X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=ofproto%2Fofproto-dpif-upcall.c;h=906bf17f157cbd3c8508fdcef36782abbd304f52;hb=73a3c4757e596ff156d40f41496a0264373e5bc4;hp=83699596803d5f45e6f196a4f982d4e464ff8de6;hpb=f8b92eb858962b75ae17519489b4cd9f29a60203;p=sliver-openvswitch.git diff --git a/ofproto/ofproto-dpif-upcall.c b/ofproto/ofproto-dpif-upcall.c index 836995968..906bf17f1 100644 --- a/ofproto/ofproto-dpif-upcall.c +++ b/ofproto/ofproto-dpif-upcall.c @@ -1,4 +1,4 @@ -/* Copyright (c) 2009, 2010, 2011, 2012, 2013 Nicira, Inc. +/* Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -32,6 +32,7 @@ #include "ofproto-dpif-ipfix.h" #include "ofproto-dpif-sflow.h" #include "ofproto-dpif-xlate.h" +#include "ovs-rcu.h" #include "packets.h" #include "poll-loop.h" #include "seq.h" @@ -44,26 +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 @@ -86,14 +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. */ @@ -102,7 +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. */ @@ -125,8 +124,12 @@ struct udpif { unsigned int avg_n_flows; /* Following fields are accessed and modified by different threads. */ - atomic_llong max_idle; /* Maximum datapath flow idle time. */ atomic_uint flow_limit; /* Datapath flow hard limit. */ + + /* n_flows_mutex prevents multiple threads updating these concurrently. */ + atomic_uint64_t n_flows; /* Number of flows in the datapath. */ + atomic_llong n_flows_timestamp; /* Last time n_flows was updated. */ + struct ovs_mutex n_flows_mutex; }; enum upcall_type { @@ -138,7 +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. */ @@ -161,8 +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 @@ -197,7 +203,6 @@ struct flow_miss { struct ofproto_dpif *ofproto; struct flow flow; - enum odp_key_fitness key_fitness; const struct nlattr *key; size_t key_len; enum dpif_upcall_type upcall_type; @@ -208,28 +213,37 @@ struct flow_miss { struct odputil_keybuf mask_buf; struct xlate_out xout; -}; -static void upcall_destroy(struct upcall *); + bool put; +}; 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(const struct udpif *); +static uint64_t udpif_get_n_flows(struct udpif *); static void revalidate_udumps(struct revalidator *, struct list *udumps); static void revalidator_sweep(struct revalidator *); +static void revalidator_purge(struct revalidator *); static void upcall_unixctl_show(struct unixctl_conn *conn, int argc, const char *argv[], void *aux); static void upcall_unixctl_disable_megaflows(struct unixctl_conn *, int argc, const char *argv[], void *aux); static void upcall_unixctl_enable_megaflows(struct unixctl_conn *, int argc, const char *argv[], void *aux); +static void upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, int argc, + const char *argv[], void *aux); static void ukey_delete(struct revalidator *, struct udpif_key *); static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true); @@ -247,18 +261,22 @@ udpif_create(struct dpif_backer *backer, struct dpif *dpif) upcall_unixctl_disable_megaflows, NULL); unixctl_command_register("upcall/enable-megaflows", "", 0, 0, upcall_unixctl_enable_megaflows, NULL); + unixctl_command_register("upcall/set-flow-limit", "", 1, 1, + upcall_unixctl_set_flow_limit, NULL); ovsthread_once_done(&once); } udpif->dpif = dpif; udpif->backer = backer; - atomic_init(&udpif->max_idle, 5000); atomic_init(&udpif->flow_limit, MIN(ofproto_flow_limit, 10000)); udpif->secret = random_uint32(); udpif->reval_seq = seq_create(); udpif->dump_seq = seq_create(); latch_init(&udpif->exit_latch); list_push_back(&all_udpifs, &udpif->list_node); + atomic_init(&udpif->n_flows, 0); + atomic_init(&udpif->n_flows_timestamp, LLONG_MIN); + ovs_mutex_init(&udpif->n_flows_mutex); return udpif; } @@ -266,28 +284,22 @@ udpif_create(struct dpif_backer *backer, struct dpif *dpif) void udpif_destroy(struct udpif *udpif) { - udpif_set_threads(udpif, 0, 0); - udpif_flush(); + udpif_stop_threads(udpif); list_remove(&udpif->list_node); latch_destroy(&udpif->exit_latch); seq_destroy(udpif->reval_seq); seq_destroy(udpif->dump_seq); + ovs_mutex_destroy(&udpif->n_flows_mutex); 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) { - /* 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); @@ -295,9 +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); } @@ -311,12 +320,10 @@ 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]; struct udpif_flow_dump *udump, *next_udump; - struct udpif_key *ukey, *next_ukey; LIST_FOR_EACH_SAFE (udump, next_udump, list_node, &revalidator->udumps) { @@ -324,10 +331,9 @@ udpif_set_threads(struct udpif *udpif, size_t n_handlers, free(udump); } - HMAP_FOR_EACH_SAFE (ukey, next_ukey, hmap_node, - &revalidator->ukeys) { - ukey_delete(revalidator, ukey); - } + /* Delete ukeys, and delete all flows from the datapath to prevent + * double-counting stats. */ + revalidator_purge(revalidator); hmap_destroy(&revalidator->ukeys); ovs_mutex_destroy(&revalidator->mutex); @@ -335,17 +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); @@ -357,9 +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; @@ -370,10 +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); } @@ -391,11 +390,61 @@ 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(); +} + +/* Waits for all ongoing upcall translations to complete. This ensures that + * there are no transient references to any removed ofprotos (or other + * objects). In particular, this should be called after an ofproto is removed + * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */ +void +udpif_synchronize(struct udpif *udpif) +{ + /* This is stronger than necessary. It would be sufficient to ensure + * (somehow) that each handler and revalidator thread had passed through + * its main loop once. */ + size_t n_handlers = udpif->n_handlers; + size_t n_revalidators = udpif->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 * xlate_actions() results invalid. */ void @@ -418,16 +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++) { @@ -443,53 +485,56 @@ udpif_get_memory_usage(struct udpif *udpif, struct simap *usage) } } -/* Removes all flows from all datapaths. */ +/* Remove flows from a single datapath. */ void -udpif_flush(void) +udpif_flush(struct udpif *udpif) { - struct udpif *udpif; + size_t n_handlers, n_revalidators; - LIST_FOR_EACH (udpif, list_node, &all_udpifs) { - dpif_flow_flush(udpif->dpif); - } + n_handlers = udpif->n_handlers; + n_revalidators = udpif->n_revalidators; + + ovsrcu_quiesce_start(); + + udpif_stop_threads(udpif); + dpif_flow_flush(udpif->dpif); + udpif_start_threads(udpif, n_handlers, n_revalidators); + + ovsrcu_quiesce_end(); } - -/* Destroys and deallocates 'upcall'. */ + +/* Removes all flows from all datapaths. */ static void -upcall_destroy(struct upcall *upcall) +udpif_flush_all_datapaths(void) { - if (upcall) { - ofpbuf_uninit(&upcall->dpif_upcall.packet); - ofpbuf_uninit(&upcall->upcall_buf); - free(upcall); + struct udpif *udpif; + + LIST_FOR_EACH (udpif, list_node, &all_udpifs) { + udpif_flush(udpif); } } + static uint64_t -udpif_get_n_flows(const struct udpif *udpif) -{ - struct dpif_dp_stats stats; - - dpif_get_dp_stats(udpif->dpif, &stats); - return stats.n_flows; -} - -/* The dispatcher thread is responsible for receiving upcalls from the kernel, - * assigning them to a upcall_handler thread. */ -static void * -udpif_dispatcher(void *arg) +udpif_get_n_flows(struct udpif *udpif) { - 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(); + long long int time, now; + uint64_t flow_count; + + now = time_msec(); + atomic_read(&udpif->n_flows_timestamp, &time); + if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) { + struct dpif_dp_stats stats; + + atomic_store(&udpif->n_flows_timestamp, now); + dpif_get_dp_stats(udpif->dpif, &stats); + flow_count = stats.n_flows; + atomic_store(&udpif->n_flows, flow_count); + ovs_mutex_unlock(&udpif->n_flows_mutex); + } else { + atomic_read(&udpif->n_flows, &flow_count); } - - return NULL; + return flow_count; } static void * @@ -505,10 +550,11 @@ udpif_flow_dumper(void *arg) struct dpif_flow_dump dump; size_t key_len, mask_len; unsigned int flow_limit; - long long int max_idle; bool need_revalidate; uint64_t reval_seq; size_t n_flows, i; + int error; + void *state = NULL; reval_seq = seq_read(udpif->reval_seq); need_revalidate = udpif->last_reval_seq != reval_seq; @@ -518,22 +564,15 @@ udpif_flow_dumper(void *arg) udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows); udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2; - atomic_read(&udpif->flow_limit, &flow_limit); - if (n_flows < flow_limit / 8) { - max_idle = 5000; - } else if (n_flows < flow_limit / 4) { - max_idle = 2000; - } else if (n_flows < flow_limit / 2) { - max_idle = 1000; - } else { - max_idle = 500; - } - atomic_store(&udpif->max_idle, max_idle); - start_time = time_msec(); - dpif_flow_dump_start(&dump, udpif->dpif); - while (dpif_flow_dump_next(&dump, &key, &key_len, &mask, &mask_len, - NULL, NULL, &stats) + error = dpif_flow_dump_start(&dump, udpif->dpif); + if (error) { + VLOG_INFO("Failed to start flow dump (%s)", ovs_strerror(error)); + goto skip; + } + dpif_flow_dump_state_init(udpif->dpif, &state); + while (dpif_flow_dump_next(&dump, state, &key, &key_len, + &mask, &mask_len, NULL, NULL, &stats) && !latch_is_set(&udpif->exit_latch)) { struct udpif_flow_dump *udump = xmalloc(sizeof *udump); struct revalidator *revalidator; @@ -564,6 +603,7 @@ udpif_flow_dumper(void *arg) xpthread_cond_signal(&revalidator->wake_cond); ovs_mutex_unlock(&revalidator->mutex); } + dpif_flow_dump_state_uninit(udpif->dpif, state); dpif_flow_dump_done(&dump); /* Let all the revalidators finish and garbage collect. */ @@ -587,8 +627,9 @@ udpif_flow_dumper(void *arg) ovs_mutex_unlock(&revalidator->mutex); } - duration = time_msec() - start_time; + duration = MAX(time_msec() - start_time, 1); udpif->dump_duration = duration; + atomic_read(&udpif->flow_limit, &flow_limit); if (duration > 2000) { flow_limit /= duration / 1000; } else if (duration > 1300) { @@ -605,7 +646,8 @@ udpif_flow_dumper(void *arg) duration); } - poll_timer_wait_until(start_time + MIN(max_idle, 500)); +skip: + poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500)); seq_wait(udpif->reval_seq, udpif->last_reval_seq); latch_wait(&udpif->exit_latch); poll_block(); @@ -614,46 +656,47 @@ udpif_flow_dumper(void *arg) 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); - for (;;) { - struct list misses = LIST_INITIALIZER(&misses); - size_t i; - - ovs_mutex_lock(&handler->mutex); - - if (latch_is_set(&handler->udpif->exit_latch)) { - ovs_mutex_unlock(&handler->mutex); - return NULL; - } - - if (!handler->n_upcalls) { - ovs_mutex_cond_wait(&handler->wake_cond, &handler->mutex); - } + while (!latch_is_set(&handler->udpif->exit_latch)) { + 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; } static void * @@ -737,16 +780,16 @@ classify_upcall(const struct upcall *upcall) } memset(&cookie, 0, sizeof cookie); memcpy(&cookie, nl_attr_get(dpif_upcall->userdata), userdata_len); - if (userdata_len == sizeof cookie.sflow + if (userdata_len == MAX(8, sizeof cookie.sflow) && cookie.type == USER_ACTION_COOKIE_SFLOW) { return SFLOW_UPCALL; - } else if (userdata_len == sizeof cookie.slow_path + } else if (userdata_len == MAX(8, sizeof cookie.slow_path) && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) { return MISS_UPCALL; - } else if (userdata_len == sizeof cookie.flow_sample + } else if (userdata_len == MAX(8, sizeof cookie.flow_sample) && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) { return FLOW_SAMPLE_UPCALL; - } else if (userdata_len == sizeof cookie.ipfix + } else if (userdata_len == MAX(8, sizeof cookie.ipfix) && cookie.type == USER_ACTION_COOKIE_IPFIX) { return IPFIX_UPCALL; } else { @@ -756,98 +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; @@ -860,7 +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); } @@ -879,26 +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 @@ -917,21 +871,33 @@ 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, &miss->key_fitness, + dupcall->key_len, &flow, &ofproto, &ipfix, &sflow, NULL, &odp_in_port); if (error) { if (error == ENODEV) { @@ -947,23 +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_from_flow(&flow); - flow_extract(packet, flow.skb_priority, flow.pkt_mark, - &flow.tunnel, &flow.in_port, &miss->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; @@ -973,16 +936,17 @@ handle_upcalls(struct handler *handler, struct list *upcalls) miss->stats.used = time_msec(); 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 |= packet_get_tcp_flags(packet, &miss->flow); - miss->stats.n_bytes += packet->size; + miss->stats.tcp_flags |= ntohs(miss->flow.tcp_flags); + miss->stats.n_bytes += ofpbuf_size(packet); miss->stats.n_packets++; upcall->flow_miss = miss; + n_upcalls++; continue; } @@ -1029,10 +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, @@ -1041,7 +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, @@ -1073,12 +1056,16 @@ 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 ofpbuf mask; struct dpif_op *op; - bool megaflow; + ovs_be16 flow_vlan_tci; + + /* Save a copy of flow.vlan_tci in case it is changed to + * generate proper mega flow masks for VLAN splinter flows. */ + flow_vlan_tci = miss->flow.vlan_tci; if (miss->xout.slow) { struct xlate_in xin; @@ -1087,58 +1074,90 @@ handle_upcalls(struct handler *handler, struct list *upcalls) xlate_actions_for_side_effects(&xin); } - atomic_read(&enable_megaflows, &megaflow); - ofpbuf_use_stack(&mask, &miss->mask_buf, sizeof miss->mask_buf); - if (megaflow) { - odp_flow_key_from_mask(&mask, &miss->xout.wc.masks, &miss->flow, - UINT32_MAX); + 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. */ + if (ofpbuf_size(&miss->xout.odp_actions)) { + eth_pop_vlan(packet); + } + + /* Remove the flow vlan tags inserted by vlan splinter logic + * to ensure megaflow masks generated match the data path flow. */ + miss->flow.vlan_tci = 0; } - if (may_put) { + /* Do not install a flow into the datapath if: + * + * - The datapath already has too many flows. + * + * - An earlier iteration of this loop already put the same flow. + * + * - We received this packet via some flow installed in the kernel + * already. */ + if (may_put + && !miss->put + && upcall->dpif_upcall.type == DPIF_UC_MISS) { + struct ofpbuf mask; + bool megaflow; + + miss->put = true; + + atomic_read(&enable_megaflows, &megaflow); + ofpbuf_use_stack(&mask, &miss->mask_buf, sizeof miss->mask_buf); + if (megaflow) { + size_t max_mpls; + + max_mpls = ofproto_dpif_get_max_mpls_depth(miss->ofproto); + odp_flow_key_from_mask(&mask, &miss->xout.wc.masks, + &miss->flow, UINT32_MAX, max_mpls); + } + op = &ops[n_ops++]; op->type = DPIF_OP_FLOW_PUT; 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); } } - if (miss->xout.odp_actions.size) { - 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); - } + /* + * The 'miss' may be shared by multiple upcalls. Restore + * the saved flow vlan_tci field before processing the next + * upcall. */ + miss->flow.vlan_tci = flow_vlan_tci; + + if (ofpbuf_size(&miss->xout.odp_actions)) { op = &ops[n_ops++]; op->type = DPIF_OP_EXECUTE; - op->u.execute.key = miss->key; - op->u.execute.key_len = miss->key_len; op->u.execute.packet = packet; - op->u.execute.actions = miss->xout.odp_actions.data; - op->u.execute.actions_len = miss->xout.odp_actions.size; + odp_key_to_pkt_metadata(miss->key, miss->key_len, + &op->u.execute.md); + 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; } } @@ -1153,20 +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); } } @@ -1176,17 +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 * @@ -1204,13 +1213,58 @@ ukey_lookup(struct revalidator *revalidator, struct udpif_flow_dump *udump) return NULL; } +static struct udpif_key * +ukey_create(const struct nlattr *key, size_t key_len, long long int used) +{ + struct udpif_key *ukey = xmalloc(sizeof *ukey); + + ukey->key = (struct nlattr *) &ukey->key_buf; + memcpy(&ukey->key_buf, key, key_len); + 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; +} + 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) @@ -1218,19 +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 @@ -1242,6 +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 @@ -1252,20 +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, - NULL, &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; @@ -1276,11 +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)) { @@ -1307,35 +1386,135 @@ 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; } +struct dump_op { + struct udpif_key *ukey; + struct udpif_flow_dump *udump; + struct dpif_flow_stats stats; /* Stats for 'op'. */ + struct dpif_op op; /* Flow del operation. */ +}; + static void -revalidate_udumps(struct revalidator *revalidator, struct list *udumps) +dump_op_init(struct dump_op *op, const struct nlattr *key, size_t key_len, + struct udpif_key *ukey, struct udpif_flow_dump *udump) +{ + op->ukey = ukey; + op->udump = udump; + op->op.type = DPIF_OP_FLOW_DEL; + op->op.u.flow_del.key = key; + op->op.u.flow_del.key_len = key_len; + op->op.u.flow_del.stats = &op->stats; +} + +static void +push_dump_ops(struct revalidator *revalidator, + struct dump_op *ops, size_t n_ops) { struct udpif *udpif = revalidator->udpif; + struct dpif_op *opsp[REVALIDATE_MAX_BATCH]; + size_t i; - struct { - struct dpif_flow_stats ukey_stats; /* Stats stored in the ukey. */ - struct dpif_flow_stats stats; /* Stats for 'op'. */ - struct dpif_op op; /* Flow del operation. */ - } ops[REVALIDATE_MAX_BATCH]; + ovs_assert(n_ops <= REVALIDATE_MAX_BATCH); + for (i = 0; i < n_ops; i++) { + opsp[i] = &ops[i].op; + } + dpif_operate(udpif->dpif, opsp, n_ops); - struct dpif_op *opsp[REVALIDATE_MAX_BATCH]; + for (i = 0; i < n_ops; i++) { + struct dump_op *op = &ops[i]; + struct dpif_flow_stats *push, *stats, push_buf; + + stats = op->op.u.flow_del.stats; + if (op->ukey) { + push = &push_buf; + push->used = MAX(stats->used, op->ukey->stats.used); + push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags; + push->n_packets = stats->n_packets - op->ukey->stats.n_packets; + push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes; + } else { + push = stats; + } + + if (push->n_packets || netflow_exists()) { + 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, + NULL, NULL, &netflow, NULL)) { + struct xlate_in xin; + + xlate_in_init(&xin, ofproto, &flow, NULL, push->tcp_flags, + NULL); + xin.resubmit_stats = push->n_packets ? push : NULL; + xin.may_learn = may_learn; + xin.skip_wildcards = true; + xlate_actions_for_side_effects(&xin); + + if (netflow) { + netflow_expire(netflow, &flow); + netflow_flow_clear(netflow, &flow); + netflow_unref(netflow); + } + } + } + } + + for (i = 0; i < n_ops; i++) { + struct udpif_key *ukey; + + /* If there's a udump, this ukey came directly from a datapath flow + * dump. Sometimes a datapath can send duplicates in flow dumps, in + * which case we wouldn't want to double-free a ukey, so avoid that by + * looking up the ukey again. + * + * If there's no udump then we know what we're doing. */ + ukey = (ops[i].udump + ? ukey_lookup(revalidator, ops[i].udump) + : ops[i].ukey); + if (ukey) { + ukey_delete(revalidator, ukey); + } + } +} + +static void +revalidate_udumps(struct revalidator *revalidator, struct list *udumps) +{ + struct udpif *udpif = revalidator->udpif; + + struct dump_op ops[REVALIDATE_MAX_BATCH]; struct udpif_flow_dump *udump, *next_udump; - size_t n_ops, i, n_flows; + size_t n_ops, n_flows; unsigned int flow_limit; long long int max_idle; bool must_del; - atomic_read(&udpif->max_idle, &max_idle); atomic_read(&udpif->flow_limit, &flow_limit); n_flows = udpif_get_n_flows(udpif); must_del = false; + max_idle = ofproto_max_idle; if (n_flows > flow_limit) { must_del = n_flows > 2 * flow_limit; max_idle = 100; @@ -1354,92 +1533,43 @@ revalidate_udumps(struct revalidator *revalidator, struct list *udumps) used = ukey->created; } - if (must_del || (used && used < now - max_idle)) { - struct dpif_flow_stats *ukey_stats = &ops[n_ops].ukey_stats; - struct dpif_op *op = &ops[n_ops].op; + 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; + } - op->type = DPIF_OP_FLOW_DEL; - op->u.flow_del.key = udump->key; - op->u.flow_del.key_len = udump->key_len; - op->u.flow_del.stats = &ops[n_ops].stats; - n_ops++; + if (must_del || (used && used < now - max_idle)) { + struct dump_op *dop = &ops[n_ops++]; if (ukey) { - *ukey_stats = ukey->stats; - ukey_delete(revalidator, ukey); - } else { - memset(ukey_stats, 0, sizeof *ukey_stats); + ukey->flow_exists = false; } - + dump_op_init(dop, udump->key, udump->key_len, ukey, udump); continue; } if (!ukey) { - ukey = xmalloc(sizeof *ukey); - - ukey->key = (struct nlattr *) &ukey->key_buf; - memcpy(ukey->key, udump->key, udump->key_len); - ukey->key_len = udump->key_len; - - ukey->created = used ? used : now; - memset(&ukey->stats, 0, sizeof ukey->stats); - - ukey->mark = false; - + ukey = ukey_create(udump->key, udump->key_len, used); hmap_insert(&revalidator->ukeys, &ukey->hmap_node, udump->key_hash); } 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); free(udump); } - for (i = 0; i < n_ops; i++) { - opsp[i] = &ops[i].op; - } - dpif_operate(udpif->dpif, opsp, n_ops); - - for (i = 0; i < n_ops; i++) { - struct dpif_flow_stats push, *stats, *ukey_stats; - - ukey_stats = &ops[i].ukey_stats; - stats = ops[i].op.u.flow_del.stats; - push.used = MAX(stats->used, ukey_stats->used); - push.tcp_flags = stats->tcp_flags | ukey_stats->tcp_flags; - push.n_packets = stats->n_packets - ukey_stats->n_packets; - push.n_bytes = stats->n_bytes - ukey_stats->n_bytes; - - if (push.n_packets || netflow_exists()) { - struct ofproto_dpif *ofproto; - struct netflow *netflow; - struct flow flow; - - if (!xlate_receive(udpif->backer, NULL, ops[i].op.u.flow_del.key, - ops[i].op.u.flow_del.key_len, &flow, NULL, - &ofproto, NULL, NULL, &netflow, NULL)) { - struct xlate_in xin; - - 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.skip_wildcards = true; - xlate_actions_for_side_effects(&xin); - - if (netflow) { - netflow_expire(netflow, &flow); - netflow_flow_clear(netflow, &flow); - netflow_unref(netflow); - } - } - } - } + push_dump_ops(revalidator, ops, n_ops); LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) { list_remove(&udump->list_node); @@ -1448,17 +1578,48 @@ revalidate_udumps(struct revalidator *revalidator, struct list *udumps) } static void -revalidator_sweep(struct revalidator *revalidator) +revalidator_sweep__(struct revalidator *revalidator, bool purge) { + struct dump_op ops[REVALIDATE_MAX_BATCH]; struct udpif_key *ukey, *next; + size_t n_ops; + + n_ops = 0; HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, &revalidator->ukeys) { - if (ukey->mark) { + if (!purge && ukey->mark) { ukey->mark = false; - } else { + } else if (!ukey->flow_exists) { ukey_delete(revalidator, ukey); + } else { + struct dump_op *op = &ops[n_ops++]; + + /* If we have previously seen a flow in the datapath, but didn't + * see it during the most recent dump, delete it. This allows us + * to clean up the ukey and keep the statistics consistent. */ + dump_op_init(op, ukey->key, ukey->key_len, ukey, NULL); + if (n_ops == REVALIDATE_MAX_BATCH) { + push_dump_ops(revalidator, ops, n_ops); + n_ops = 0; + } } } + + if (n_ops) { + push_dump_ops(revalidator, ops, n_ops); + } +} + +static void +revalidator_sweep(struct revalidator *revalidator) +{ + revalidator_sweep__(revalidator, false); +} + +static void +revalidator_purge(struct revalidator *revalidator) +{ + revalidator_sweep__(revalidator, true); } static void @@ -1470,29 +1631,16 @@ upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED, LIST_FOR_EACH (udpif, list_node, &all_udpifs) { unsigned int flow_limit; - long long int max_idle; size_t i; atomic_read(&udpif->flow_limit, &flow_limit); - atomic_read(&udpif->max_idle, &max_idle); ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif)); ds_put_format(&ds, "\tflows : (current %"PRIu64")" " (avg %u) (max %u) (limit %u)\n", udpif_get_n_flows(udpif), udpif->avg_n_flows, udpif->max_n_flows, flow_limit); - ds_put_format(&ds, "\tmax idle : %lldms\n", max_idle); 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]; @@ -1522,7 +1670,7 @@ upcall_unixctl_disable_megaflows(struct unixctl_conn *conn, void *aux OVS_UNUSED) { atomic_store(&enable_megaflows, false); - udpif_flush(); + udpif_flush_all_datapaths(); unixctl_command_reply(conn, "megaflows disabled"); } @@ -1537,6 +1685,28 @@ upcall_unixctl_enable_megaflows(struct unixctl_conn *conn, void *aux OVS_UNUSED) { atomic_store(&enable_megaflows, true); - udpif_flush(); + udpif_flush_all_datapaths(); unixctl_command_reply(conn, "megaflows enabled"); } + +/* Set the flow limit. + * + * This command is only needed for advanced debugging, so it's not + * documented in the man page. */ +static void +upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, + int argc OVS_UNUSED, + const char *argv[] OVS_UNUSED, + void *aux OVS_UNUSED) +{ + struct ds ds = DS_EMPTY_INITIALIZER; + struct udpif *udpif; + unsigned int flow_limit = atoi(argv[1]); + + LIST_FOR_EACH (udpif, list_node, &all_udpifs) { + atomic_store(&udpif->flow_limit, flow_limit); + } + ds_put_format(&ds, "set flow_limit to %u\n", flow_limit); + unixctl_command_reply(conn, ds_cstr(&ds)); + ds_destroy(&ds); +}