X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=ofproto%2Fofproto-dpif-upcall.c;h=906bf17f157cbd3c8508fdcef36782abbd304f52;hb=73a3c4757e596ff156d40f41496a0264373e5bc4;hp=db81dbecc0617459ab52b0632da6932431fc45b0;hpb=e0eecb1ca122bd8a22058e42161bd93c8474a74f;p=sliver-openvswitch.git diff --git a/ofproto/ofproto-dpif-upcall.c b/ofproto/ofproto-dpif-upcall.c index db81dbecc..906bf17f1 100644 --- a/ofproto/ofproto-dpif-upcall.c +++ b/ofproto/ofproto-dpif-upcall.c @@ -45,26 +45,15 @@ VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall); -COVERAGE_DEFINE(upcall_queue_overflow); +COVERAGE_DEFINE(upcall_duplicate_flow); -/* A thread that processes each upcall handed to it by the dispatcher thread, - * forwards the upcall's packet, and possibly sets up a kernel flow as a - * cache. */ +/* A thread that reads upcalls from dpif, forwards each upcall's packet, + * and possibly sets up a kernel flow as a cache. */ struct handler { struct udpif *udpif; /* Parent udpif. */ pthread_t thread; /* Thread ID. */ char *name; /* Thread name. */ - - struct ovs_mutex mutex; /* Mutex guarding the following. */ - - /* Atomic queue of unprocessed upcalls. */ - struct list upcalls OVS_GUARDED; - size_t n_upcalls OVS_GUARDED; - - bool need_signal; /* Only changed by the dispatcher. */ - - pthread_cond_t wake_cond; /* Wakes 'thread' while holding - 'mutex'. */ + uint32_t handler_id; /* Handler id. */ }; /* A thread that processes each kernel flow handed to it by the flow_dumper @@ -87,14 +76,24 @@ struct revalidator { /* An upcall handler for ofproto_dpif. * - * udpif has two logically separate pieces: + * udpif keeps records of two kind of logically separate units: + * + * upcall handling + * --------------- * - * - A "dispatcher" thread that reads upcalls from the kernel and dispatches - * them to one of several "handler" threads (see struct handler). + * - An array of 'struct handler's for upcall handling and flow + * installation. + * + * flow revalidation + * ----------------- + * + * - An array of 'struct revalidator's for flow revalidation and + * stats collection. * * - A "flow_dumper" thread that reads the kernel flow table and dispatches * flows to one of several "revalidator" threads (see struct - * revalidator). */ + * revalidator). + * */ struct udpif { struct list list_node; /* In all_udpifs list. */ @@ -103,7 +102,6 @@ struct udpif { uint32_t secret; /* Random seed for upcall hash. */ - pthread_t dispatcher; /* Dispatcher thread ID. */ pthread_t flow_dumper; /* Flow dumper thread ID. */ struct handler *handlers; /* Upcall handlers. */ @@ -143,7 +141,6 @@ enum upcall_type { }; struct upcall { - struct list list_node; /* For queuing upcalls. */ struct flow_miss *flow_miss; /* This upcall's flow_miss. */ /* Raw upcall plus data for keeping track of the memory backing it. */ @@ -166,8 +163,12 @@ struct udpif_key { long long int created; /* Estimation of creation time. */ bool mark; /* Used by mark and sweep GC algorithm. */ + bool flow_exists; /* Ensures flows are only deleted once. */ struct odputil_keybuf key_buf; /* Memory for 'key'. */ + struct xlate_cache *xcache; /* Cache for xlate entries that + * are affected by this ukey. + * Used for stats and learning.*/ }; /* 'udpif_flow_dump's hold the state associated with one iteration in a flow @@ -216,15 +217,19 @@ struct flow_miss { bool put; }; -static void upcall_destroy(struct upcall *); - static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); static struct list all_udpifs = LIST_INITIALIZER(&all_udpifs); -static void recv_upcalls(struct udpif *); -static void handle_upcalls(struct handler *handler, struct list *upcalls); +static size_t read_upcalls(struct handler *, + struct upcall upcalls[FLOW_MISS_MAX_BATCH], + struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH], + struct hmap *); +static void handle_upcalls(struct handler *, struct hmap *, struct upcall *, + size_t n_upcalls); +static void udpif_stop_threads(struct udpif *); +static void udpif_start_threads(struct udpif *, size_t n_handlers, + size_t n_revalidators); static void *udpif_flow_dumper(void *); -static void *udpif_dispatcher(void *); static void *udpif_upcall_handler(void *); static void *udpif_revalidator(void *); static uint64_t udpif_get_n_flows(struct udpif *); @@ -279,8 +284,7 @@ udpif_create(struct dpif_backer *backer, struct dpif *dpif) void udpif_destroy(struct udpif *udpif) { - udpif_set_threads(udpif, 0, 0); - udpif_flush(udpif); + udpif_stop_threads(udpif); list_remove(&udpif->list_node); latch_destroy(&udpif->exit_latch); @@ -290,20 +294,12 @@ udpif_destroy(struct udpif *udpif) free(udpif); } -/* Tells 'udpif' how many threads it should use to handle upcalls. Disables - * all threads if 'n_handlers' and 'n_revalidators' is zero. 'udpif''s - * datapath handle must have packet reception enabled before starting threads. - */ -void -udpif_set_threads(struct udpif *udpif, size_t n_handlers, - size_t n_revalidators) +/* Stops the handler and revalidator threads, must be enclosed in + * ovsrcu quiescent state unless when destroying udpif. */ +static void +udpif_stop_threads(struct udpif *udpif) { - ovsrcu_quiesce_start(); - - /* Stop the old threads (if any). */ - if (udpif->handlers && - (udpif->n_handlers != n_handlers - || udpif->n_revalidators != n_revalidators)) { + if (udpif && (udpif->n_handlers != 0 || udpif->n_revalidators != 0)) { size_t i; latch_set(&udpif->exit_latch); @@ -311,9 +307,6 @@ udpif_set_threads(struct udpif *udpif, size_t n_handlers, for (i = 0; i < udpif->n_handlers; i++) { struct handler *handler = &udpif->handlers[i]; - ovs_mutex_lock(&handler->mutex); - xpthread_cond_signal(&handler->wake_cond); - ovs_mutex_unlock(&handler->mutex); xpthread_join(handler->thread, NULL); } @@ -327,7 +320,6 @@ udpif_set_threads(struct udpif *udpif, size_t n_handlers, } xpthread_join(udpif->flow_dumper, NULL); - xpthread_join(udpif->dispatcher, NULL); for (i = 0; i < udpif->n_revalidators; i++) { struct revalidator *revalidator = &udpif->revalidators[i]; @@ -349,17 +341,7 @@ udpif_set_threads(struct udpif *udpif, size_t n_handlers, } for (i = 0; i < udpif->n_handlers; i++) { - struct handler *handler = &udpif->handlers[i]; - struct upcall *miss, *next; - - LIST_FOR_EACH_SAFE (miss, next, list_node, &handler->upcalls) { - list_remove(&miss->list_node); - upcall_destroy(miss); - } - ovs_mutex_destroy(&handler->mutex); - - xpthread_cond_destroy(&handler->wake_cond); - free(handler->name); + free(udpif->handlers[i].name); } latch_poll(&udpif->exit_latch); @@ -371,9 +353,15 @@ udpif_set_threads(struct udpif *udpif, size_t n_handlers, udpif->handlers = NULL; udpif->n_handlers = 0; } +} - /* Start new threads (if necessary). */ - if (!udpif->handlers && n_handlers) { +/* Starts the handler and revalidator threads, must be enclosed in + * ovsrcu quiescent state. */ +static void +udpif_start_threads(struct udpif *udpif, size_t n_handlers, + size_t n_revalidators) +{ + if (udpif && (!udpif->handlers && !udpif->revalidators)) { size_t i; udpif->n_handlers = n_handlers; @@ -384,10 +372,7 @@ udpif_set_threads(struct udpif *udpif, size_t n_handlers, struct handler *handler = &udpif->handlers[i]; handler->udpif = udpif; - list_init(&handler->upcalls); - handler->need_signal = false; - xpthread_cond_init(&handler->wake_cond, NULL); - ovs_mutex_init(&handler->mutex); + handler->handler_id = i; xpthread_create(&handler->thread, NULL, udpif_upcall_handler, handler); } @@ -405,10 +390,39 @@ udpif_set_threads(struct udpif *udpif, size_t n_handlers, xpthread_create(&revalidator->thread, NULL, udpif_revalidator, revalidator); } - xpthread_create(&udpif->dispatcher, NULL, udpif_dispatcher, udpif); xpthread_create(&udpif->flow_dumper, NULL, udpif_flow_dumper, udpif); } +} + +/* Tells 'udpif' how many threads it should use to handle upcalls. + * 'n_handlers' and 'n_revalidators' can never be zero. 'udpif''s + * datapath handle must have packet reception enabled before starting + * threads. */ +void +udpif_set_threads(struct udpif *udpif, size_t n_handlers, + size_t n_revalidators) +{ + int error; + + ovs_assert(udpif); + ovs_assert(n_handlers && n_revalidators); + + ovsrcu_quiesce_start(); + if (udpif->n_handlers != n_handlers + || udpif->n_revalidators != n_revalidators) { + udpif_stop_threads(udpif); + } + error = dpif_handlers_set(udpif->dpif, n_handlers); + if (error) { + VLOG_ERR("failed to configure handlers in dpif %s: %s", + dpif_name(udpif->dpif), ovs_strerror(error)); + return; + } + + if (!udpif->handlers && !udpif->revalidators) { + udpif_start_threads(udpif, n_handlers, n_revalidators); + } ovsrcu_quiesce_end(); } @@ -424,8 +438,11 @@ udpif_synchronize(struct udpif *udpif) * its main loop once. */ size_t n_handlers = udpif->n_handlers; size_t n_revalidators = udpif->n_revalidators; - udpif_set_threads(udpif, 0, 0); - udpif_set_threads(udpif, n_handlers, n_revalidators); + + ovsrcu_quiesce_start(); + udpif_stop_threads(udpif); + udpif_start_threads(udpif, n_handlers, n_revalidators); + ovsrcu_quiesce_end(); } /* Notifies 'udpif' that something changed which may render previous @@ -450,16 +467,9 @@ udpif_get_memory_usage(struct udpif *udpif, struct simap *usage) { size_t i; - simap_increase(usage, "dispatchers", 1); simap_increase(usage, "flow_dumpers", 1); simap_increase(usage, "handlers", udpif->n_handlers); - for (i = 0; i < udpif->n_handlers; i++) { - struct handler *handler = &udpif->handlers[i]; - ovs_mutex_lock(&handler->mutex); - simap_increase(usage, "handler upcalls", handler->n_upcalls); - ovs_mutex_unlock(&handler->mutex); - } simap_increase(usage, "revalidators", udpif->n_revalidators); for (i = 0; i < udpif->n_revalidators; i++) { @@ -484,9 +494,13 @@ udpif_flush(struct udpif *udpif) n_handlers = udpif->n_handlers; n_revalidators = udpif->n_revalidators; - udpif_set_threads(udpif, 0, 0); + ovsrcu_quiesce_start(); + + udpif_stop_threads(udpif); dpif_flow_flush(udpif->dpif); - udpif_set_threads(udpif, n_handlers, n_revalidators); + udpif_start_threads(udpif, n_handlers, n_revalidators); + + ovsrcu_quiesce_end(); } /* Removes all flows from all datapaths. */ @@ -501,17 +515,6 @@ udpif_flush_all_datapaths(void) } -/* Destroys and deallocates 'upcall'. */ -static void -upcall_destroy(struct upcall *upcall) -{ - if (upcall) { - ofpbuf_uninit(&upcall->dpif_upcall.packet); - ofpbuf_uninit(&upcall->upcall_buf); - free(upcall); - } -} - static uint64_t udpif_get_n_flows(struct udpif *udpif) { @@ -534,24 +537,6 @@ udpif_get_n_flows(struct udpif *udpif) return flow_count; } -/* The dispatcher thread is responsible for receiving upcalls from the kernel, - * assigning them to a upcall_handler thread. */ -static void * -udpif_dispatcher(void *arg) -{ - struct udpif *udpif = arg; - - set_subprogram_name("dispatcher"); - while (!latch_is_set(&udpif->exit_latch)) { - recv_upcalls(udpif); - dpif_recv_wait(udpif->dpif); - latch_wait(&udpif->exit_latch); - poll_block(); - } - - return NULL; -} - static void * udpif_flow_dumper(void *arg) { @@ -671,40 +656,45 @@ skip: return NULL; } -/* The miss handler thread is responsible for processing miss upcalls retrieved - * by the dispatcher thread. Once finished it passes the processed miss - * upcalls to ofproto-dpif where they're installed in the datapath. */ +/* The upcall handler thread tries to read a batch of FLOW_MISS_MAX_BATCH + * upcalls from dpif, processes the batch and installs corresponding flows + * in dpif. */ static void * udpif_upcall_handler(void *arg) { struct handler *handler = arg; + struct udpif *udpif = handler->udpif; + struct hmap misses = HMAP_INITIALIZER(&misses); handler->name = xasprintf("handler_%u", ovsthread_id_self()); set_subprogram_name("%s", handler->name); while (!latch_is_set(&handler->udpif->exit_latch)) { - struct list misses = LIST_INITIALIZER(&misses); - size_t i; - - ovs_mutex_lock(&handler->mutex); - if (!handler->n_upcalls) { - ovs_mutex_cond_wait(&handler->wake_cond, &handler->mutex); - } + struct upcall upcalls[FLOW_MISS_MAX_BATCH]; + struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH]; + struct flow_miss *miss; + size_t n_upcalls, i; + + n_upcalls = read_upcalls(handler, upcalls, miss_buf, &misses); + if (!n_upcalls) { + dpif_recv_wait(udpif->dpif, handler->handler_id); + latch_wait(&udpif->exit_latch); + poll_block(); + } else { + handle_upcalls(handler, &misses, upcalls, n_upcalls); - for (i = 0; i < FLOW_MISS_MAX_BATCH; i++) { - if (handler->n_upcalls) { - handler->n_upcalls--; - list_push_back(&misses, list_pop_front(&handler->upcalls)); - } else { - break; + HMAP_FOR_EACH (miss, hmap_node, &misses) { + xlate_out_uninit(&miss->xout); + } + hmap_clear(&misses); + for (i = 0; i < n_upcalls; i++) { + ofpbuf_uninit(&upcalls[i].dpif_upcall.packet); + ofpbuf_uninit(&upcalls[i].upcall_buf); } } - ovs_mutex_unlock(&handler->mutex); - - handle_upcalls(handler, &misses); - coverage_clear(); } + hmap_destroy(&misses); return NULL; } @@ -809,98 +799,12 @@ classify_upcall(const struct upcall *upcall) } } -static void -recv_upcalls(struct udpif *udpif) -{ - int n; - - for (;;) { - uint32_t hash = udpif->secret; - struct handler *handler; - struct upcall *upcall; - size_t n_bytes, left; - struct nlattr *nla; - int error; - - upcall = xmalloc(sizeof *upcall); - ofpbuf_use_stub(&upcall->upcall_buf, upcall->upcall_stub, - sizeof upcall->upcall_stub); - error = dpif_recv(udpif->dpif, &upcall->dpif_upcall, - &upcall->upcall_buf); - if (error) { - /* upcall_destroy() can only be called on successfully received - * upcalls. */ - ofpbuf_uninit(&upcall->upcall_buf); - free(upcall); - break; - } - - n_bytes = 0; - NL_ATTR_FOR_EACH (nla, left, upcall->dpif_upcall.key, - upcall->dpif_upcall.key_len) { - enum ovs_key_attr type = nl_attr_type(nla); - if (type == OVS_KEY_ATTR_IN_PORT - || type == OVS_KEY_ATTR_TCP - || type == OVS_KEY_ATTR_UDP) { - if (nl_attr_get_size(nla) == 4) { - hash = mhash_add(hash, nl_attr_get_u32(nla)); - n_bytes += 4; - } else { - VLOG_WARN_RL(&rl, - "Netlink attribute with incorrect size."); - } - } - } - hash = mhash_finish(hash, n_bytes); - - handler = &udpif->handlers[hash % udpif->n_handlers]; - - ovs_mutex_lock(&handler->mutex); - if (handler->n_upcalls < MAX_QUEUE_LENGTH) { - list_push_back(&handler->upcalls, &upcall->list_node); - if (handler->n_upcalls == 0) { - handler->need_signal = true; - } - handler->n_upcalls++; - if (handler->need_signal && - handler->n_upcalls >= FLOW_MISS_MAX_BATCH) { - handler->need_signal = false; - xpthread_cond_signal(&handler->wake_cond); - } - ovs_mutex_unlock(&handler->mutex); - if (!VLOG_DROP_DBG(&rl)) { - struct ds ds = DS_EMPTY_INITIALIZER; - - odp_flow_key_format(upcall->dpif_upcall.key, - upcall->dpif_upcall.key_len, - &ds); - VLOG_DBG("dispatcher: enqueue (%s)", ds_cstr(&ds)); - ds_destroy(&ds); - } - } else { - ovs_mutex_unlock(&handler->mutex); - COVERAGE_INC(upcall_queue_overflow); - upcall_destroy(upcall); - } - } - - for (n = 0; n < udpif->n_handlers; ++n) { - struct handler *handler = &udpif->handlers[n]; - - if (handler->need_signal) { - handler->need_signal = false; - ovs_mutex_lock(&handler->mutex); - xpthread_cond_signal(&handler->wake_cond); - ovs_mutex_unlock(&handler->mutex); - } - } -} - /* Calculates slow path actions for 'xout'. 'buf' must statically be * initialized with at least 128 bytes of space. */ static void compose_slow_path(struct udpif *udpif, struct xlate_out *xout, - odp_port_t odp_in_port, struct ofpbuf *buf) + struct flow *flow, odp_port_t odp_in_port, + struct ofpbuf *buf) { union user_action_cookie cookie; odp_port_t port; @@ -913,7 +817,7 @@ compose_slow_path(struct udpif *udpif, struct xlate_out *xout, port = xout->slow & (SLOW_CFM | SLOW_BFD | SLOW_LACP | SLOW_STP) ? ODPP_NONE : odp_in_port; - pid = dpif_port_get_pid(udpif->dpif, port); + pid = dpif_port_get_pid(udpif->dpif, port, flow_hash_5tuple(flow, 0)); odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, buf); } @@ -932,26 +836,23 @@ flow_miss_find(struct hmap *todo, const struct ofproto_dpif *ofproto, return NULL; } -static void -handle_upcalls(struct handler *handler, struct list *upcalls) +/* Reads and classifies upcalls. Returns the number of upcalls successfully + * read. */ +static size_t +read_upcalls(struct handler *handler, + struct upcall upcalls[FLOW_MISS_MAX_BATCH], + struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH], + struct hmap *misses) { - struct hmap misses = HMAP_INITIALIZER(&misses); struct udpif *udpif = handler->udpif; + size_t i; + size_t n_misses = 0; + size_t n_upcalls = 0; - struct flow_miss miss_buf[FLOW_MISS_MAX_BATCH]; - struct dpif_op *opsp[FLOW_MISS_MAX_BATCH * 2]; - struct dpif_op ops[FLOW_MISS_MAX_BATCH * 2]; - struct flow_miss *miss, *next_miss; - struct upcall *upcall, *next; - size_t n_misses, n_ops, i; - unsigned int flow_limit; - bool fail_open, may_put; - enum upcall_type type; - - atomic_read(&udpif->flow_limit, &flow_limit); - may_put = udpif_get_n_flows(udpif) < flow_limit; - - /* Extract the flow from each upcall. Construct in 'misses' a hash table + /* + * Try reading FLOW_MISS_MAX_BATCH upcalls from dpif. + * + * Extract the flow from each upcall. Construct in 'misses' a hash table * that maps each unique flow to a 'struct flow_miss'. * * Most commonly there is a single packet per flow_miss, but there are @@ -970,19 +871,31 @@ handle_upcalls(struct handler *handler, struct list *upcalls) * other end of the connection, which gives OVS a chance to set up a * datapath flow.) */ - n_misses = 0; - LIST_FOR_EACH_SAFE (upcall, next, list_node, upcalls) { - struct dpif_upcall *dupcall = &upcall->dpif_upcall; + for (i = 0; i < FLOW_MISS_MAX_BATCH; i++) { + struct upcall *upcall = &upcalls[n_upcalls]; struct flow_miss *miss = &miss_buf[n_misses]; - struct ofpbuf *packet = &dupcall->packet; + struct dpif_upcall *dupcall; + struct ofpbuf *packet; struct flow_miss *existing_miss; struct ofproto_dpif *ofproto; struct dpif_sflow *sflow; struct dpif_ipfix *ipfix; - odp_port_t odp_in_port; struct flow flow; + enum upcall_type type; + odp_port_t odp_in_port; int error; + ofpbuf_use_stub(&upcall->upcall_buf, upcall->upcall_stub, + sizeof upcall->upcall_stub); + error = dpif_recv(udpif->dpif, handler->handler_id, + &upcall->dpif_upcall, &upcall->upcall_buf); + if (error) { + ofpbuf_uninit(&upcall->upcall_buf); + break; + } + + dupcall = &upcall->dpif_upcall; + packet = &dupcall->packet; error = xlate_receive(udpif->backer, packet, dupcall->key, dupcall->key_len, &flow, &ofproto, &ipfix, &sflow, NULL, &odp_in_port); @@ -1000,22 +913,20 @@ handle_upcalls(struct handler *handler, struct list *upcalls) dupcall->key, dupcall->key_len, NULL, 0, NULL, 0, NULL); } - list_remove(&upcall->list_node); - upcall_destroy(upcall); - continue; + goto destroy_upcall; } type = classify_upcall(upcall); if (type == MISS_UPCALL) { uint32_t hash; - struct pkt_metadata md = PKT_METADATA_INITIALIZER_FLOW(&flow); + struct pkt_metadata md = pkt_metadata_from_flow(&flow); flow_extract(packet, &md, &miss->flow); hash = flow_hash(&miss->flow, 0); - existing_miss = flow_miss_find(&misses, ofproto, &miss->flow, + existing_miss = flow_miss_find(misses, ofproto, &miss->flow, hash); if (!existing_miss) { - hmap_insert(&misses, &miss->hmap_node, hash); + hmap_insert(misses, &miss->hmap_node, hash); miss->ofproto = ofproto; miss->key = dupcall->key; miss->key_len = dupcall->key_len; @@ -1026,16 +937,16 @@ handle_upcalls(struct handler *handler, struct list *upcalls) miss->stats.tcp_flags = 0; miss->odp_in_port = odp_in_port; miss->put = false; - n_misses++; } else { miss = existing_miss; } miss->stats.tcp_flags |= ntohs(miss->flow.tcp_flags); - miss->stats.n_bytes += packet->size; + miss->stats.n_bytes += ofpbuf_size(packet); miss->stats.n_packets++; upcall->flow_miss = miss; + n_upcalls++; continue; } @@ -1082,10 +993,29 @@ handle_upcalls(struct handler *handler, struct list *upcalls) dpif_ipfix_unref(ipfix); dpif_sflow_unref(sflow); - list_remove(&upcall->list_node); - upcall_destroy(upcall); +destroy_upcall: + ofpbuf_uninit(&upcall->dpif_upcall.packet); + ofpbuf_uninit(&upcall->upcall_buf); } + return n_upcalls; +} + +static void +handle_upcalls(struct handler *handler, struct hmap *misses, + struct upcall *upcalls, size_t n_upcalls) +{ + struct udpif *udpif = handler->udpif; + struct dpif_op *opsp[FLOW_MISS_MAX_BATCH * 2]; + struct dpif_op ops[FLOW_MISS_MAX_BATCH * 2]; + struct flow_miss *miss; + size_t n_ops, i; + unsigned int flow_limit; + bool fail_open, may_put; + + atomic_read(&udpif->flow_limit, &flow_limit); + may_put = udpif_get_n_flows(udpif) < flow_limit; + /* Initialize each 'struct flow_miss's ->xout. * * We do this per-flow_miss rather than per-packet because, most commonly, @@ -1094,7 +1024,7 @@ handle_upcalls(struct handler *handler, struct list *upcalls) * We can't do this in the previous loop because we need the TCP flags for * all the packets in each miss. */ fail_open = false; - HMAP_FOR_EACH (miss, hmap_node, &misses) { + HMAP_FOR_EACH (miss, hmap_node, misses) { struct xlate_in xin; xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL, @@ -1126,7 +1056,8 @@ handle_upcalls(struct handler *handler, struct list *upcalls) * The loop fills 'ops' with an array of operations to execute in the * datapath. */ n_ops = 0; - LIST_FOR_EACH (upcall, list_node, upcalls) { + for (i = 0; i < n_upcalls; i++) { + struct upcall *upcall = &upcalls[i]; struct flow_miss *miss = upcall->flow_miss; struct ofpbuf *packet = &upcall->dpif_upcall.packet; struct dpif_op *op; @@ -1153,7 +1084,7 @@ handle_upcalls(struct handler *handler, struct list *upcalls) * the packet contained no VLAN. So, we must remove the * VLAN header from the packet before trying to execute the * actions. */ - if (miss->xout.odp_actions.size) { + if (ofpbuf_size(&miss->xout.odp_actions)) { eth_pop_vlan(packet); } @@ -1193,21 +1124,22 @@ handle_upcalls(struct handler *handler, struct list *upcalls) op->u.flow_put.flags = DPIF_FP_CREATE | DPIF_FP_MODIFY; op->u.flow_put.key = miss->key; op->u.flow_put.key_len = miss->key_len; - op->u.flow_put.mask = mask.data; - op->u.flow_put.mask_len = mask.size; + op->u.flow_put.mask = ofpbuf_data(&mask); + op->u.flow_put.mask_len = ofpbuf_size(&mask); op->u.flow_put.stats = NULL; if (!miss->xout.slow) { - op->u.flow_put.actions = miss->xout.odp_actions.data; - op->u.flow_put.actions_len = miss->xout.odp_actions.size; + op->u.flow_put.actions = ofpbuf_data(&miss->xout.odp_actions); + op->u.flow_put.actions_len = ofpbuf_size(&miss->xout.odp_actions); } else { struct ofpbuf buf; ofpbuf_use_stack(&buf, miss->slow_path_buf, sizeof miss->slow_path_buf); - compose_slow_path(udpif, &miss->xout, miss->odp_in_port, &buf); - op->u.flow_put.actions = buf.data; - op->u.flow_put.actions_len = buf.size; + compose_slow_path(udpif, &miss->xout, &miss->flow, + miss->odp_in_port, &buf); + op->u.flow_put.actions = ofpbuf_data(&buf); + op->u.flow_put.actions_len = ofpbuf_size(&buf); } } @@ -1217,15 +1149,15 @@ handle_upcalls(struct handler *handler, struct list *upcalls) * upcall. */ miss->flow.vlan_tci = flow_vlan_tci; - if (miss->xout.odp_actions.size) { + if (ofpbuf_size(&miss->xout.odp_actions)) { op = &ops[n_ops++]; op->type = DPIF_OP_EXECUTE; op->u.execute.packet = packet; odp_key_to_pkt_metadata(miss->key, miss->key_len, &op->u.execute.md); - op->u.execute.actions = miss->xout.odp_actions.data; - op->u.execute.actions_len = miss->xout.odp_actions.size; + op->u.execute.actions = ofpbuf_data(&miss->xout.odp_actions); + op->u.execute.actions_len = ofpbuf_size(&miss->xout.odp_actions); op->u.execute.needs_help = (miss->xout.slow & SLOW_ACTION) != 0; } } @@ -1240,20 +1172,21 @@ handle_upcalls(struct handler *handler, struct list *upcalls) * * Copy packets before they are modified by execution. */ if (fail_open) { - LIST_FOR_EACH (upcall, list_node, upcalls) { + for (i = 0; i < n_upcalls; i++) { + struct upcall *upcall = &upcalls[i]; struct flow_miss *miss = upcall->flow_miss; struct ofpbuf *packet = &upcall->dpif_upcall.packet; struct ofproto_packet_in *pin; pin = xmalloc(sizeof *pin); - pin->up.packet = xmemdup(packet->data, packet->size); - pin->up.packet_len = packet->size; + pin->up.packet = xmemdup(ofpbuf_data(packet), ofpbuf_size(packet)); + pin->up.packet_len = ofpbuf_size(packet); pin->up.reason = OFPR_NO_MATCH; pin->up.table_id = 0; pin->up.cookie = OVS_BE64_MAX; flow_get_metadata(&miss->flow, &pin->up.fmd); pin->send_len = 0; /* Not used for flow table misses. */ - pin->generated_by_table_miss = false; + pin->miss_type = OFPROTO_PACKET_IN_NO_MISS; ofproto_dpif_send_packet_in(miss->ofproto, pin); } } @@ -1263,17 +1196,6 @@ handle_upcalls(struct handler *handler, struct list *upcalls) opsp[i] = &ops[i]; } dpif_operate(udpif->dpif, opsp, n_ops); - - HMAP_FOR_EACH_SAFE (miss, next_miss, hmap_node, &misses) { - hmap_remove(&misses, &miss->hmap_node); - xlate_out_uninit(&miss->xout); - } - hmap_destroy(&misses); - - LIST_FOR_EACH_SAFE (upcall, next, list_node, upcalls) { - list_remove(&upcall->list_node); - upcall_destroy(upcall); - } } static struct udpif_key * @@ -1301,8 +1223,10 @@ ukey_create(const struct nlattr *key, size_t key_len, long long int used) ukey->key_len = key_len; ukey->mark = false; + ukey->flow_exists = true; ukey->created = used ? used : time_msec(); memset(&ukey->stats, 0, sizeof ukey->stats); + ukey->xcache = NULL; return ukey; } @@ -1311,9 +1235,36 @@ static void ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey) { hmap_remove(&revalidator->ukeys, &ukey->hmap_node); + xlate_cache_delete(ukey->xcache); free(ukey); } +static bool +should_revalidate(uint64_t packets, long long int used) +{ + long long int metric, now, duration; + + /* Calculate the mean time between seeing these packets. If this + * exceeds the threshold, then delete the flow rather than performing + * costly revalidation for flows that aren't being hit frequently. + * + * This is targeted at situations where the dump_duration is high (~1s), + * and revalidation is triggered by a call to udpif_revalidate(). In + * these situations, revalidation of all flows causes fluctuations in the + * flow_limit due to the interaction with the dump_duration and max_idle. + * This tends to result in deletion of low-throughput flows anyway, so + * skip the revalidation and just delete those flows. */ + packets = MAX(packets, 1); + now = MAX(used, time_msec()); + duration = now - used; + metric = duration / packets; + + if (metric > 200) { + return false; + } + return true; +} + static bool revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump, struct udpif_key *ukey) @@ -1321,19 +1272,22 @@ revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump, struct ofpbuf xout_actions, *actions; uint64_t slow_path_buf[128 / 8]; struct xlate_out xout, *xoutp; + struct netflow *netflow; struct flow flow, udump_mask; struct ofproto_dpif *ofproto; struct dpif_flow_stats push; uint32_t *udump32, *xout32; odp_port_t odp_in_port; struct xlate_in xin; + long long int last_used; int error; size_t i; - bool ok; + bool may_learn, ok; ok = false; xoutp = NULL; actions = NULL; + netflow = NULL; /* If we don't need to revalidate, we can simply push the stats contained * in the udump, otherwise we'll have to get the actions so we can check @@ -1345,6 +1299,7 @@ revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump, } } + last_used = ukey->stats.used; push.used = udump->stats.used; push.tcp_flags = udump->stats.tcp_flags; push.n_packets = udump->stats.n_packets > ukey->stats.n_packets @@ -1355,20 +1310,41 @@ revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump, : 0; ukey->stats = udump->stats; + if (udump->need_revalidate && last_used + && !should_revalidate(push.n_packets, last_used)) { + ok = false; + goto exit; + } + if (!push.n_packets && !udump->need_revalidate) { ok = true; goto exit; } + may_learn = push.n_packets > 0; + if (ukey->xcache && !udump->need_revalidate) { + xlate_push_stats(ukey->xcache, may_learn, &push); + ok = true; + goto exit; + } + error = xlate_receive(udpif->backer, NULL, ukey->key, ukey->key_len, &flow, - &ofproto, NULL, NULL, NULL, &odp_in_port); + &ofproto, NULL, NULL, &netflow, &odp_in_port); if (error) { goto exit; } + if (udump->need_revalidate) { + xlate_cache_clear(ukey->xcache); + } + if (!ukey->xcache) { + ukey->xcache = xlate_cache_new(); + } + xlate_in_init(&xin, ofproto, &flow, NULL, push.tcp_flags, NULL); xin.resubmit_stats = push.n_packets ? &push : NULL; - xin.may_learn = push.n_packets > 0; + xin.xcache = ukey->xcache; + xin.may_learn = may_learn; xin.skip_wildcards = !udump->need_revalidate; xlate_actions(&xin, &xout); xoutp = &xout; @@ -1379,11 +1355,11 @@ revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump, } if (!xout.slow) { - ofpbuf_use_const(&xout_actions, xout.odp_actions.data, - xout.odp_actions.size); + ofpbuf_use_const(&xout_actions, ofpbuf_data(&xout.odp_actions), + ofpbuf_size(&xout.odp_actions)); } else { ofpbuf_use_stack(&xout_actions, slow_path_buf, sizeof slow_path_buf); - compose_slow_path(udpif, &xout, odp_in_port, &xout_actions); + compose_slow_path(udpif, &xout, &flow, odp_in_port, &xout_actions); } if (!ofpbuf_equal(&xout_actions, actions)) { @@ -1410,6 +1386,13 @@ revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump, ok = true; exit: + if (netflow) { + if (!ok) { + netflow_expire(netflow, &flow); + netflow_flow_clear(netflow, &flow); + } + netflow_unref(netflow); + } ofpbuf_delete(actions); xlate_out_uninit(xoutp); return ok; @@ -1467,6 +1450,13 @@ push_dump_ops(struct revalidator *revalidator, struct ofproto_dpif *ofproto; struct netflow *netflow; struct flow flow; + bool may_learn; + + may_learn = push->n_packets > 0; + if (op->ukey && op->ukey->xcache) { + xlate_push_stats(op->ukey->xcache, may_learn, push); + continue; + } if (!xlate_receive(udpif->backer, NULL, op->op.u.flow_del.key, op->op.u.flow_del.key_len, &flow, &ofproto, @@ -1476,7 +1466,7 @@ push_dump_ops(struct revalidator *revalidator, xlate_in_init(&xin, ofproto, &flow, NULL, push->tcp_flags, NULL); xin.resubmit_stats = push->n_packets ? push : NULL; - xin.may_learn = push->n_packets > 0; + xin.may_learn = may_learn; xin.skip_wildcards = true; xlate_actions_for_side_effects(&xin); @@ -1543,9 +1533,20 @@ revalidate_udumps(struct revalidator *revalidator, struct list *udumps) used = ukey->created; } + if (ukey && (ukey->mark || !ukey->flow_exists)) { + /* The flow has already been dumped. This can occasionally occur + * if the datapath is changed in the middle of a flow dump. Rather + * than perform the same work twice, skip the flow this time. */ + COVERAGE_INC(upcall_duplicate_flow); + continue; + } + if (must_del || (used && used < now - max_idle)) { struct dump_op *dop = &ops[n_ops++]; + if (ukey) { + ukey->flow_exists = false; + } dump_op_init(dop, udump->key, udump->key_len, ukey, udump); continue; } @@ -1558,8 +1559,10 @@ revalidate_udumps(struct revalidator *revalidator, struct list *udumps) ukey->mark = true; if (!revalidate_ukey(udpif, udump, ukey)) { + ukey->flow_exists = false; dpif_flow_del(udpif->dpif, udump->key, udump->key_len, NULL); - ukey_delete(revalidator, ukey); + /* The ukey will be cleaned up by revalidator_sweep(). + * This helps to avoid deleting the same flow twice. */ } list_remove(&udump->list_node); @@ -1586,6 +1589,8 @@ revalidator_sweep__(struct revalidator *revalidator, bool purge) HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, &revalidator->ukeys) { if (!purge && ukey->mark) { ukey->mark = false; + } else if (!ukey->flow_exists) { + ukey_delete(revalidator, ukey); } else { struct dump_op *op = &ops[n_ops++]; @@ -1636,16 +1641,6 @@ upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED, udpif->avg_n_flows, udpif->max_n_flows, flow_limit); ds_put_format(&ds, "\tdump duration : %lldms\n", udpif->dump_duration); - ds_put_char(&ds, '\n'); - for (i = 0; i < udpif->n_handlers; i++) { - struct handler *handler = &udpif->handlers[i]; - - ovs_mutex_lock(&handler->mutex); - ds_put_format(&ds, "\t%s: (upcall queue %"PRIuSIZE")\n", - handler->name, handler->n_upcalls); - ovs_mutex_unlock(&handler->mutex); - } - ds_put_char(&ds, '\n'); for (i = 0; i < n_revalidators; i++) { struct revalidator *revalidator = &udpif->revalidators[i];