X-Git-Url: http://git.cascardo.eti.br/?a=blobdiff_plain;f=ofproto%2Fofproto-dpif-upcall.c;h=1f9c484e5242c17be4b1e38a625d2f061d7fed46;hb=10ab94fef2fc44902c1b16e06e4f4edcc3da7583;hp=1e0c12ce41243a1f099255d1257068f2d1538f16;hpb=73e141f977163d93e60c30a177a99c44ecf72fe2;p=cascardo%2Fovs.git diff --git a/ofproto/ofproto-dpif-upcall.c b/ofproto/ofproto-dpif-upcall.c index 1e0c12ce4..1f9c484e5 100644 --- a/ofproto/ofproto-dpif-upcall.c +++ b/ofproto/ofproto-dpif-upcall.c @@ -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" @@ -39,61 +40,49 @@ #include "vlog.h" #define MAX_QUEUE_LENGTH 512 -#define FLOW_MISS_MAX_BATCH 50 +#define UPCALL_MAX_BATCH 64 #define REVALIDATE_MAX_BATCH 50 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall); -COVERAGE_DEFINE(upcall_queue_overflow); +COVERAGE_DEFINE(dumped_duplicate_flow); +COVERAGE_DEFINE(dumped_new_flow); +COVERAGE_DEFINE(revalidate_missed_dp_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 - * thread, updates OpenFlow statistics, and updates or removes the kernel flow - * as necessary. */ +/* A thread that processes datapath flows, updates OpenFlow statistics, and + * updates or removes them if necessary. */ struct revalidator { struct udpif *udpif; /* Parent udpif. */ - char *name; /* Thread name. */ - pthread_t thread; /* Thread ID. */ - struct hmap ukeys; /* Datapath flow keys. */ - - uint64_t dump_seq; - - struct ovs_mutex mutex; /* Mutex guarding the following. */ - pthread_cond_t wake_cond; - struct list udumps OVS_GUARDED; /* Unprocessed udumps. */ - size_t n_udumps OVS_GUARDED; /* Number of unprocessed udumps. */ + unsigned int id; /* ovsthread_id_self(). */ + struct hmap *ukeys; /* Points into udpif->ukeys for this + revalidator. Used for GC phase. */ }; /* An upcall handler for ofproto_dpif. * - * udpif has two logically separate pieces: + * udpif keeps records of two kind of logically separate units: * - * - A "dispatcher" thread that reads upcalls from the kernel and dispatches - * them to one of several "handler" threads (see struct handler). + * upcall handling + * --------------- * - * - A "flow_dumper" thread that reads the kernel flow table and dispatches - * flows to one of several "revalidator" threads (see struct - * revalidator). */ + * - An array of 'struct handler's for upcall handling and flow + * installation. + * + * flow revalidation + * ----------------- + * + * - Revalidation threads which read the datapath flow table and maintains + * them. + */ struct udpif { struct list list_node; /* In all_udpifs list. */ @@ -102,31 +91,51 @@ 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. */ size_t n_handlers; struct revalidator *revalidators; /* Flow revalidators. */ size_t n_revalidators; - uint64_t last_reval_seq; /* 'reval_seq' at last revalidation. */ - struct seq *reval_seq; /* Incremented to force revalidation. */ - - struct seq *dump_seq; /* Increments each dump iteration. */ - struct latch exit_latch; /* Tells child threads to exit. */ + /* Revalidation. */ + struct seq *reval_seq; /* Incremented to force revalidation. */ + bool need_revalidate; /* As indicated by 'reval_seq'. */ + bool reval_exit; /* Set by leader on 'exit_latch. */ + struct ovs_barrier reval_barrier; /* Barrier used by revalidators. */ + struct dpif_flow_dump *dump; /* DPIF flow dump state. */ long long int dump_duration; /* Duration of the last flow dump. */ + struct seq *dump_seq; /* Increments each dump iteration. */ + + /* There are 'n_revalidators' ukey hmaps. Each revalidator retains a + * reference to one of these for garbage collection. + * + * During the flow dump phase, revalidators insert into these with a random + * distribution. During the garbage collection phase, each revalidator + * takes care of garbage collecting one of these hmaps. */ + struct { + struct ovs_mutex mutex; /* Guards the following. */ + struct hmap hmap OVS_GUARDED; /* Datapath flow keys. */ + } *ukeys; /* Datapath flow statistics. */ unsigned int max_n_flows; 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_uint 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; + + /* Following fields are accessed and modified only from the main thread. */ + struct unixctl_conn **conns; /* Connections waiting on dump_seq. */ + uint64_t conn_seq; /* Corresponds to 'dump_seq' when + conns[n_conns-1] was stored. */ + size_t n_conns; /* Number of connections waiting. */ }; enum upcall_type { @@ -138,101 +147,112 @@ 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. */ - struct dpif_upcall dpif_upcall; /* As returned by dpif_recv() */ - struct ofpbuf upcall_buf; /* Owns some data in 'dpif_upcall'. */ - uint64_t upcall_stub[512 / 8]; /* Buffer to reduce need for malloc(). */ -}; - -/* 'udpif_key's are responsible for tracking the little bit of state udpif - * needs to do flow expiration which can't be pulled directly from the - * datapath. They are owned, created by, maintained, and destroyed by a single - * revalidator making them easy to efficiently handle with multiple threads. */ -struct udpif_key { - struct hmap_node hmap_node; /* In parent revalidator 'ukeys' map. */ - - struct nlattr *key; /* Datapath flow key. */ - size_t key_len; /* Length of 'key'. */ + struct ofproto_dpif *ofproto; /* Parent ofproto. */ - struct dpif_flow_stats stats; /* Stats at most recent flow dump. */ - long long int created; /* Estimation of creation time. */ + /* The flow and packet are only required to be constant when using + * dpif-netdev. If a modification is absolutely necessary, a const cast + * may be used with other datapaths. */ + const struct flow *flow; /* Parsed representation of the packet. */ + const struct ofpbuf *packet; /* Packet associated with this upcall. */ + ofp_port_t in_port; /* OpenFlow in port, or OFPP_NONE. */ - bool mark; /* Used by mark and sweep GC algorithm. */ + enum dpif_upcall_type type; /* Datapath type of the upcall. */ + const struct nlattr *userdata; /* Userdata for DPIF_UC_ACTION Upcalls. */ - struct odputil_keybuf key_buf; /* Memory for 'key'. */ -}; - -/* 'udpif_flow_dump's hold the state associated with one iteration in a flow - * dump operation. This is created by the flow_dumper thread and handed to the - * appropriate revalidator thread to be processed. */ -struct udpif_flow_dump { - struct list list_node; - - struct nlattr *key; /* Datapath flow key. */ - size_t key_len; /* Length of 'key'. */ - uint32_t key_hash; /* Hash of 'key'. */ + bool xout_initialized; /* True if 'xout' must be uninitialized. */ + struct xlate_out xout; /* Result of xlate_actions(). */ + struct ofpbuf put_actions; /* Actions 'put' in the fastapath. */ - struct odputil_keybuf mask_buf; - struct nlattr *mask; /* Datapath mask for 'key'. */ - size_t mask_len; /* Length of 'mask'. */ + struct dpif_ipfix *ipfix; /* IPFIX pointer or NULL. */ + struct dpif_sflow *sflow; /* SFlow pointer or NULL. */ - struct dpif_flow_stats stats; /* Stats pulled from the datapath. */ + bool vsp_adjusted; /* 'packet' and 'flow' were adjusted for + VLAN splinters if true. */ - bool need_revalidate; /* Key needs revalidation? */ - - struct odputil_keybuf key_buf; + /* Not used by the upcall callback interface. */ + const struct nlattr *key; /* Datapath flow key. */ + size_t key_len; /* Datapath flow key length. */ + const struct nlattr *out_tun_key; /* Datapath output tunnel key. */ }; -/* Flow miss batching. +/* 'udpif_key's are responsible for tracking the little bit of state udpif + * needs to do flow expiration which can't be pulled directly from the + * datapath. They may be created or maintained by any revalidator during + * the dump phase, but are owned by a single revalidator, and are destroyed + * by that revalidator during the garbage-collection phase. * - * 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. */ -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; - enum dpif_upcall_type upcall_type; - struct dpif_flow_stats stats; - odp_port_t odp_in_port; - - uint64_t slow_path_buf[128 / 8]; - struct odputil_keybuf mask_buf; + * While some elements of a udpif_key are protected by a mutex, the ukey itself + * is not. Therefore it is not safe to destroy a udpif_key except when all + * revalidators are in garbage collection phase, or they aren't running. */ +struct udpif_key { + struct hmap_node hmap_node; /* In parent revalidator 'ukeys' map. */ - struct xlate_out xout; + /* These elements are read only once created, and therefore aren't + * protected by a mutex. */ + const struct nlattr *key; /* Datapath flow key. */ + size_t key_len; /* Length of 'key'. */ - bool put; + struct ovs_mutex mutex; /* Guards the following. */ + struct dpif_flow_stats stats OVS_GUARDED; /* Last known stats.*/ + long long int created OVS_GUARDED; /* Estimate of creation time. */ + uint64_t dump_seq OVS_GUARDED; /* Tracks udpif->dump_seq. */ + bool flow_exists OVS_GUARDED; /* Ensures flows are only deleted + once. */ + + struct xlate_cache *xcache OVS_GUARDED; /* Cache for xlate entries that + * are affected by this ukey. + * Used for stats and learning.*/ + union { + struct odputil_keybuf key_buf; /* Memory for 'key'. */ + struct nlattr key_buf_nla; + }; }; -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 void *udpif_flow_dumper(void *); -static void *udpif_dispatcher(void *); +static size_t recv_upcalls(struct handler *); +static int process_upcall(struct udpif *, struct upcall *, + struct ofpbuf *odp_actions); +static void handle_upcalls(struct udpif *, 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_upcall_handler(void *); static void *udpif_revalidator(void *); -static uint64_t udpif_get_n_flows(const struct udpif *); -static void revalidate_udumps(struct revalidator *, struct list *udumps); +static unsigned long udpif_get_n_flows(struct udpif *); +static void revalidate(struct revalidator *); 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 upcall_unixctl_dump_wait(struct unixctl_conn *conn, int argc, + const char *argv[], void *aux); + +static struct udpif_key *ukey_create(const struct nlattr *key, size_t key_len, + long long int used); +static struct udpif_key *ukey_lookup(struct udpif *udpif, + const struct nlattr *key, size_t key_len, + uint32_t hash); +static bool ukey_acquire(struct udpif *udpif, const struct nlattr *key, + size_t key_len, long long int used, + struct udpif_key **result); static void ukey_delete(struct revalidator *, struct udpif_key *); +static enum upcall_type classify_upcall(enum dpif_upcall_type type, + const struct nlattr *userdata); + +static int upcall_receive(struct upcall *, const struct dpif_backer *, + const struct ofpbuf *packet, enum dpif_upcall_type, + const struct nlattr *userdata, const struct flow *); +static void upcall_uninit(struct upcall *); + +static upcall_callback upcall_cb; static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true); @@ -249,49 +269,64 @@ 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); + unixctl_command_register("revalidator/wait", "", 0, 0, + upcall_unixctl_dump_wait, 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); + + dpif_register_upcall_cb(dpif, upcall_cb, udpif); return udpif; } +void +udpif_run(struct udpif *udpif) +{ + if (udpif->conns && udpif->conn_seq != seq_read(udpif->dump_seq)) { + int i; + + for (i = 0; i < udpif->n_conns; i++) { + unixctl_command_reply(udpif->conns[i], NULL); + } + free(udpif->conns); + udpif->conns = NULL; + udpif->n_conns = 0; + } +} + 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); - atomic_destroy(&udpif->max_idle); - atomic_destroy(&udpif->flow_limit); + 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); @@ -299,60 +334,30 @@ 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); } for (i = 0; i < udpif->n_revalidators; i++) { - struct revalidator *revalidator = &udpif->revalidators[i]; - - ovs_mutex_lock(&revalidator->mutex); - xpthread_cond_signal(&revalidator->wake_cond); - ovs_mutex_unlock(&revalidator->mutex); - xpthread_join(revalidator->thread, NULL); + xpthread_join(udpif->revalidators[i].thread, NULL); } - xpthread_join(udpif->flow_dumper, NULL); - xpthread_join(udpif->dispatcher, NULL); + dpif_disable_upcall(udpif->dpif); 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) { - list_remove(&udump->list_node); - free(udump); - } - HMAP_FOR_EACH_SAFE (ukey, next_ukey, hmap_node, - &revalidator->ukeys) { - ukey_delete(revalidator, ukey); - } - hmap_destroy(&revalidator->ukeys); - ovs_mutex_destroy(&revalidator->mutex); + /* Delete ukeys, and delete all flows from the datapath to prevent + * double-counting stats. */ + revalidator_purge(revalidator); - free(revalidator->name); + hmap_destroy(&udpif->ukeys[i].hmap); + ovs_mutex_destroy(&udpif->ukeys[i].mutex); } - 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); - } latch_poll(&udpif->exit_latch); + ovs_barrier_destroy(&udpif->reval_barrier); + free(udpif->revalidators); udpif->revalidators = NULL; udpif->n_revalidators = 0; @@ -360,10 +365,19 @@ udpif_set_threads(struct udpif *udpif, size_t n_handlers, free(udpif->handlers); udpif->handlers = NULL; udpif->n_handlers = 0; + + free(udpif->ukeys); + udpif->ukeys = NULL; } +} - /* 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 && n_handlers && n_revalidators) { size_t i; udpif->n_handlers = n_handlers; @@ -374,30 +388,80 @@ 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); - xpthread_create(&handler->thread, NULL, udpif_upcall_handler, - handler); + handler->handler_id = i; + handler->thread = ovs_thread_create( + "handler", udpif_upcall_handler, handler); } + dpif_enable_upcall(udpif->dpif); + + ovs_barrier_init(&udpif->reval_barrier, udpif->n_revalidators); + udpif->reval_exit = false; udpif->revalidators = xzalloc(udpif->n_revalidators * sizeof *udpif->revalidators); + udpif->ukeys = xmalloc(sizeof *udpif->ukeys * n_revalidators); for (i = 0; i < udpif->n_revalidators; i++) { struct revalidator *revalidator = &udpif->revalidators[i]; revalidator->udpif = udpif; - list_init(&revalidator->udumps); - hmap_init(&revalidator->ukeys); - ovs_mutex_init(&revalidator->mutex); - xpthread_cond_init(&revalidator->wake_cond, NULL); - xpthread_create(&revalidator->thread, NULL, udpif_revalidator, - revalidator); + hmap_init(&udpif->ukeys[i].hmap); + ovs_mutex_init(&udpif->ukeys[i].mutex); + revalidator->ukeys = &udpif->ukeys[i].hmap; + revalidator->thread = ovs_thread_create( + "revalidator", udpif_revalidator, revalidator); + } + } +} + +/* 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) +{ + 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); + } + + if (!udpif->handlers && !udpif->revalidators) { + int error; + + 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; } - xpthread_create(&udpif->dispatcher, NULL, udpif_dispatcher, udpif); - xpthread_create(&udpif->flow_dumper, NULL, udpif_flow_dumper, udpif); + + 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 @@ -422,279 +486,258 @@ 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++) { - struct revalidator *revalidator = &udpif->revalidators[i]; - ovs_mutex_lock(&revalidator->mutex); - simap_increase(usage, "revalidator dumps", revalidator->n_udumps); - - /* XXX: This isn't technically thread safe because the revalidator - * ukeys maps isn't protected by a mutex since it's per thread. */ - simap_increase(usage, "revalidator keys", - hmap_count(&revalidator->ukeys)); - ovs_mutex_unlock(&revalidator->mutex); + ovs_mutex_lock(&udpif->ukeys[i].mutex); + simap_increase(usage, "udpif keys", hmap_count(&udpif->ukeys[i].hmap)); + ovs_mutex_unlock(&udpif->ukeys[i].mutex); } } -/* 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) + +static unsigned long +udpif_get_n_flows(struct udpif *udpif) { - struct dpif_dp_stats stats; - - dpif_get_dp_stats(udpif->dpif, &stats); - return stats.n_flows; + long long int time, now; + unsigned long flow_count; + + now = time_msec(); + atomic_read_relaxed(&udpif->n_flows_timestamp, &time); + if (time < now - 100 && !ovs_mutex_trylock(&udpif->n_flows_mutex)) { + struct dpif_dp_stats stats; + + atomic_store_relaxed(&udpif->n_flows_timestamp, now); + dpif_get_dp_stats(udpif->dpif, &stats); + flow_count = stats.n_flows; + atomic_store_relaxed(&udpif->n_flows, flow_count); + ovs_mutex_unlock(&udpif->n_flows_mutex); + } else { + atomic_read_relaxed(&udpif->n_flows, &flow_count); + } + return flow_count; } -/* The dispatcher thread is responsible for receiving upcalls from the kernel, - * assigning them to a upcall_handler thread. */ +/* The upcall handler thread tries to read a batch of UPCALL_MAX_BATCH + * upcalls from dpif, processes the batch and installs corresponding flows + * in dpif. */ static void * -udpif_dispatcher(void *arg) +udpif_upcall_handler(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(); + struct handler *handler = arg; + struct udpif *udpif = handler->udpif; + + while (!latch_is_set(&handler->udpif->exit_latch)) { + if (!recv_upcalls(handler)) { + dpif_recv_wait(udpif->dpif, handler->handler_id); + latch_wait(&udpif->exit_latch); + poll_block(); + } + coverage_clear(); } return NULL; } -static void * -udpif_flow_dumper(void *arg) +static size_t +recv_upcalls(struct handler *handler) { - struct udpif *udpif = arg; - - set_subprogram_name("flow_dumper"); - while (!latch_is_set(&udpif->exit_latch)) { - const struct dpif_flow_stats *stats; - long long int start_time, duration; - const struct nlattr *key, *mask; - 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; - - reval_seq = seq_read(udpif->reval_seq); - need_revalidate = udpif->last_reval_seq != reval_seq; - udpif->last_reval_seq = reval_seq; - - n_flows = udpif_get_n_flows(udpif); - 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) - && !latch_is_set(&udpif->exit_latch)) { - struct udpif_flow_dump *udump = xmalloc(sizeof *udump); - struct revalidator *revalidator; - - udump->key_hash = hash_bytes(key, key_len, udpif->secret); - memcpy(&udump->key_buf, key, key_len); - udump->key = (struct nlattr *) &udump->key_buf; - udump->key_len = key_len; - - memcpy(&udump->mask_buf, mask, mask_len); - udump->mask = (struct nlattr *) &udump->mask_buf; - udump->mask_len = mask_len; - - udump->stats = *stats; - udump->need_revalidate = need_revalidate; - - revalidator = &udpif->revalidators[udump->key_hash - % udpif->n_revalidators]; - - ovs_mutex_lock(&revalidator->mutex); - while (revalidator->n_udumps >= REVALIDATE_MAX_BATCH * 3 - && !latch_is_set(&udpif->exit_latch)) { - ovs_mutex_cond_wait(&revalidator->wake_cond, - &revalidator->mutex); - } - list_push_back(&revalidator->udumps, &udump->list_node); - revalidator->n_udumps++; - xpthread_cond_signal(&revalidator->wake_cond); - ovs_mutex_unlock(&revalidator->mutex); - } - dpif_flow_dump_done(&dump); - - /* Let all the revalidators finish and garbage collect. */ - seq_change(udpif->dump_seq); - for (i = 0; i < udpif->n_revalidators; i++) { - struct revalidator *revalidator = &udpif->revalidators[i]; - ovs_mutex_lock(&revalidator->mutex); - xpthread_cond_signal(&revalidator->wake_cond); - ovs_mutex_unlock(&revalidator->mutex); - } - - for (i = 0; i < udpif->n_revalidators; i++) { - struct revalidator *revalidator = &udpif->revalidators[i]; + struct udpif *udpif = handler->udpif; + uint64_t recv_stubs[UPCALL_MAX_BATCH][512 / 8]; + struct ofpbuf recv_bufs[UPCALL_MAX_BATCH]; + struct dpif_upcall dupcalls[UPCALL_MAX_BATCH]; + struct upcall upcalls[UPCALL_MAX_BATCH]; + struct flow flows[UPCALL_MAX_BATCH]; + size_t n_upcalls, i; + + n_upcalls = 0; + while (n_upcalls < UPCALL_MAX_BATCH) { + struct ofpbuf *recv_buf = &recv_bufs[n_upcalls]; + struct dpif_upcall *dupcall = &dupcalls[n_upcalls]; + struct upcall *upcall = &upcalls[n_upcalls]; + struct flow *flow = &flows[n_upcalls]; + struct pkt_metadata md; + int error; - ovs_mutex_lock(&revalidator->mutex); - while (revalidator->dump_seq != seq_read(udpif->dump_seq) - && !latch_is_set(&udpif->exit_latch)) { - ovs_mutex_cond_wait(&revalidator->wake_cond, - &revalidator->mutex); - } - ovs_mutex_unlock(&revalidator->mutex); + ofpbuf_use_stub(recv_buf, recv_stubs[n_upcalls], + sizeof recv_stubs[n_upcalls]); + if (dpif_recv(udpif->dpif, handler->handler_id, dupcall, recv_buf)) { + ofpbuf_uninit(recv_buf); + break; } - duration = time_msec() - start_time; - udpif->dump_duration = duration; - if (duration > 2000) { - flow_limit /= duration / 1000; - } else if (duration > 1300) { - flow_limit = flow_limit * 3 / 4; - } else if (duration < 1000 && n_flows > 2000 - && flow_limit < n_flows * 1000 / duration) { - flow_limit += 1000; + if (odp_flow_key_to_flow(dupcall->key, dupcall->key_len, flow) + == ODP_FIT_ERROR) { + goto free_dupcall; } - flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000)); - atomic_store(&udpif->flow_limit, flow_limit); - if (duration > 2000) { - VLOG_INFO("Spent an unreasonably long %lldms dumping flows", - duration); + error = upcall_receive(upcall, udpif->backer, &dupcall->packet, + dupcall->type, dupcall->userdata, flow); + if (error) { + if (error == ENODEV) { + /* 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. */ + dpif_flow_put(udpif->dpif, DPIF_FP_CREATE, dupcall->key, + dupcall->key_len, NULL, 0, NULL, 0, NULL); + VLOG_INFO_RL(&rl, "received packet on unassociated datapath " + "port %"PRIu32, flow->in_port.odp_port); + } + goto free_dupcall; } - poll_timer_wait_until(start_time + MIN(max_idle, 500)); - seq_wait(udpif->reval_seq, udpif->last_reval_seq); - latch_wait(&udpif->exit_latch); - poll_block(); - } - - return NULL; -} + upcall->key = dupcall->key; + upcall->key_len = dupcall->key_len; -/* 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. */ -static void * -udpif_upcall_handler(void *arg) -{ - struct handler *handler = arg; + upcall->out_tun_key = dupcall->out_tun_key; - handler->name = xasprintf("handler_%u", ovsthread_id_self()); - set_subprogram_name("%s", handler->name); + if (vsp_adjust_flow(upcall->ofproto, flow, &dupcall->packet)) { + upcall->vsp_adjusted = true; + } - while (!latch_is_set(&handler->udpif->exit_latch)) { - struct list misses = LIST_INITIALIZER(&misses); - size_t i; + md = pkt_metadata_from_flow(flow); + flow_extract(&dupcall->packet, &md, flow); - ovs_mutex_lock(&handler->mutex); - if (!handler->n_upcalls) { - ovs_mutex_cond_wait(&handler->wake_cond, &handler->mutex); + error = process_upcall(udpif, upcall, NULL); + if (error) { + goto cleanup; } - 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; - } - } - ovs_mutex_unlock(&handler->mutex); + n_upcalls++; + continue; - handle_upcalls(handler, &misses); +cleanup: + upcall_uninit(upcall); +free_dupcall: + ofpbuf_uninit(&dupcall->packet); + ofpbuf_uninit(recv_buf); + } - coverage_clear(); + if (n_upcalls) { + handle_upcalls(handler->udpif, upcalls, n_upcalls); + for (i = 0; i < n_upcalls; i++) { + ofpbuf_uninit(&dupcalls[i].packet); + ofpbuf_uninit(&recv_bufs[i]); + upcall_uninit(&upcalls[i]); + } } - return NULL; + return n_upcalls; } static void * udpif_revalidator(void *arg) { + /* Used by all revalidators. */ struct revalidator *revalidator = arg; + struct udpif *udpif = revalidator->udpif; + bool leader = revalidator == &udpif->revalidators[0]; + + /* Used only by the leader. */ + long long int start_time = 0; + uint64_t last_reval_seq = 0; + size_t n_flows = 0; - revalidator->name = xasprintf("revalidator_%u", ovsthread_id_self()); - set_subprogram_name("%s", revalidator->name); + revalidator->id = ovsthread_id_self(); for (;;) { - struct list udumps = LIST_INITIALIZER(&udumps); - struct udpif *udpif = revalidator->udpif; - size_t i; + if (leader) { + uint64_t reval_seq; - ovs_mutex_lock(&revalidator->mutex); - if (latch_is_set(&udpif->exit_latch)) { - ovs_mutex_unlock(&revalidator->mutex); - return NULL; - } + reval_seq = seq_read(udpif->reval_seq); + udpif->need_revalidate = last_reval_seq != reval_seq; + last_reval_seq = reval_seq; - if (!revalidator->n_udumps) { - if (revalidator->dump_seq != seq_read(udpif->dump_seq)) { - revalidator->dump_seq = seq_read(udpif->dump_seq); - revalidator_sweep(revalidator); - } else { - ovs_mutex_cond_wait(&revalidator->wake_cond, - &revalidator->mutex); + n_flows = udpif_get_n_flows(udpif); + udpif->max_n_flows = MAX(n_flows, udpif->max_n_flows); + udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2; + + /* Only the leader checks the exit latch to prevent a race where + * some threads think it's true and exit and others think it's + * false and block indefinitely on the reval_barrier */ + udpif->reval_exit = latch_is_set(&udpif->exit_latch); + + start_time = time_msec(); + if (!udpif->reval_exit) { + udpif->dump = dpif_flow_dump_create(udpif->dpif); } } - for (i = 0; i < REVALIDATE_MAX_BATCH && revalidator->n_udumps; i++) { - list_push_back(&udumps, list_pop_front(&revalidator->udumps)); - revalidator->n_udumps--; + /* Wait for the leader to start the flow dump. */ + ovs_barrier_block(&udpif->reval_barrier); + if (udpif->reval_exit) { + break; } + revalidate(revalidator); + + /* Wait for all flows to have been dumped before we garbage collect. */ + ovs_barrier_block(&udpif->reval_barrier); + revalidator_sweep(revalidator); + + /* Wait for all revalidators to finish garbage collection. */ + ovs_barrier_block(&udpif->reval_barrier); + + if (leader) { + unsigned int flow_limit; + long long int duration; + + atomic_read_relaxed(&udpif->flow_limit, &flow_limit); - /* Wake up the flow dumper. */ - xpthread_cond_signal(&revalidator->wake_cond); - ovs_mutex_unlock(&revalidator->mutex); + dpif_flow_dump_destroy(udpif->dump); + seq_change(udpif->dump_seq); - if (!list_is_empty(&udumps)) { - revalidate_udumps(revalidator, &udumps); + duration = MAX(time_msec() - start_time, 1); + udpif->dump_duration = duration; + if (duration > 2000) { + flow_limit /= duration / 1000; + } else if (duration > 1300) { + flow_limit = flow_limit * 3 / 4; + } else if (duration < 1000 && n_flows > 2000 + && flow_limit < n_flows * 1000 / duration) { + flow_limit += 1000; + } + flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000)); + atomic_store_relaxed(&udpif->flow_limit, flow_limit); + + if (duration > 2000) { + VLOG_INFO("Spent an unreasonably long %lldms dumping flows", + duration); + } + + poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500)); + seq_wait(udpif->reval_seq, last_reval_seq); + latch_wait(&udpif->exit_latch); + poll_block(); } } @@ -702,14 +745,13 @@ udpif_revalidator(void *arg) } static enum upcall_type -classify_upcall(const struct upcall *upcall) +classify_upcall(enum dpif_upcall_type type, const struct nlattr *userdata) { - const struct dpif_upcall *dpif_upcall = &upcall->dpif_upcall; union user_action_cookie cookie; size_t userdata_len; /* First look at the upcall type. */ - switch (dpif_upcall->type) { + switch (type) { case DPIF_UC_ACTION: break; @@ -718,17 +760,16 @@ classify_upcall(const struct upcall *upcall) case DPIF_N_UC_TYPES: default: - VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32, - dpif_upcall->type); + VLOG_WARN_RL(&rl, "upcall has unexpected type %"PRIu32, type); return BAD_UPCALL; } /* "action" upcalls need a closer look. */ - if (!dpif_upcall->userdata) { + if (!userdata) { VLOG_WARN_RL(&rl, "action upcall missing cookie"); return BAD_UPCALL; } - userdata_len = nl_attr_get_size(dpif_upcall->userdata); + userdata_len = nl_attr_get_size(userdata); if (userdata_len < sizeof cookie.type || userdata_len > sizeof cookie) { VLOG_WARN_RL(&rl, "action upcall cookie has unexpected size %"PRIuSIZE, @@ -736,17 +777,17 @@ classify_upcall(const struct upcall *upcall) return BAD_UPCALL; } memset(&cookie, 0, sizeof cookie); - memcpy(&cookie, nl_attr_get(dpif_upcall->userdata), userdata_len); - if (userdata_len == sizeof cookie.sflow + memcpy(&cookie, nl_attr_get(userdata), userdata_len); + 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 +797,12 @@ 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) + const struct flow *flow, odp_port_t odp_in_port, + struct ofpbuf *buf) { union user_action_cookie cookie; odp_port_t port; @@ -860,210 +815,256 @@ 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); - odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, buf); + pid = dpif_port_get_pid(udpif->dpif, port, flow_hash_5tuple(flow, 0)); + odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, ODPP_NONE, + buf); } -static struct flow_miss * -flow_miss_find(struct hmap *todo, const struct ofproto_dpif *ofproto, - const struct flow *flow, uint32_t hash) +/* If there is no error, the upcall must be destroyed with upcall_uninit() + * before quiescing, as the referred objects are guaranteed to exist only + * until the calling thread quiesces. Otherwise, do not call upcall_uninit() + * since the 'upcall->put_actions' remains uninitialized. */ +static int +upcall_receive(struct upcall *upcall, const struct dpif_backer *backer, + const struct ofpbuf *packet, enum dpif_upcall_type type, + const struct nlattr *userdata, const struct flow *flow) { - struct flow_miss *miss; + int error; - HMAP_FOR_EACH_WITH_HASH (miss, hmap_node, hash, todo) { - if (miss->ofproto == ofproto && flow_equal(&miss->flow, flow)) { - return miss; - } + error = xlate_lookup(backer, flow, &upcall->ofproto, &upcall->ipfix, + &upcall->sflow, NULL, &upcall->in_port); + if (error) { + return error; } - return NULL; + upcall->flow = flow; + upcall->packet = packet; + upcall->type = type; + upcall->userdata = userdata; + ofpbuf_init(&upcall->put_actions, 0); + + upcall->xout_initialized = false; + upcall->vsp_adjusted = false; + + upcall->key = NULL; + upcall->key_len = 0; + + upcall->out_tun_key = NULL; + + return 0; } static void -handle_upcalls(struct handler *handler, struct list *upcalls) +upcall_xlate(struct udpif *udpif, struct upcall *upcall, + struct ofpbuf *odp_actions) { - struct hmap misses = HMAP_INITIALIZER(&misses); - struct udpif *udpif = handler->udpif; + struct dpif_flow_stats stats; + struct xlate_in xin; - 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; + stats.n_packets = 1; + stats.n_bytes = ofpbuf_size(upcall->packet); + stats.used = time_msec(); + stats.tcp_flags = ntohs(upcall->flow->tcp_flags); - atomic_read(&udpif->flow_limit, &flow_limit); - may_put = udpif_get_n_flows(udpif) < flow_limit; + xlate_in_init(&xin, upcall->ofproto, upcall->flow, upcall->in_port, NULL, + stats.tcp_flags, upcall->packet); + xin.odp_actions = odp_actions; - /* 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 - * several reasons why there might be more than one, e.g.: + if (upcall->type == DPIF_UC_MISS) { + xin.resubmit_stats = &stats; + } else { + /* For non-miss upcalls, there's a flow in the datapath which this + * packet was accounted to. Presumably the revalidators will deal + * with pushing its stats eventually. */ + } + + xlate_actions(&xin, &upcall->xout); + upcall->xout_initialized = true; + + /* Special case for fail-open mode. * - * - The dpif packet interface does not support TSO (or UFO, etc.), so a - * large packet sent to userspace is split into a sequence of smaller - * ones. + * If we are in fail-open mode, but we are connected to a controller too, + * then 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. * - * - A stream of quickly arriving packets in an established "slow-pathed" - * flow. + * See the top-level comment in fail-open.c for more information. * - * - Rarely, a stream of quickly arriving packets in a flow not yet - * established. (This is rare because most protocols do not send - * multiple back-to-back packets before receiving a reply from the - * 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; - struct flow_miss *miss = &miss_buf[n_misses]; - struct ofpbuf *packet = &dupcall->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; - int error; + * Copy packets before they are modified by execution. */ + if (upcall->xout.fail_open) { + const struct ofpbuf *packet = upcall->packet; + struct ofproto_packet_in *pin; + + pin = xmalloc(sizeof *pin); + 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(upcall->flow, &pin->up.fmd); + pin->send_len = 0; /* Not used for flow table misses. */ + pin->miss_type = OFPROTO_PACKET_IN_NO_MISS; + ofproto_dpif_send_packet_in(upcall->ofproto, pin); + } - error = xlate_receive(udpif->backer, packet, dupcall->key, - dupcall->key_len, &flow, &miss->key_fitness, - &ofproto, &ipfix, &sflow, NULL, &odp_in_port); - if (error) { - if (error == ENODEV) { - /* 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); - dpif_flow_put(udpif->dpif, DPIF_FP_CREATE | DPIF_FP_MODIFY, - dupcall->key, dupcall->key_len, NULL, 0, NULL, 0, - NULL); - } - list_remove(&upcall->list_node); - upcall_destroy(upcall); - continue; + if (!upcall->xout.slow) { + ofpbuf_use_const(&upcall->put_actions, + ofpbuf_data(upcall->xout.odp_actions), + ofpbuf_size(upcall->xout.odp_actions)); + } else { + ofpbuf_init(&upcall->put_actions, 0); + compose_slow_path(udpif, &upcall->xout, upcall->flow, + upcall->flow->in_port.odp_port, + &upcall->put_actions); + } +} + +static void +upcall_uninit(struct upcall *upcall) +{ + if (upcall) { + if (upcall->xout_initialized) { + xlate_out_uninit(&upcall->xout); } + ofpbuf_uninit(&upcall->put_actions); + } +} - type = classify_upcall(upcall); - if (type == MISS_UPCALL) { - uint32_t hash; - - flow_extract(packet, flow.skb_priority, flow.pkt_mark, - &flow.tunnel, &flow.in_port, &miss->flow); - - hash = flow_hash(&miss->flow, 0); - existing_miss = flow_miss_find(&misses, ofproto, &miss->flow, - hash); - if (!existing_miss) { - hmap_insert(&misses, &miss->hmap_node, hash); - miss->ofproto = ofproto; - miss->key = dupcall->key; - miss->key_len = dupcall->key_len; - miss->upcall_type = dupcall->type; - miss->stats.n_packets = 0; - miss->stats.n_bytes = 0; - 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.n_packets++; +static int +upcall_cb(const struct ofpbuf *packet, const struct flow *flow, + enum dpif_upcall_type type, const struct nlattr *userdata, + struct ofpbuf *actions, struct flow_wildcards *wc, + struct ofpbuf *put_actions, void *aux) +{ + struct udpif *udpif = aux; + unsigned int flow_limit; + struct upcall upcall; + bool megaflow; + int error; + + atomic_read_relaxed(&enable_megaflows, &megaflow); + atomic_read_relaxed(&udpif->flow_limit, &flow_limit); + + error = upcall_receive(&upcall, udpif->backer, packet, type, userdata, + flow); + if (error) { + return error; + } + + error = process_upcall(udpif, &upcall, actions); + if (error) { + goto out; + } + + if (upcall.xout.slow && put_actions) { + ofpbuf_put(put_actions, ofpbuf_data(&upcall.put_actions), + ofpbuf_size(&upcall.put_actions)); + } - upcall->flow_miss = miss; - continue; + if (OVS_LIKELY(wc)) { + if (megaflow) { + /* XXX: This could be avoided with sufficient API changes. */ + *wc = upcall.xout.wc; + } else { + flow_wildcards_init_for_packet(wc, flow); } + } - switch (type) { - case SFLOW_UPCALL: - if (sflow) { - union user_action_cookie cookie; + if (udpif_get_n_flows(udpif) >= flow_limit) { + error = ENOSPC; + } - memset(&cookie, 0, sizeof cookie); - memcpy(&cookie, nl_attr_get(dupcall->userdata), - sizeof cookie.sflow); - dpif_sflow_received(sflow, packet, &flow, odp_in_port, - &cookie); - } - break; - case IPFIX_UPCALL: - if (ipfix) { - dpif_ipfix_bridge_sample(ipfix, packet, &flow); - } - break; - case FLOW_SAMPLE_UPCALL: - if (ipfix) { - union user_action_cookie cookie; - - memset(&cookie, 0, sizeof cookie); - memcpy(&cookie, nl_attr_get(dupcall->userdata), - sizeof cookie.flow_sample); - - /* The flow reflects exactly the contents of the packet. - * Sample the packet using it. */ - dpif_ipfix_flow_sample(ipfix, packet, &flow, - cookie.flow_sample.collector_set_id, - cookie.flow_sample.probability, - cookie.flow_sample.obs_domain_id, - cookie.flow_sample.obs_point_id); - } - break; - case BAD_UPCALL: - break; - case MISS_UPCALL: - OVS_NOT_REACHED(); +out: + upcall_uninit(&upcall); + return error; +} + +static int +process_upcall(struct udpif *udpif, struct upcall *upcall, + struct ofpbuf *odp_actions) +{ + const struct nlattr *userdata = upcall->userdata; + const struct ofpbuf *packet = upcall->packet; + const struct flow *flow = upcall->flow; + + switch (classify_upcall(upcall->type, userdata)) { + case MISS_UPCALL: + upcall_xlate(udpif, upcall, odp_actions); + return 0; + + case SFLOW_UPCALL: + if (upcall->sflow) { + union user_action_cookie cookie; + + memset(&cookie, 0, sizeof cookie); + memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.sflow); + dpif_sflow_received(upcall->sflow, packet, flow, + flow->in_port.odp_port, &cookie); } + break; - dpif_ipfix_unref(ipfix); - dpif_sflow_unref(sflow); + case IPFIX_UPCALL: + if (upcall->ipfix) { + union user_action_cookie cookie; + struct flow_tnl output_tunnel_key; - list_remove(&upcall->list_node); - upcall_destroy(upcall); - } + memset(&cookie, 0, sizeof cookie); + memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.ipfix); - /* Initialize each 'struct flow_miss's ->xout. - * - * We do this per-flow_miss rather than per-packet because, most commonly, - * all the packets in a flow can use the same translation. - * - * 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) { - struct xlate_in xin; - - xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL, - miss->stats.tcp_flags, NULL); - xin.may_learn = true; + if (upcall->out_tun_key) { + memset(&output_tunnel_key, 0, sizeof output_tunnel_key); + odp_tun_key_from_attr(upcall->out_tun_key, + &output_tunnel_key); + } + dpif_ipfix_bridge_sample(upcall->ipfix, packet, flow, + flow->in_port.odp_port, + cookie.ipfix.output_odp_port, + upcall->out_tun_key ? + &output_tunnel_key : NULL); + } + break; - if (miss->upcall_type == DPIF_UC_MISS) { - xin.resubmit_stats = &miss->stats; - } else { - /* For non-miss upcalls, there's a flow in the datapath which this - * packet was accounted to. Presumably the revalidators will deal - * with pushing its stats eventually. */ + case FLOW_SAMPLE_UPCALL: + if (upcall->ipfix) { + union user_action_cookie cookie; + + memset(&cookie, 0, sizeof cookie); + memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.flow_sample); + + /* The flow reflects exactly the contents of the packet. + * Sample the packet using it. */ + dpif_ipfix_flow_sample(upcall->ipfix, packet, flow, + cookie.flow_sample.collector_set_id, + cookie.flow_sample.probability, + cookie.flow_sample.obs_domain_id, + cookie.flow_sample.obs_point_id); } + break; - xlate_actions(&xin, &miss->xout); - fail_open = fail_open || miss->xout.fail_open; + case BAD_UPCALL: + break; } - /* Now handle the packets individually in order of arrival. In the common - * case each packet of a miss can share the same actions, but slow-pathed - * packets need to be translated individually: + return EAGAIN; +} + +static void +handle_upcalls(struct udpif *udpif, struct upcall *upcalls, + size_t n_upcalls) +{ + struct odputil_keybuf mask_bufs[UPCALL_MAX_BATCH]; + struct dpif_op *opsp[UPCALL_MAX_BATCH * 2]; + struct dpif_op ops[UPCALL_MAX_BATCH * 2]; + unsigned int flow_limit; + size_t n_ops, i; + bool may_put; + bool megaflow; + + atomic_read_relaxed(&udpif->flow_limit, &flow_limit); + atomic_read_relaxed(&enable_megaflows, &megaflow); + + may_put = udpif_get_n_flows(udpif) < flow_limit; + + /* Handle the packets individually in order of arrival. * * - For SLOW_CFM, SLOW_LACP, SLOW_STP, and SLOW_BFD, translation is what * processes received packets for these protocols. @@ -1074,132 +1075,70 @@ 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) { - struct flow_miss *miss = upcall->flow_miss; - struct ofpbuf *packet = &upcall->dpif_upcall.packet; + for (i = 0; i < n_upcalls; i++) { + struct upcall *upcall = &upcalls[i]; + const struct ofpbuf *packet = upcall->packet; struct dpif_op *op; - 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; - - xlate_in_init(&xin, miss->ofproto, &miss->flow, NULL, 0, packet); - xlate_actions_for_side_effects(&xin); - } - 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 (miss->xout.odp_actions.size) { - eth_pop_vlan(packet); + if (upcall->vsp_adjusted) { + /* 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(upcall->xout.odp_actions)) { + eth_pop_vlan(CONST_CAST(struct ofpbuf *, upcall->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; + CONST_CAST(struct flow *, upcall->flow)->vlan_tci = 0; } /* 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) { + if (may_put && upcall->type == DPIF_UC_MISS) { struct ofpbuf mask; - bool megaflow; - miss->put = true; + ofpbuf_use_stack(&mask, &mask_bufs[i], sizeof mask_bufs[i]); - 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); + size_t max_mpls; + bool recirc; + + recirc = ofproto_dpif_get_enable_recirc(upcall->ofproto); + max_mpls = ofproto_dpif_get_max_mpls_depth(upcall->ofproto); + odp_flow_key_from_mask(&mask, &upcall->xout.wc.masks, + upcall->flow, UINT32_MAX, max_mpls, + recirc); } 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.flags = DPIF_FP_CREATE; + op->u.flow_put.key = upcall->key; + op->u.flow_put.key_len = upcall->key_len; + 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; - } 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; - } + op->u.flow_put.actions = ofpbuf_data(&upcall->put_actions); + op->u.flow_put.actions_len = ofpbuf_size(&upcall->put_actions); } - /* - * 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 (miss->xout.odp_actions.size) { - + if (ofpbuf_size(upcall->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.packet = CONST_CAST(struct ofpbuf *, packet); + odp_key_to_pkt_metadata(upcall->key, upcall->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.needs_help = (miss->xout.slow & SLOW_ACTION) != 0; - } - } - - /* Special case for fail-open mode. - * - * If we are in fail-open mode, but we are connected to a controller too, - * then 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. - * - * Copy packets before they are modified by execution. */ - if (fail_open) { - LIST_FOR_EACH (upcall, list_node, upcalls) { - 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.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; - ofproto_dpif_send_packet_in(miss->ofproto, pin); + op->u.execute.actions = ofpbuf_data(upcall->xout.odp_actions); + op->u.execute.actions_len = ofpbuf_size(upcall->xout.odp_actions); + op->u.execute.needs_help = (upcall->xout.slow & SLOW_ACTION) != 0; + op->u.execute.probe = false; } } @@ -1208,118 +1147,231 @@ 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); - } } +/* Must be called with udpif->ukeys[hash % udpif->n_revalidators].mutex. */ static struct udpif_key * -ukey_lookup(struct revalidator *revalidator, struct udpif_flow_dump *udump) +ukey_lookup(struct udpif *udpif, const struct nlattr *key, size_t key_len, + uint32_t hash) + OVS_REQUIRES(udpif->ukeys->mutex) { struct udpif_key *ukey; + struct hmap *hmap = &udpif->ukeys[hash % udpif->n_revalidators].hmap; - HMAP_FOR_EACH_WITH_HASH (ukey, hmap_node, udump->key_hash, - &revalidator->ukeys) { - if (ukey->key_len == udump->key_len - && !memcmp(ukey->key, udump->key, udump->key_len)) { + HMAP_FOR_EACH_WITH_HASH (ukey, hmap_node, hash, hmap) { + if (ukey->key_len == key_len && !memcmp(ukey->key, key, key_len)) { return ukey; } } return NULL; } +/* Creates a ukey for 'key' and 'key_len', returning it with ukey->mutex in + * a locked state. */ +static struct udpif_key * +ukey_create(const struct nlattr *key, size_t key_len, long long int used) + OVS_NO_THREAD_SAFETY_ANALYSIS +{ + struct udpif_key *ukey = xmalloc(sizeof *ukey); + + ovs_mutex_init(&ukey->mutex); + ukey->key = &ukey->key_buf_nla; + memcpy(&ukey->key_buf, key, key_len); + ukey->key_len = key_len; + + ovs_mutex_lock(&ukey->mutex); + ukey->dump_seq = 0; + ukey->flow_exists = true; + ukey->created = used ? used : time_msec(); + memset(&ukey->stats, 0, sizeof ukey->stats); + ukey->xcache = NULL; + + return ukey; +} + +/* Searches for a ukey in 'udpif->ukeys' that matches 'key' and 'key_len' and + * attempts to lock the ukey. If the ukey does not exist, create it. + * + * Returns true on success, setting *result to the matching ukey and returning + * it in a locked state. Otherwise, returns false and clears *result. */ +static bool +ukey_acquire(struct udpif *udpif, const struct nlattr *key, size_t key_len, + long long int used, struct udpif_key **result) + OVS_TRY_LOCK(true, (*result)->mutex) +{ + struct udpif_key *ukey; + uint32_t hash, idx; + bool locked = false; + + hash = hash_bytes(key, key_len, udpif->secret); + idx = hash % udpif->n_revalidators; + + ovs_mutex_lock(&udpif->ukeys[idx].mutex); + ukey = ukey_lookup(udpif, key, key_len, hash); + if (!ukey) { + ukey = ukey_create(key, key_len, used); + hmap_insert(&udpif->ukeys[idx].hmap, &ukey->hmap_node, hash); + locked = true; + } else if (!ovs_mutex_trylock(&ukey->mutex)) { + locked = true; + } + ovs_mutex_unlock(&udpif->ukeys[idx].mutex); + + if (locked) { + *result = ukey; + } else { + *result = NULL; + } + return locked; +} + static void ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey) + OVS_NO_THREAD_SAFETY_ANALYSIS { - hmap_remove(&revalidator->ukeys, &ukey->hmap_node); + if (revalidator) { + hmap_remove(revalidator->ukeys, &ukey->hmap_node); + } + xlate_cache_delete(ukey->xcache); + ovs_mutex_destroy(&ukey->mutex); free(ukey); } static bool -revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump, - struct udpif_key *ukey) +should_revalidate(const struct udpif *udpif, uint64_t packets, + long long int used) +{ + long long int metric, now, duration; + + if (udpif->dump_duration < 200) { + /* We are likely to handle full revalidation for the flows. */ + return true; + } + + /* 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) { + /* The flow is receiving more than ~5pps, so keep it. */ + return true; + } + return false; +} + +static bool +revalidate_ukey(struct udpif *udpif, struct udpif_key *ukey, + const struct dpif_flow *f) + OVS_REQUIRES(ukey->mutex) { - struct ofpbuf xout_actions, *actions; uint64_t slow_path_buf[128 / 8]; struct xlate_out xout, *xoutp; - struct flow flow, udump_mask; + struct netflow *netflow; struct ofproto_dpif *ofproto; struct dpif_flow_stats push; - uint32_t *udump32, *xout32; - odp_port_t odp_in_port; + struct ofpbuf xout_actions; + struct flow flow, dp_mask; + uint32_t *dp32, *xout32; + ofp_port_t ofp_in_port; struct xlate_in xin; + long long int last_used; int error; size_t i; bool ok; ok = false; xoutp = NULL; - actions = 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 - * them. */ - if (udump->need_revalidate) { - if (dpif_flow_get(udpif->dpif, ukey->key, ukey->key_len, &actions, - &udump->stats)) { - goto exit; - } + netflow = NULL; + + last_used = ukey->stats.used; + push.used = f->stats.used; + push.tcp_flags = f->stats.tcp_flags; + push.n_packets = (f->stats.n_packets > ukey->stats.n_packets + ? f->stats.n_packets - ukey->stats.n_packets + : 0); + push.n_bytes = (f->stats.n_bytes > ukey->stats.n_bytes + ? f->stats.n_bytes - ukey->stats.n_bytes + : 0); + + if (udpif->need_revalidate && last_used + && !should_revalidate(udpif, push.n_packets, last_used)) { + ok = false; + goto exit; } - push.used = udump->stats.used; - push.tcp_flags = udump->stats.tcp_flags; - push.n_packets = udump->stats.n_packets > ukey->stats.n_packets - ? udump->stats.n_packets - ukey->stats.n_packets - : 0; - push.n_bytes = udump->stats.n_bytes > ukey->stats.n_bytes - ? udump->stats.n_bytes - ukey->stats.n_bytes - : 0; - ukey->stats = udump->stats; - - if (!push.n_packets && !udump->need_revalidate) { + /* We will push the stats, so update the ukey stats cache. */ + ukey->stats = f->stats; + if (!push.n_packets && !udpif->need_revalidate) { ok = true; goto exit; } - error = xlate_receive(udpif->backer, NULL, ukey->key, ukey->key_len, &flow, - NULL, &ofproto, NULL, NULL, NULL, &odp_in_port); + if (ukey->xcache && !udpif->need_revalidate) { + xlate_push_stats(ukey->xcache, &push); + ok = true; + goto exit; + } + + if (odp_flow_key_to_flow(ukey->key, ukey->key_len, &flow) + == ODP_FIT_ERROR) { + goto exit; + } + + error = xlate_lookup(udpif->backer, &flow, &ofproto, NULL, NULL, &netflow, + &ofp_in_port); if (error) { goto exit; } - 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 = !udump->need_revalidate; + if (udpif->need_revalidate) { + xlate_cache_clear(ukey->xcache); + } + if (!ukey->xcache) { + ukey->xcache = xlate_cache_new(); + } + + xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL, push.tcp_flags, + NULL); + if (push.n_packets) { + xin.resubmit_stats = &push; + xin.may_learn = true; + } + xin.xcache = ukey->xcache; + xin.skip_wildcards = !udpif->need_revalidate; xlate_actions(&xin, &xout); xoutp = &xout; - if (!udump->need_revalidate) { + if (!udpif->need_revalidate) { ok = true; goto exit; } 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, flow.in_port.odp_port, + &xout_actions); } - if (!ofpbuf_equal(&xout_actions, actions)) { + if (f->actions_len != ofpbuf_size(&xout_actions) + || memcmp(ofpbuf_data(&xout_actions), f->actions, f->actions_len)) { goto exit; } - if (odp_flow_key_to_mask(udump->mask, udump->mask_len, &udump_mask, &flow) + if (odp_flow_key_to_mask(f->mask, f->mask_len, &dp_mask, &flow) == ODP_FIT_ERROR) { goto exit; } @@ -1329,168 +1381,286 @@ revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump, * mask in the kernel is more specific i.e. less wildcarded, than what * we've calculated here. This guarantees we don't catch any packets we * shouldn't with the megaflow. */ - udump32 = (uint32_t *) &udump_mask; + dp32 = (uint32_t *) &dp_mask; xout32 = (uint32_t *) &xout.wc.masks; for (i = 0; i < FLOW_U32S; i++) { - if ((udump32[i] | xout32[i]) != udump32[i]) { + if ((dp32[i] | xout32[i]) != dp32[i]) { goto exit; } } ok = true; exit: - ofpbuf_delete(actions); + if (netflow && !ok) { + netflow_flow_clear(netflow, &flow); + } xlate_out_uninit(xoutp); return ok; } +struct dump_op { + struct udpif_key *ukey; + 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 *udpif = revalidator->udpif; - - 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]; + op->ukey = ukey; + 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 udpif *udpif, struct dump_op *ops, size_t n_ops) +{ struct dpif_op *opsp[REVALIDATE_MAX_BATCH]; - struct udpif_flow_dump *udump, *next_udump; - size_t n_ops, i, 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); + size_t i; - must_del = false; - if (n_flows > flow_limit) { - must_del = n_flows > 2 * flow_limit; - max_idle = 100; + 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); - n_ops = 0; - LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) { - long long int used, now; - struct udpif_key *ukey; - - now = time_msec(); - ukey = ukey_lookup(revalidator, udump); + for (i = 0; i < n_ops; i++) { + struct dump_op *op = &ops[i]; + struct dpif_flow_stats *push, *stats, push_buf; - used = udump->stats.used; - if (!used && ukey) { - used = ukey->created; - } + stats = op->op.u.flow_del.stats; + push = &push_buf; - 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; + ovs_mutex_lock(&op->ukey->mutex); + 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; + ovs_mutex_unlock(&op->ukey->mutex); - 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 (push->n_packets || netflow_exists()) { + struct ofproto_dpif *ofproto; + struct netflow *netflow; + ofp_port_t ofp_in_port; + struct flow flow; + int error; - if (ukey) { - *ukey_stats = ukey->stats; - ukey_delete(revalidator, ukey); - } else { - memset(ukey_stats, 0, sizeof *ukey_stats); + ovs_mutex_lock(&op->ukey->mutex); + if (op->ukey->xcache) { + xlate_push_stats(op->ukey->xcache, push); + ovs_mutex_unlock(&op->ukey->mutex); + continue; } + ovs_mutex_unlock(&op->ukey->mutex); - 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); + if (odp_flow_key_to_flow(op->op.u.flow_del.key, + op->op.u.flow_del.key_len, &flow) + == ODP_FIT_ERROR) { + continue; + } - ukey->mark = false; + error = xlate_lookup(udpif->backer, &flow, &ofproto, + NULL, NULL, &netflow, &ofp_in_port); + if (!error) { + struct xlate_in xin; - hmap_insert(&revalidator->ukeys, &ukey->hmap_node, - udump->key_hash); - } - ukey->mark = true; + xlate_in_init(&xin, ofproto, &flow, ofp_in_port, 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 (!revalidate_ukey(udpif, udump, ukey)) { - dpif_flow_del(udpif->dpif, udump->key, udump->key_len, NULL); - ukey_delete(revalidator, ukey); + if (netflow) { + netflow_flow_clear(netflow, &flow); + } + } } - - list_remove(&udump->list_node); - free(udump); } +} + +static void +push_dump_ops(struct revalidator *revalidator, + struct dump_op *ops, size_t n_ops) +{ + int i; + push_dump_ops__(revalidator->udpif, ops, n_ops); for (i = 0; i < n_ops; i++) { - opsp[i] = &ops[i].op; + ukey_delete(revalidator, ops[i].ukey); } - dpif_operate(udpif->dpif, opsp, n_ops); +} - for (i = 0; i < n_ops; i++) { - struct dpif_flow_stats push, *stats, *ukey_stats; +static void +revalidate(struct revalidator *revalidator) +{ + struct udpif *udpif = revalidator->udpif; + struct dpif_flow_dump_thread *dump_thread; + uint64_t dump_seq; + unsigned int flow_limit; - 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; + dump_seq = seq_read(udpif->dump_seq); + atomic_read_relaxed(&udpif->flow_limit, &flow_limit); + dump_thread = dpif_flow_dump_thread_create(udpif->dump); + for (;;) { + struct dump_op ops[REVALIDATE_MAX_BATCH]; + int n_ops = 0; - if (push.n_packets || netflow_exists()) { - struct ofproto_dpif *ofproto; - struct netflow *netflow; - struct flow flow; + struct dpif_flow flows[REVALIDATE_MAX_BATCH]; + const struct dpif_flow *f; + int n_dumped; - 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; + long long int max_idle; + long long int now; + size_t n_dp_flows; + bool kill_them_all; - 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); + n_dumped = dpif_flow_dump_next(dump_thread, flows, ARRAY_SIZE(flows)); + if (!n_dumped) { + break; + } - if (netflow) { - netflow_expire(netflow, &flow); - netflow_flow_clear(netflow, &flow); - netflow_unref(netflow); + now = time_msec(); + + /* In normal operation we want to keep flows around until they have + * been idle for 'ofproto_max_idle' milliseconds. However: + * + * - If the number of datapath flows climbs above 'flow_limit', + * drop that down to 100 ms to try to bring the flows down to + * the limit. + * + * - If the number of datapath flows climbs above twice + * 'flow_limit', delete all the datapath flows as an emergency + * measure. (We reassess this condition for the next batch of + * datapath flows, so we will recover before all the flows are + * gone.) */ + n_dp_flows = udpif_get_n_flows(udpif); + kill_them_all = n_dp_flows > flow_limit * 2; + max_idle = n_dp_flows > flow_limit ? 100 : ofproto_max_idle; + + for (f = flows; f < &flows[n_dumped]; f++) { + long long int used = f->stats.used; + struct udpif_key *ukey; + bool already_dumped, keep; + + if (!ukey_acquire(udpif, f->key, f->key_len, used, &ukey)) { + /* We couldn't acquire the ukey. This means that + * another revalidator is processing this flow + * concurrently, so don't bother processing it. */ + COVERAGE_INC(dumped_duplicate_flow); + continue; + } + + already_dumped = ukey->dump_seq == dump_seq; + if (already_dumped) { + /* The flow has already been handled during this flow dump + * operation. Skip it. */ + if (ukey->xcache) { + COVERAGE_INC(dumped_duplicate_flow); + } else { + COVERAGE_INC(dumped_new_flow); } + ovs_mutex_unlock(&ukey->mutex); + continue; + } + + if (!used) { + used = ukey->created; + } + if (kill_them_all || (used && used < now - max_idle)) { + keep = false; + } else { + keep = revalidate_ukey(udpif, ukey, f); + } + ukey->dump_seq = dump_seq; + ukey->flow_exists = keep; + + if (!keep) { + dump_op_init(&ops[n_ops++], f->key, f->key_len, ukey); } + ovs_mutex_unlock(&ukey->mutex); + } + + if (n_ops) { + push_dump_ops__(udpif, ops, n_ops); } } + dpif_flow_dump_thread_destroy(dump_thread); +} + +/* Called with exclusive access to 'revalidator' and 'ukey'. */ +static bool +handle_missed_revalidation(struct revalidator *revalidator, + struct udpif_key *ukey) + OVS_NO_THREAD_SAFETY_ANALYSIS +{ + struct udpif *udpif = revalidator->udpif; + struct dpif_flow flow; + struct ofpbuf buf; + uint64_t stub[DPIF_FLOW_BUFSIZE / 8]; + bool keep = false; - LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) { - list_remove(&udump->list_node); - free(udump); + COVERAGE_INC(revalidate_missed_dp_flow); + + ofpbuf_use_stub(&buf, &stub, sizeof stub); + if (!dpif_flow_get(udpif->dpif, ukey->key, ukey->key_len, &buf, &flow)) { + keep = revalidate_ukey(udpif, ukey, &flow); } + ofpbuf_uninit(&buf); + + return keep; } static void -revalidator_sweep(struct revalidator *revalidator) +revalidator_sweep__(struct revalidator *revalidator, bool purge) + OVS_NO_THREAD_SAFETY_ANALYSIS { + struct dump_op ops[REVALIDATE_MAX_BATCH]; struct udpif_key *ukey, *next; + size_t n_ops; + uint64_t dump_seq; - HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, &revalidator->ukeys) { - if (ukey->mark) { - ukey->mark = false; - } else { + n_ops = 0; + dump_seq = seq_read(revalidator->udpif->dump_seq); + + /* During garbage collection, this revalidator completely owns its ukeys + * map, and therefore doesn't need to do any locking. */ + HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, revalidator->ukeys) { + if (ukey->flow_exists + && (purge + || (ukey->dump_seq != dump_seq + && revalidator->udpif->need_revalidate + && !handle_missed_revalidation(revalidator, ukey)))) { + struct dump_op *op = &ops[n_ops++]; + + dump_op_init(op, ukey->key, ukey->key_len, ukey); + if (n_ops == REVALIDATE_MAX_BATCH) { + push_dump_ops(revalidator, ops, n_ops); + n_ops = 0; + } + } else if (!ukey->flow_exists) { ukey_delete(revalidator, ukey); } } + + 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 @@ -1502,40 +1672,24 @@ 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); + atomic_read_relaxed(&udpif->flow_limit, &flow_limit); ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif)); - ds_put_format(&ds, "\tflows : (current %"PRIu64")" + ds_put_format(&ds, "\tflows : (current %lu)" " (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]; - /* XXX: The result of hmap_count(&revalidator->ukeys) may not be - * accurate because it's not protected by the revalidator mutex. */ - ovs_mutex_lock(&revalidator->mutex); - ds_put_format(&ds, "\t%s: (dump queue %"PRIuSIZE") (keys %"PRIuSIZE - ")\n", revalidator->name, revalidator->n_udumps, - hmap_count(&revalidator->ukeys)); - ovs_mutex_unlock(&revalidator->mutex); + ovs_mutex_lock(&udpif->ukeys[i].mutex); + ds_put_format(&ds, "\t%u: (keys %"PRIuSIZE")\n", + revalidator->id, hmap_count(&udpif->ukeys[i].hmap)); + ovs_mutex_unlock(&udpif->ukeys[i].mutex); } } @@ -1553,8 +1707,8 @@ upcall_unixctl_disable_megaflows(struct unixctl_conn *conn, const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED) { - atomic_store(&enable_megaflows, false); - udpif_flush(); + atomic_store_relaxed(&enable_megaflows, false); + udpif_flush_all_datapaths(); unixctl_command_reply(conn, "megaflows disabled"); } @@ -1568,7 +1722,49 @@ upcall_unixctl_enable_megaflows(struct unixctl_conn *conn, const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED) { - atomic_store(&enable_megaflows, true); - udpif_flush(); + atomic_store_relaxed(&enable_megaflows, true); + 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_relaxed(&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); +} + +static void +upcall_unixctl_dump_wait(struct unixctl_conn *conn, + int argc OVS_UNUSED, + const char *argv[] OVS_UNUSED, + void *aux OVS_UNUSED) +{ + if (list_is_singleton(&all_udpifs)) { + struct udpif *udpif = NULL; + size_t len; + + udpif = OBJECT_CONTAINING(list_front(&all_udpifs), udpif, list_node); + len = (udpif->n_conns + 1) * sizeof *udpif->conns; + udpif->conn_seq = seq_read(udpif->dump_seq); + udpif->conns = xrealloc(udpif->conns, len); + udpif->conns[udpif->n_conns++] = conn; + } else { + unixctl_command_reply_error(conn, "can't wait on multiple udpifs."); + } +}