_stats */ };