X-Git-Url: http://git.cascardo.eti.br/?a=blobdiff_plain;f=ofproto%2Fofproto-dpif-upcall.c;h=c1919275b668bde54dd33cc4d83ab437aa893cd3;hb=b476e2f21b1f20bd0d24c0522e4d745603599582;hp=3a75690a5dfaa0729da774348e227d1511a9b823;hpb=feca8bd78a2e8bf5bb37ac8d07e6040d6ff75bd5;p=cascardo%2Fovs.git diff --git a/ofproto/ofproto-dpif-upcall.c b/ofproto/ofproto-dpif-upcall.c index 3a75690a5..c1919275b 100644 --- a/ofproto/ofproto-dpif-upcall.c +++ b/ofproto/ofproto-dpif-upcall.c @@ -1,4 +1,4 @@ -/* Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc. +/* Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2015 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -21,6 +21,7 @@ #include "connmgr.h" #include "coverage.h" +#include "cmap.h" #include "dpif.h" #include "dynamic-string.h" #include "fail-open.h" @@ -37,15 +38,19 @@ #include "poll-loop.h" #include "seq.h" #include "unixctl.h" -#include "vlog.h" +#include "openvswitch/vlog.h" #define MAX_QUEUE_LENGTH 512 -#define UPCALL_MAX_BATCH 50 +#define UPCALL_MAX_BATCH 64 #define REVALIDATE_MAX_BATCH 50 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall); -COVERAGE_DEFINE(upcall_duplicate_flow); +COVERAGE_DEFINE(dumped_duplicate_flow); +COVERAGE_DEFINE(dumped_new_flow); +COVERAGE_DEFINE(handler_duplicate_upcall); +COVERAGE_DEFINE(upcall_ukey_contention); +COVERAGE_DEFINE(revalidate_missed_dp_flow); /* A thread that reads upcalls from dpif, forwards each upcall's packet, * and possibly sets up a kernel flow as a cache. */ @@ -55,14 +60,21 @@ struct handler { uint32_t handler_id; /* Handler id. */ }; +/* In the absence of a multiple-writer multiple-reader datastructure for + * storing ukeys, we use a large number of cmaps, each with its own lock for + * writing. */ +#define N_UMAPS 512 /* per udpif. */ +struct umap { + struct ovs_mutex mutex; /* Take for writing to the following. */ + struct cmap cmap; /* Datapath flow keys. */ +}; + /* A thread that processes datapath flows, updates OpenFlow statistics, and * updates or removes them if necessary. */ struct revalidator { struct udpif *udpif; /* Parent udpif. */ pthread_t thread; /* Thread ID. */ 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. @@ -82,13 +94,11 @@ struct revalidator { * them. */ struct udpif { - struct list list_node; /* In all_udpifs list. */ + struct ovs_list list_node; /* In all_udpifs list. */ struct dpif *dpif; /* Datapath handle. */ struct dpif_backer *backer; /* Opaque dpif_backer pointer. */ - uint32_t secret; /* Random seed for upcall hash. */ - struct handler *handlers; /* Upcall handlers. */ size_t n_handlers; @@ -99,23 +109,33 @@ struct udpif { /* 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. */ - pthread_barrier_t reval_barrier; /* Barrier used by revalidators. */ + 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. + atomic_bool enable_ufid; /* If true, skip dumping flow attrs. */ + + /* These variables provide a mechanism for the main thread to pause + * all revalidation without having to completely shut the threads down. + * 'pause_latch' is shared between the main thread and the lead + * revalidator thread, so when it is desirable to halt revalidation, the + * main thread will set the latch. 'pause' and 'pause_barrier' are shared + * by revalidator threads. The lead revalidator will set 'pause' when it + * observes the latch has been set, and this will cause all revalidator + * threads to wait on 'pause_barrier' at the beginning of the next + * revalidation round. */ + bool pause; /* Set by leader on 'pause_latch. */ + struct latch pause_latch; /* Set to force revalidators pause. */ + struct ovs_barrier pause_barrier; /* Barrier used to pause all */ + /* revalidators by main thread. */ + + /* There are 'N_UMAPS' maps containing 'struct udpif_key' elements. * * 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; + * takes care of garbage collecting a slice of these maps. */ + struct umap *ukeys; /* Datapath flow statistics. */ unsigned int max_n_flows; @@ -125,9 +145,15 @@ struct udpif { atomic_uint flow_limit; /* Datapath flow hard limit. */ /* n_flows_mutex prevents multiple threads updating these concurrently. */ - atomic_ulong n_flows; /* Number of flows in the datapath. */ + 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 { @@ -138,71 +164,131 @@ enum upcall_type { IPFIX_UPCALL /* Per-bridge sampling. */ }; -struct upcall { - struct ofproto_dpif *ofproto; - - struct flow flow; - 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; +enum reval_result { + UKEY_KEEP, + UKEY_DELETE, + UKEY_MODIFY +}; - struct xlate_out xout; +struct upcall { + struct ofproto_dpif *ofproto; /* Parent ofproto. */ + const struct recirc_id_node *recirc; /* Recirculation context. */ + bool have_recirc_ref; /* Reference held on recirc ctx? */ + + /* 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 ovs_u128 *ufid; /* Unique identifier for 'flow'. */ + unsigned pmd_id; /* Datapath poll mode driver id. */ + const struct dp_packet *packet; /* Packet associated with this upcall. */ + ofp_port_t in_port; /* OpenFlow in port, or OFPP_NONE. */ + uint16_t mru; /* If !0, Maximum receive unit of + fragmented IP packet */ + + enum dpif_upcall_type type; /* Datapath type of the upcall. */ + const struct nlattr *userdata; /* Userdata for DPIF_UC_ACTION Upcalls. */ + const struct nlattr *actions; /* Flow actions in DPIF_UC_ACTION Upcalls. */ + + bool xout_initialized; /* True if 'xout' must be uninitialized. */ + struct xlate_out xout; /* Result of xlate_actions(). */ + struct ofpbuf odp_actions; /* Datapath actions from xlate_actions(). */ + struct flow_wildcards wc; /* Dependencies that megaflow must match. */ + struct ofpbuf put_actions; /* Actions 'put' in the fastpath. */ + + struct dpif_ipfix *ipfix; /* IPFIX pointer or NULL. */ + struct dpif_sflow *sflow; /* SFlow pointer or NULL. */ + + bool vsp_adjusted; /* 'packet' and 'flow' were adjusted for + VLAN splinters if true. */ + + struct udpif_key *ukey; /* Revalidator flow cache. */ + bool ukey_persists; /* Set true to keep 'ukey' beyond the + lifetime of this upcall. */ + + uint64_t dump_seq; /* udpif->dump_seq at translation time. */ + uint64_t reval_seq; /* udpif->reval_seq at translation time. */ + + /* 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. */ - /* 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(). */ + uint64_t odp_actions_stub[1024 / 8]; /* Stub for odp_actions. */ }; /* '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. + * datapath. They may be created by any handler or revalidator thread at any + * time, and read by any revalidator during the dump phase. They are however + * each owned by a single revalidator which takes care of destroying them + * during the garbage-collection phase. * - * 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. */ + * The mutex within the ukey protects some members of the ukey. The ukey + * itself is protected by RCU and is held within a umap in the parent udpif. + * Adding or removing a ukey from a umap is only safe when holding the + * corresponding umap lock. */ struct udpif_key { - struct hmap_node hmap_node; /* In parent revalidator 'ukeys' map. */ + struct cmap_node cmap_node; /* In parent revalidator 'ukeys' map. */ /* 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'. */ + const struct nlattr *mask; /* Datapath flow mask. */ + size_t mask_len; /* Length of 'mask'. */ + ovs_u128 ufid; /* Unique flow identifier. */ + bool ufid_present; /* True if 'ufid' is in datapath. */ + uint32_t hash; /* Pre-computed hash for 'key'. */ + unsigned pmd_id; /* Datapath poll mode driver id. */ 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. */ - bool mark OVS_GUARDED; /* For mark and sweep garbage - collection. */ + uint64_t dump_seq OVS_GUARDED; /* Tracks udpif->dump_seq. */ + uint64_t reval_seq OVS_GUARDED; /* Tracks udpif->reval_seq. */ bool flow_exists OVS_GUARDED; /* Ensures flows are only deleted once. */ + /* Datapath flow actions as nlattrs. Protected by RCU. Read with + * ukey_get_actions(), and write with ukey_set_actions(). */ + OVSRCU_TYPE(struct ofpbuf *) actions; struct xlate_cache *xcache OVS_GUARDED; /* Cache for xlate entries that * are affected by this ukey. * Used for stats and learning.*/ - struct odputil_keybuf key_buf; /* Memory for 'key'. */ + union { + struct odputil_keybuf buf; + struct nlattr nla; + } keybuf, maskbuf; + + uint32_t key_recirc_id; /* Non-zero if reference is held by the ukey. */ + struct recirc_refs recircs; /* Action recirc IDs with references held. */ +}; + +/* Datapath operation with optional ukey attached. */ +struct ukey_op { + struct udpif_key *ukey; + struct dpif_flow_stats stats; /* Stats for 'op'. */ + struct dpif_op dop; /* Flow operation. */ }; static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); -static struct list all_udpifs = LIST_INITIALIZER(&all_udpifs); +static struct ovs_list all_udpifs = OVS_LIST_INITIALIZER(&all_udpifs); -static size_t read_upcalls(struct handler *, - struct upcall upcalls[UPCALL_MAX_BATCH]); -static void handle_upcalls(struct handler *, struct upcall *, size_t n_upcalls); +static size_t recv_upcalls(struct handler *); +static int process_upcall(struct udpif *, struct upcall *, + struct ofpbuf *odp_actions, struct flow_wildcards *); +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_pause_revalidators(struct udpif *); +static void udpif_resume_revalidators(struct udpif *); static void *udpif_upcall_handler(void *); static void *udpif_revalidator(void *); static unsigned long udpif_get_n_flows(struct udpif *); static void revalidate(struct revalidator *); +static void revalidator_pause(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, @@ -211,27 +297,53 @@ 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_disable_ufid(struct unixctl_conn *, int argc, + const char *argv[], void *aux); +static void upcall_unixctl_enable_ufid(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 struct udpif_key *ukey_create(const struct nlattr *key, size_t key_len, - long long int used); +static void upcall_unixctl_dump_wait(struct unixctl_conn *conn, int argc, + const char *argv[], void *aux); +static void upcall_unixctl_purge(struct unixctl_conn *conn, int argc, + const char *argv[], void *aux); + +static struct udpif_key *ukey_create_from_upcall(struct upcall *, + struct flow_wildcards *); +static int ukey_create_from_dpif_flow(const struct udpif *, + const struct dpif_flow *, + struct udpif_key **); +static void ukey_get_actions(struct udpif_key *, const struct nlattr **actions, + size_t *size); +static bool ukey_install_start(struct udpif *, struct udpif_key *ukey); +static bool ukey_install_finish(struct udpif_key *ukey, int error); +static bool ukey_install(struct udpif *udpif, struct udpif_key *ukey); 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 *); + const ovs_u128 *ufid); +static int ukey_acquire(struct udpif *, const struct dpif_flow *, + struct udpif_key **result, int *error); +static void ukey_delete__(struct udpif_key *); +static void ukey_delete(struct umap *, 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 dp_packet *packet, enum dpif_upcall_type, + const struct nlattr *userdata, const struct flow *, + const unsigned int mru, + const ovs_u128 *ufid, const unsigned pmd_id); +static void upcall_uninit(struct upcall *); + +static upcall_callback upcall_cb; +static dp_purge_callback dp_purge_cb; static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true); +static atomic_bool enable_ufid = ATOMIC_VAR_INIT(true); -struct udpif * -udpif_create(struct dpif_backer *backer, struct dpif *dpif) +void +udpif_init(void) { static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER; - struct udpif *udpif = xzalloc(sizeof *udpif); - if (ovsthread_once_start(&once)) { unixctl_command_register("upcall/show", "", 0, 0, upcall_unixctl_show, NULL); @@ -239,33 +351,79 @@ 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/disable-ufid", "", 0, 0, + upcall_unixctl_disable_ufid, NULL); + unixctl_command_register("upcall/enable-ufid", "", 0, 0, + upcall_unixctl_enable_ufid, 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); + unixctl_command_register("revalidator/purge", "", 0, 0, + upcall_unixctl_purge, NULL); ovsthread_once_done(&once); } +} + +struct udpif * +udpif_create(struct dpif_backer *backer, struct dpif *dpif) +{ + struct udpif *udpif = xzalloc(sizeof *udpif); udpif->dpif = dpif; udpif->backer = backer; 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); + latch_init(&udpif->pause_latch); list_push_back(&all_udpifs, &udpif->list_node); + atomic_init(&udpif->enable_ufid, false); atomic_init(&udpif->n_flows, 0); atomic_init(&udpif->n_flows_timestamp, LLONG_MIN); ovs_mutex_init(&udpif->n_flows_mutex); + udpif->ukeys = xmalloc(N_UMAPS * sizeof *udpif->ukeys); + for (int i = 0; i < N_UMAPS; i++) { + cmap_init(&udpif->ukeys[i].cmap); + ovs_mutex_init(&udpif->ukeys[i].mutex); + } + + dpif_register_upcall_cb(dpif, upcall_cb, udpif); + dpif_register_dp_purge_cb(dpif, dp_purge_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_stop_threads(udpif); + for (int i = 0; i < N_UMAPS; i++) { + cmap_destroy(&udpif->ukeys[i].cmap); + ovs_mutex_destroy(&udpif->ukeys[i].mutex); + } + free(udpif->ukeys); + udpif->ukeys = NULL; + list_remove(&udpif->list_node); latch_destroy(&udpif->exit_latch); + latch_destroy(&udpif->pause_latch); seq_destroy(udpif->reval_seq); seq_destroy(udpif->dump_seq); ovs_mutex_destroy(&udpif->n_flows_mutex); @@ -292,20 +450,20 @@ udpif_stop_threads(struct udpif *udpif) xpthread_join(udpif->revalidators[i].thread, NULL); } + dpif_disable_upcall(udpif->dpif); + for (i = 0; i < udpif->n_revalidators; i++) { struct revalidator *revalidator = &udpif->revalidators[i]; /* Delete ukeys, and delete all flows from the datapath to prevent * double-counting stats. */ revalidator_purge(revalidator); - - hmap_destroy(&udpif->ukeys[i].hmap); - ovs_mutex_destroy(&udpif->ukeys[i].mutex); } latch_poll(&udpif->exit_latch); - xpthread_barrier_destroy(&udpif->reval_barrier); + ovs_barrier_destroy(&udpif->reval_barrier); + ovs_barrier_destroy(&udpif->pause_barrier); free(udpif->revalidators); udpif->revalidators = NULL; @@ -314,9 +472,6 @@ udpif_stop_threads(struct udpif *udpif) free(udpif->handlers); udpif->handlers = NULL; udpif->n_handlers = 0; - - free(udpif->ukeys); - udpif->ukeys = NULL; } } @@ -328,6 +483,7 @@ udpif_start_threads(struct udpif *udpif, size_t n_handlers, { if (udpif && n_handlers && n_revalidators) { size_t i; + bool enable_ufid; udpif->n_handlers = n_handlers; udpif->n_revalidators = n_revalidators; @@ -342,25 +498,49 @@ udpif_start_threads(struct udpif *udpif, size_t n_handlers, "handler", udpif_upcall_handler, handler); } - xpthread_barrier_init(&udpif->reval_barrier, NULL, - udpif->n_revalidators); + enable_ufid = ofproto_dpif_get_enable_ufid(udpif->backer); + atomic_init(&udpif->enable_ufid, enable_ufid); + dpif_enable_upcall(udpif->dpif); + + ovs_barrier_init(&udpif->reval_barrier, udpif->n_revalidators); + ovs_barrier_init(&udpif->pause_barrier, udpif->n_revalidators + 1); udpif->reval_exit = false; + udpif->pause = 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; - 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); } } } +/* Pauses all revalidators. Should only be called by the main thread. + * When function returns, all revalidators are paused and will proceed + * only after udpif_resume_revalidators() is called. */ +static void +udpif_pause_revalidators(struct udpif *udpif) +{ + if (ofproto_dpif_backer_enabled(udpif->backer)) { + latch_set(&udpif->pause_latch); + ovs_barrier_block(&udpif->pause_barrier); + } +} + +/* Resumes the pausing of revalidators. Should only be called by the + * main thread. */ +static void +udpif_resume_revalidators(struct udpif *udpif) +{ + if (ofproto_dpif_backer_enabled(udpif->backer)) { + latch_poll(&udpif->pause_latch); + ovs_barrier_block(&udpif->pause_barrier); + } +} + /* 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 @@ -437,10 +617,8 @@ udpif_get_memory_usage(struct udpif *udpif, struct simap *usage) simap_increase(usage, "handlers", udpif->n_handlers); simap_increase(usage, "revalidators", udpif->n_revalidators); - for (i = 0; i < udpif->n_revalidators; i++) { - 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); + for (i = 0; i < N_UMAPS; i++) { + simap_increase(usage, "udpif keys", cmap_count(&udpif->ukeys[i].cmap)); } } @@ -473,6 +651,15 @@ udpif_flush_all_datapaths(void) } } +static bool +udpif_use_ufid(struct udpif *udpif) +{ + bool enable; + + atomic_read_relaxed(&enable_ufid, &enable); + return enable && ofproto_dpif_get_enable_ufid(udpif->backer); +} + static unsigned long udpif_get_n_flows(struct udpif *udpif) @@ -481,17 +668,17 @@ udpif_get_n_flows(struct udpif *udpif) unsigned long flow_count; now = time_msec(); - atomic_read(&udpif->n_flows_timestamp, &time); + 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(&udpif->n_flows_timestamp, now); + atomic_store_relaxed(&udpif->n_flows_timestamp, now); dpif_get_dp_stats(udpif->dpif, &stats); flow_count = stats.n_flows; - atomic_store(&udpif->n_flows, flow_count); + atomic_store_relaxed(&udpif->n_flows, flow_count); ovs_mutex_unlock(&udpif->n_flows_mutex); } else { - atomic_read(&udpif->n_flows, &flow_count); + atomic_read_relaxed(&udpif->n_flows, &flow_count); } return flow_count; } @@ -506,27 +693,114 @@ udpif_upcall_handler(void *arg) struct udpif *udpif = handler->udpif; while (!latch_is_set(&handler->udpif->exit_latch)) { - struct upcall upcalls[UPCALL_MAX_BATCH]; - size_t n_upcalls, i; - - n_upcalls = read_upcalls(handler, upcalls); - if (!n_upcalls) { + if (recv_upcalls(handler)) { + poll_immediate_wake(); + } else { dpif_recv_wait(udpif->dpif, handler->handler_id); latch_wait(&udpif->exit_latch); - poll_block(); + } + poll_block(); + } + + return NULL; +} + +static size_t +recv_upcalls(struct handler *handler) +{ + 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]; + unsigned int mru; + int error; + + 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; + } + + if (odp_flow_key_to_flow(dupcall->key, dupcall->key_len, flow) + == ODP_FIT_ERROR) { + goto free_dupcall; + } + + if (dupcall->mru) { + mru = nl_attr_get_u16(dupcall->mru); } else { - handle_upcalls(handler, upcalls, n_upcalls); + mru = 0; + } - for (i = 0; i < n_upcalls; i++) { - xlate_out_uninit(&upcalls[i].xout); - ofpbuf_uninit(&upcalls[i].dpif_upcall.packet); - ofpbuf_uninit(&upcalls[i].upcall_buf); + error = upcall_receive(upcall, udpif->backer, &dupcall->packet, + dupcall->type, dupcall->userdata, flow, mru, + &dupcall->ufid, PMD_ID_NULL); + 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, + &dupcall->ufid, PMD_ID_NULL, NULL); + VLOG_INFO_RL(&rl, "received packet on unassociated datapath " + "port %"PRIu32, flow->in_port.odp_port); } + goto free_dupcall; } - coverage_clear(); + + upcall->key = dupcall->key; + upcall->key_len = dupcall->key_len; + upcall->ufid = &dupcall->ufid; + + upcall->out_tun_key = dupcall->out_tun_key; + upcall->actions = dupcall->actions; + + if (vsp_adjust_flow(upcall->ofproto, flow, &dupcall->packet)) { + upcall->vsp_adjusted = true; + } + + pkt_metadata_from_flow(&dupcall->packet.md, flow); + flow_extract(&dupcall->packet, flow); + + error = process_upcall(udpif, upcall, + &upcall->odp_actions, &upcall->wc); + if (error) { + goto cleanup; + } + + n_upcalls++; + continue; + +cleanup: + upcall_uninit(upcall); +free_dupcall: + dp_packet_uninit(&dupcall->packet); + ofpbuf_uninit(recv_buf); } - return NULL; + if (n_upcalls) { + handle_upcalls(handler->udpif, upcalls, n_upcalls); + for (i = 0; i < n_upcalls; i++) { + dp_packet_uninit(&dupcalls[i].packet); + ofpbuf_uninit(&recv_bufs[i]); + upcall_uninit(&upcalls[i]); + } + } + + return n_upcalls; } static void * @@ -540,7 +814,6 @@ udpif_revalidator(void *arg) /* Used only by the leader. */ long long int start_time = 0; uint64_t last_reval_seq = 0; - unsigned int flow_limit = 0; size_t n_flows = 0; revalidator->id = ovsthread_id_self(); @@ -548,14 +821,21 @@ udpif_revalidator(void *arg) if (leader) { uint64_t reval_seq; + recirc_run(); /* Recirculation cleanup. */ + reval_seq = seq_read(udpif->reval_seq); - udpif->need_revalidate = last_reval_seq != reval_seq; 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; + /* Only the leader checks the pause latch to prevent a race where + * some threads think it's false and proceed to block on + * reval_barrier and others think it's true and block indefinitely + * on the pause_barrier */ + udpif->pause = latch_is_set(&udpif->pause_latch); + /* 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 */ @@ -563,32 +843,41 @@ udpif_revalidator(void *arg) start_time = time_msec(); if (!udpif->reval_exit) { - udpif->dump = dpif_flow_dump_create(udpif->dpif); + bool terse_dump; + + terse_dump = udpif_use_ufid(udpif); + udpif->dump = dpif_flow_dump_create(udpif->dpif, terse_dump); } } /* Wait for the leader to start the flow dump. */ - xpthread_barrier_wait(&udpif->reval_barrier); + ovs_barrier_block(&udpif->reval_barrier); + if (udpif->pause) { + revalidator_pause(revalidator); + } + if (udpif->reval_exit) { break; } revalidate(revalidator); /* Wait for all flows to have been dumped before we garbage collect. */ - xpthread_barrier_wait(&udpif->reval_barrier); + ovs_barrier_block(&udpif->reval_barrier); revalidator_sweep(revalidator); /* Wait for all revalidators to finish garbage collection. */ - xpthread_barrier_wait(&udpif->reval_barrier); + ovs_barrier_block(&udpif->reval_barrier); if (leader) { + unsigned int flow_limit; long long int duration; + atomic_read_relaxed(&udpif->flow_limit, &flow_limit); + dpif_flow_dump_destroy(udpif->dump); seq_change(udpif->dump_seq); duration = MAX(time_msec() - start_time, 1); - atomic_read(&udpif->flow_limit, &flow_limit); udpif->dump_duration = duration; if (duration > 2000) { flow_limit /= duration / 1000; @@ -599,7 +888,7 @@ udpif_revalidator(void *arg) flow_limit += 1000; } flow_limit = MIN(ofproto_flow_limit, MAX(flow_limit, 1000)); - atomic_store(&udpif->flow_limit, flow_limit); + atomic_store_relaxed(&udpif->flow_limit, flow_limit); if (duration > 2000) { VLOG_INFO("Spent an unreasonably long %lldms dumping flows", @@ -609,6 +898,7 @@ udpif_revalidator(void *arg) poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500)); seq_wait(udpif->reval_seq, last_reval_seq); latch_wait(&udpif->exit_latch); + latch_wait(&udpif->pause_latch); poll_block(); } } @@ -617,14 +907,13 @@ 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; @@ -633,17 +922,16 @@ 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, @@ -651,7 +939,7 @@ classify_upcall(const struct upcall *upcall) return BAD_UPCALL; } memset(&cookie, 0, sizeof cookie); - memcpy(&cookie, nl_attr_get(dpif_upcall->userdata), userdata_len); + 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; @@ -675,7 +963,7 @@ classify_upcall(const struct upcall *upcall) * initialized with at least 128 bytes of space. */ static void compose_slow_path(struct udpif *udpif, struct xlate_out *xout, - struct flow *flow, odp_port_t odp_in_port, + const struct flow *flow, odp_port_t odp_in_port, struct ofpbuf *buf) { union user_action_cookie cookie; @@ -690,167 +978,323 @@ compose_slow_path(struct udpif *udpif, struct xlate_out *xout, ? ODPP_NONE : odp_in_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); + odp_put_userspace_action(pid, &cookie, sizeof cookie.slow_path, + ODPP_NONE, false, buf); } -static void -upcall_init(struct upcall *upcall, struct flow *flow, struct ofpbuf *packet, - struct ofproto_dpif *ofproto, struct dpif_upcall *dupcall, - odp_port_t odp_in_port) +/* 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 dp_packet *packet, enum dpif_upcall_type type, + const struct nlattr *userdata, const struct flow *flow, + const unsigned int mru, + const ovs_u128 *ufid, const unsigned pmd_id) { - struct pkt_metadata md = pkt_metadata_from_flow(flow); - struct xlate_in xin; + int error; - flow_extract(packet, &md, &upcall->flow); + error = xlate_lookup(backer, flow, &upcall->ofproto, &upcall->ipfix, + &upcall->sflow, NULL, &upcall->in_port); + if (error) { + return error; + } - upcall->ofproto = ofproto; - upcall->key = dupcall->key; - upcall->key_len = dupcall->key_len; - upcall->upcall_type = dupcall->type; - upcall->stats.n_packets = 1; - upcall->stats.n_bytes = ofpbuf_size(packet); - upcall->stats.used = time_msec(); - upcall->stats.tcp_flags = ntohs(upcall->flow.tcp_flags); - upcall->odp_in_port = odp_in_port; + upcall->recirc = NULL; + upcall->have_recirc_ref = false; + upcall->flow = flow; + upcall->packet = packet; + upcall->ufid = ufid; + upcall->pmd_id = pmd_id; + upcall->type = type; + upcall->userdata = userdata; + ofpbuf_use_stub(&upcall->odp_actions, upcall->odp_actions_stub, + sizeof upcall->odp_actions_stub); + ofpbuf_init(&upcall->put_actions, 0); + + upcall->xout_initialized = false; + upcall->vsp_adjusted = false; + upcall->ukey_persists = false; + + upcall->ukey = NULL; + upcall->key = NULL; + upcall->key_len = 0; + upcall->mru = mru; + + upcall->out_tun_key = NULL; + upcall->actions = NULL; + + return 0; +} - xlate_in_init(&xin, upcall->ofproto, &upcall->flow, NULL, - upcall->stats.tcp_flags, packet); +static void +upcall_xlate(struct udpif *udpif, struct upcall *upcall, + struct ofpbuf *odp_actions, struct flow_wildcards *wc) +{ + struct dpif_flow_stats stats; + struct xlate_in xin; - if (upcall->upcall_type == DPIF_UC_MISS) { - xin.resubmit_stats = &upcall->stats; + stats.n_packets = 1; + stats.n_bytes = dp_packet_size(upcall->packet); + stats.used = time_msec(); + stats.tcp_flags = ntohs(upcall->flow->tcp_flags); + + xlate_in_init(&xin, upcall->ofproto, upcall->flow, upcall->in_port, NULL, + stats.tcp_flags, upcall->packet, wc, odp_actions); + + if (upcall->type == DPIF_UC_MISS) { + xin.resubmit_stats = &stats; + + if (xin.recirc) { + /* We may install a datapath flow only if we get a reference to the + * recirculation context (otherwise we could have recirculation + * upcalls using recirculation ID for which no context can be + * found). We may still execute the flow's actions even if we + * don't install the flow. */ + upcall->recirc = xin.recirc; + upcall->have_recirc_ref = recirc_id_node_try_ref_rcu(xin.recirc); + } } else { - /* For non-miss upcalls, there's a flow in the datapath which this - * packet was accounted to. Presumably the revalidators will deal + /* For non-miss upcalls, we are either executing actions (one of which + * is an userspace action) for an upcall, in which case the stats have + * already been taken care of, or there's a flow in the datapath which + * this packet was accounted to. Presumably the revalidators will deal * with pushing its stats eventually. */ } + upcall->dump_seq = seq_read(udpif->dump_seq); + upcall->reval_seq = seq_read(udpif->reval_seq); xlate_actions(&xin, &upcall->xout); -} + upcall->xout_initialized = true; -/* Reads and classifies upcalls. Returns the number of upcalls successfully - * read. */ -static size_t -read_upcalls(struct handler *handler, - struct upcall upcalls[UPCALL_MAX_BATCH]) -{ - struct udpif *udpif = handler->udpif; - size_t i; - size_t n_upcalls = 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 (upcall->xout.fail_open) { + const struct dp_packet *packet = upcall->packet; + struct ofproto_packet_in *pin; + + pin = xmalloc(sizeof *pin); + pin->up.packet = xmemdup(dp_packet_data(packet), dp_packet_size(packet)); + pin->up.packet_len = dp_packet_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.flow_metadata); + 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); + } - /* Try reading UPCALL_MAX_BATCH upcalls from dpif. */ - for (i = 0; i < UPCALL_MAX_BATCH; i++) { - struct upcall *upcall = &upcalls[n_upcalls]; - struct dpif_upcall *dupcall; - struct ofpbuf *packet; - struct ofproto_dpif *ofproto; - struct dpif_sflow *sflow; - struct dpif_ipfix *ipfix; - struct flow flow; - enum upcall_type type; - odp_port_t odp_in_port; - int error; + if (!upcall->xout.slow) { + ofpbuf_use_const(&upcall->put_actions, + odp_actions->data, odp_actions->size); + } else { + /* upcall->put_actions already initialized by upcall_receive(). */ + compose_slow_path(udpif, &upcall->xout, upcall->flow, + upcall->flow->in_port.odp_port, + &upcall->put_actions); + } - 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; - } + /* This function is also called for slow-pathed flows. As we are only + * going to create new datapath flows for actual datapath misses, there is + * no point in creating a ukey otherwise. */ + if (upcall->type == DPIF_UC_MISS) { + upcall->ukey = ukey_create_from_upcall(upcall, wc); + } +} - dupcall = &upcall->dpif_upcall; - packet = &dupcall->packet; - error = xlate_receive(udpif->backer, packet, dupcall->key, - dupcall->key_len, &flow, - &ofproto, &ipfix, &sflow, NULL, &odp_in_port); - 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); +static void +upcall_uninit(struct upcall *upcall) +{ + if (upcall) { + if (upcall->xout_initialized) { + xlate_out_uninit(&upcall->xout); + } + ofpbuf_uninit(&upcall->odp_actions); + ofpbuf_uninit(&upcall->put_actions); + if (upcall->ukey) { + if (!upcall->ukey_persists) { + ukey_delete__(upcall->ukey); } - goto destroy_upcall; + } else if (upcall->have_recirc_ref) { + /* The reference was transferred to the ukey if one was created. */ + recirc_id_node_unref(upcall->recirc); } + } +} - type = classify_upcall(upcall); - if (type == MISS_UPCALL) { - upcall_init(upcall, &flow, packet, ofproto, dupcall, odp_in_port); - n_upcalls++; - continue; - } +static int +upcall_cb(const struct dp_packet *packet, const struct flow *flow, ovs_u128 *ufid, + unsigned pmd_id, enum dpif_upcall_type type, + const struct nlattr *userdata, struct ofpbuf *actions, + struct flow_wildcards *wc, struct ofpbuf *put_actions, void *aux) +{ + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1); + 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); - switch (type) { - case SFLOW_UPCALL: - if (sflow) { - union user_action_cookie cookie; + error = upcall_receive(&upcall, udpif->backer, packet, type, userdata, + flow, 0, ufid, pmd_id); + if (error) { + return error; + } + + error = process_upcall(udpif, &upcall, actions, wc); + if (error) { + goto out; + } + + if (upcall.xout.slow && put_actions) { + ofpbuf_put(put_actions, upcall.put_actions.data, + upcall.put_actions.size); + } - 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); + if (OVS_UNLIKELY(!megaflow)) { + flow_wildcards_init_for_packet(wc, flow); + } + + if (udpif_get_n_flows(udpif) >= flow_limit) { + VLOG_WARN_RL(&rl, "upcall_cb failure: datapath flow limit reached"); + error = ENOSPC; + goto out; + } + + /* Prevent miss flow installation if the key has recirculation ID but we + * were not able to get a reference on it. */ + if (type == DPIF_UC_MISS && upcall.recirc && !upcall.have_recirc_ref) { + VLOG_WARN_RL(&rl, "upcall_cb failure: no reference for recirc flow"); + error = ENOSPC; + goto out; + } + + if (upcall.ukey && !ukey_install(udpif, upcall.ukey)) { + VLOG_WARN_RL(&rl, "upcall_cb failure: ukey installation fails"); + error = ENOSPC; + } +out: + if (!error) { + upcall.ukey_persists = true; + } + upcall_uninit(&upcall); + return error; +} + +static int +process_upcall(struct udpif *udpif, struct upcall *upcall, + struct ofpbuf *odp_actions, struct flow_wildcards *wc) +{ + const struct nlattr *userdata = upcall->userdata; + const struct dp_packet *packet = upcall->packet; + const struct flow *flow = upcall->flow; + + switch (classify_upcall(upcall->type, userdata)) { + case MISS_UPCALL: + upcall_xlate(udpif, upcall, odp_actions, wc); + return 0; + + case SFLOW_UPCALL: + if (upcall->sflow) { + union user_action_cookie cookie; + const struct nlattr *actions; + size_t actions_len = 0; + struct dpif_sflow_actions sflow_actions; + memset(&sflow_actions, 0, sizeof sflow_actions); + memset(&cookie, 0, sizeof cookie); + memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.sflow); + if (upcall->actions) { + /* Actions were passed up from datapath. */ + actions = nl_attr_get(upcall->actions); + actions_len = nl_attr_get_size(upcall->actions); + if (actions && actions_len) { + dpif_sflow_read_actions(flow, actions, actions_len, + &sflow_actions); + } } - break; - case IPFIX_UPCALL: - if (ipfix) { - dpif_ipfix_bridge_sample(ipfix, packet, &flow); + if (actions_len == 0) { + /* Lookup actions in userspace cache. */ + struct udpif_key *ukey = ukey_lookup(udpif, upcall->ufid); + if (ukey) { + ukey_get_actions(ukey, &actions, &actions_len); + dpif_sflow_read_actions(flow, actions, actions_len, + &sflow_actions); + } } - 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); + dpif_sflow_received(upcall->sflow, packet, flow, + flow->in_port.odp_port, &cookie, + actions_len > 0 ? &sflow_actions : NULL); + } + break; + + case IPFIX_UPCALL: + if (upcall->ipfix) { + union user_action_cookie cookie; + struct flow_tnl output_tunnel_key; + + memset(&cookie, 0, sizeof cookie); + memcpy(&cookie, nl_attr_get(userdata), sizeof cookie.ipfix); + + if (upcall->out_tun_key) { + odp_tun_key_from_attr(upcall->out_tun_key, false, + &output_tunnel_key); } - break; - case BAD_UPCALL: - break; - case MISS_UPCALL: - OVS_NOT_REACHED(); + 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; + + case FLOW_SAMPLE_UPCALL: + if (upcall->ipfix) { + union user_action_cookie cookie; - dpif_ipfix_unref(ipfix); - dpif_sflow_unref(sflow); + 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; -destroy_upcall: - ofpbuf_uninit(&upcall->dpif_upcall.packet); - ofpbuf_uninit(&upcall->upcall_buf); + case BAD_UPCALL: + break; } - return n_upcalls; + return EAGAIN; } static void -handle_upcalls(struct handler *handler, struct upcall *upcalls, +handle_upcalls(struct udpif *udpif, struct upcall *upcalls, size_t n_upcalls) { - struct udpif *udpif = handler->udpif; struct dpif_op *opsp[UPCALL_MAX_BATCH * 2]; - struct dpif_op ops[UPCALL_MAX_BATCH * 2]; - size_t n_ops, i; + struct ukey_op ops[UPCALL_MAX_BATCH * 2]; unsigned int flow_limit; - bool fail_open, may_put; + size_t n_ops, n_opsp, i; + bool may_put; + + atomic_read_relaxed(&udpif->flow_limit, &flow_limit); - atomic_read(&udpif->flow_limit, &flow_limit); may_put = udpif_get_n_flows(udpif) < flow_limit; /* Handle the packets individually in order of arrival. @@ -863,32 +1307,25 @@ handle_upcalls(struct handler *handler, struct upcall *upcalls, * * The loop fills 'ops' with an array of operations to execute in the * datapath. */ - fail_open = false; n_ops = 0; for (i = 0; i < n_upcalls; i++) { struct upcall *upcall = &upcalls[i]; - struct ofpbuf *packet = &upcall->dpif_upcall.packet; - struct dpif_op *op; - - fail_open = fail_open || upcall->xout.fail_open; - - if (upcall->flow.in_port.ofp_port - != vsp_realdev_to_vlandev(upcall->ofproto, - upcall->flow.in_port.ofp_port, - upcall->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(&upcall->xout.odp_actions)) { - eth_pop_vlan(packet); + const struct dp_packet *packet = upcall->packet; + struct ukey_op *op; + + 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 (upcall->odp_actions.size) { + eth_pop_vlan(CONST_CAST(struct dp_packet *, upcall->packet)); } /* Remove the flow vlan tags inserted by vlan splinter logic * to ensure megaflow masks generated match the data path flow. */ - upcall->flow.vlan_tci = 0; + CONST_CAST(struct flow *, upcall->flow)->vlan_tci = 0; } /* Do not install a flow into the datapath if: @@ -896,190 +1333,417 @@ handle_upcalls(struct handler *handler, struct upcall *upcalls, * - The datapath already has too many flows. * * - We received this packet via some flow installed in the kernel - * already. */ - if (may_put - && upcall->dpif_upcall.type == DPIF_UC_MISS) { - struct ofpbuf mask; - bool megaflow; - - atomic_read(&enable_megaflows, &megaflow); - ofpbuf_use_stack(&mask, &upcall->mask_buf, sizeof upcall->mask_buf); - if (megaflow) { - 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); - } + * already. + * + * - Upcall was a recirculation but we do not have a reference to + * to the recirculation ID. */ + if (may_put && upcall->type == DPIF_UC_MISS && + (!upcall->recirc || upcall->have_recirc_ref)) { + struct udpif_key *ukey = upcall->ukey; + upcall->ukey_persists = true; 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 = 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 (!upcall->xout.slow) { - op->u.flow_put.actions = ofpbuf_data(&upcall->xout.odp_actions); - op->u.flow_put.actions_len = ofpbuf_size(&upcall->xout.odp_actions); - } else { - struct ofpbuf buf; - - ofpbuf_use_stack(&buf, upcall->slow_path_buf, - sizeof upcall->slow_path_buf); - compose_slow_path(udpif, &upcall->xout, &upcall->flow, - upcall->odp_in_port, &buf); - op->u.flow_put.actions = ofpbuf_data(&buf); - op->u.flow_put.actions_len = ofpbuf_size(&buf); - } - } - if (ofpbuf_size(&upcall->xout.odp_actions)) { + op->ukey = ukey; + op->dop.type = DPIF_OP_FLOW_PUT; + op->dop.u.flow_put.flags = DPIF_FP_CREATE; + op->dop.u.flow_put.key = ukey->key; + op->dop.u.flow_put.key_len = ukey->key_len; + op->dop.u.flow_put.mask = ukey->mask; + op->dop.u.flow_put.mask_len = ukey->mask_len; + op->dop.u.flow_put.ufid = upcall->ufid; + op->dop.u.flow_put.stats = NULL; + ukey_get_actions(ukey, &op->dop.u.flow_put.actions, + &op->dop.u.flow_put.actions_len); + } + if (upcall->odp_actions.size) { op = &ops[n_ops++]; - op->type = DPIF_OP_EXECUTE; - op->u.execute.packet = packet; + op->ukey = NULL; + op->dop.type = DPIF_OP_EXECUTE; + op->dop.u.execute.packet = CONST_CAST(struct dp_packet *, packet); odp_key_to_pkt_metadata(upcall->key, upcall->key_len, - &op->u.execute.md); - 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->dop.u.execute.packet->md); + op->dop.u.execute.actions = upcall->odp_actions.data; + op->dop.u.execute.actions_len = upcall->odp_actions.size; + op->dop.u.execute.needs_help = (upcall->xout.slow & SLOW_ACTION) != 0; + op->dop.u.execute.probe = false; + op->dop.u.execute.mtu = upcall->mru; } } - /* Special case for fail-open mode. + /* Execute batch. * - * 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) { - for (i = 0; i < n_upcalls; i++) { - struct upcall *upcall = &upcalls[i]; - struct ofpbuf *packet = &upcall->dpif_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); + * We install ukeys before installing the flows, locking them for exclusive + * access by this thread for the period of installation. This ensures that + * other threads won't attempt to delete the flows as we are creating them. + */ + n_opsp = 0; + for (i = 0; i < n_ops; i++) { + struct udpif_key *ukey = ops[i].ukey; + + if (ukey) { + /* If we can't install the ukey, don't install the flow. */ + if (!ukey_install_start(udpif, ukey)) { + ukey_delete__(ukey); + ops[i].ukey = NULL; + continue; + } } + opsp[n_opsp++] = &ops[i].dop; } - - /* Execute batch. */ + dpif_operate(udpif->dpif, opsp, n_opsp); for (i = 0; i < n_ops; i++) { - opsp[i] = &ops[i]; + if (ops[i].ukey) { + ukey_install_finish(ops[i].ukey, ops[i].dop.error); + } } - dpif_operate(udpif->dpif, opsp, n_ops); } -/* Must be called with udpif->ukeys[hash % udpif->n_revalidators].mutex. */ +static uint32_t +get_ufid_hash(const ovs_u128 *ufid) +{ + return ufid->u32[0]; +} + static struct udpif_key * -ukey_lookup(struct udpif *udpif, const struct nlattr *key, size_t key_len, - uint32_t hash) - OVS_REQUIRES(udpif->ukeys->mutex) +ukey_lookup(struct udpif *udpif, const ovs_u128 *ufid) { struct udpif_key *ukey; - struct hmap *hmap = &udpif->ukeys[hash % udpif->n_revalidators].hmap; + int idx = get_ufid_hash(ufid) % N_UMAPS; + struct cmap *cmap = &udpif->ukeys[idx].cmap; - HMAP_FOR_EACH_WITH_HASH (ukey, hmap_node, hash, hmap) { - if (ukey->key_len == key_len && !memcmp(ukey->key, key, key_len)) { + CMAP_FOR_EACH_WITH_HASH (ukey, cmap_node, get_ufid_hash(ufid), cmap) { + if (ovs_u128_equals(&ukey->ufid, ufid)) { return ukey; } } return NULL; } -/* Creates a ukey for 'key' and 'key_len', returning it with ukey->mutex in - * a locked state. */ +/* Provides safe lockless access of RCU protected 'ukey->actions'. Callers may + * alternatively access the field directly if they take 'ukey->mutex'. */ +static void +ukey_get_actions(struct udpif_key *ukey, const struct nlattr **actions, size_t *size) +{ + const struct ofpbuf *buf = ovsrcu_get(struct ofpbuf *, &ukey->actions); + *actions = buf->data; + *size = buf->size; +} + +static void +ukey_set_actions(struct udpif_key *ukey, const struct ofpbuf *actions) +{ + ovsrcu_postpone(ofpbuf_delete, + ovsrcu_get_protected(struct ofpbuf *, &ukey->actions)); + ovsrcu_set(&ukey->actions, ofpbuf_clone(actions)); +} + static struct udpif_key * -ukey_create(const struct nlattr *key, size_t key_len, long long int used) +ukey_create__(const struct nlattr *key, size_t key_len, + const struct nlattr *mask, size_t mask_len, + bool ufid_present, const ovs_u128 *ufid, + const unsigned pmd_id, const struct ofpbuf *actions, + uint64_t dump_seq, uint64_t reval_seq, long long int used, + uint32_t key_recirc_id, struct xlate_out *xout) OVS_NO_THREAD_SAFETY_ANALYSIS { struct udpif_key *ukey = xmalloc(sizeof *ukey); - ovs_mutex_init(&ukey->mutex); - ukey->key = (struct nlattr *) &ukey->key_buf; - memcpy(&ukey->key_buf, key, key_len); + memcpy(&ukey->keybuf, key, key_len); + ukey->key = &ukey->keybuf.nla; ukey->key_len = key_len; + memcpy(&ukey->maskbuf, mask, mask_len); + ukey->mask = &ukey->maskbuf.nla; + ukey->mask_len = mask_len; + ukey->ufid_present = ufid_present; + ukey->ufid = *ufid; + ukey->pmd_id = pmd_id; + ukey->hash = get_ufid_hash(&ukey->ufid); - ovs_mutex_lock(&ukey->mutex); - ukey->mark = false; - ukey->flow_exists = true; - ukey->created = used ? used : time_msec(); + ovsrcu_init(&ukey->actions, NULL); + ukey_set_actions(ukey, actions); + + ovs_mutex_init(&ukey->mutex); + ukey->dump_seq = dump_seq; + ukey->reval_seq = reval_seq; + ukey->flow_exists = false; + ukey->created = time_msec(); memset(&ukey->stats, 0, sizeof ukey->stats); + ukey->stats.used = used; ukey->xcache = NULL; + ukey->key_recirc_id = key_recirc_id; + recirc_refs_init(&ukey->recircs); + if (xout) { + /* Take ownership of the action recirc id references. */ + recirc_refs_swap(&ukey->recircs, &xout->recircs); + } + 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. +static struct udpif_key * +ukey_create_from_upcall(struct upcall *upcall, struct flow_wildcards *wc) +{ + struct odputil_keybuf keystub, maskstub; + struct ofpbuf keybuf, maskbuf; + bool megaflow; + struct odp_flow_key_parms odp_parms = { + .flow = upcall->flow, + .mask = &wc->masks, + }; + + odp_parms.support = ofproto_dpif_get_support(upcall->ofproto)->odp; + if (upcall->key_len) { + ofpbuf_use_const(&keybuf, upcall->key, upcall->key_len); + } else { + /* dpif-netdev doesn't provide a netlink-formatted flow key in the + * upcall, so convert the upcall's flow here. */ + ofpbuf_use_stack(&keybuf, &keystub, sizeof keystub); + odp_parms.odp_in_port = upcall->flow->in_port.odp_port; + odp_flow_key_from_flow(&odp_parms, &keybuf); + } + + atomic_read_relaxed(&enable_megaflows, &megaflow); + ofpbuf_use_stack(&maskbuf, &maskstub, sizeof maskstub); + if (megaflow) { + odp_parms.odp_in_port = ODPP_NONE; + odp_parms.key_buf = &keybuf; + + odp_flow_key_from_mask(&odp_parms, &maskbuf); + } + + return ukey_create__(keybuf.data, keybuf.size, maskbuf.data, maskbuf.size, + true, upcall->ufid, upcall->pmd_id, + &upcall->put_actions, upcall->dump_seq, + upcall->reval_seq, 0, + upcall->have_recirc_ref ? upcall->recirc->id : 0, + &upcall->xout); +} + +static int +ukey_create_from_dpif_flow(const struct udpif *udpif, + const struct dpif_flow *flow, + struct udpif_key **ukey) +{ + struct dpif_flow full_flow; + struct ofpbuf actions; + uint64_t dump_seq, reval_seq; + uint64_t stub[DPIF_FLOW_BUFSIZE / 8]; + const struct nlattr *a; + unsigned int left; + + if (!flow->key_len || !flow->actions_len) { + struct ofpbuf buf; + int err; + + /* If the key or actions were not provided by the datapath, fetch the + * full flow. */ + ofpbuf_use_stack(&buf, &stub, sizeof stub); + err = dpif_flow_get(udpif->dpif, NULL, 0, &flow->ufid, + flow->pmd_id, &buf, &full_flow); + if (err) { + return err; + } + flow = &full_flow; + } + + /* Check the flow actions for recirculation action. As recirculation + * relies on OVS userspace internal state, we need to delete all old + * datapath flows with either a non-zero recirc_id in the key, or any + * recirculation actions upon OVS restart. */ + NL_ATTR_FOR_EACH_UNSAFE (a, left, flow->key, flow->key_len) { + if (nl_attr_type(a) == OVS_KEY_ATTR_RECIRC_ID + && nl_attr_get_u32(a) != 0) { + return EINVAL; + } + } + NL_ATTR_FOR_EACH_UNSAFE (a, left, flow->actions, flow->actions_len) { + if (nl_attr_type(a) == OVS_ACTION_ATTR_RECIRC) { + return EINVAL; + } + } + + dump_seq = seq_read(udpif->dump_seq); + reval_seq = seq_read(udpif->reval_seq); + ofpbuf_use_const(&actions, &flow->actions, flow->actions_len); + *ukey = ukey_create__(flow->key, flow->key_len, + flow->mask, flow->mask_len, flow->ufid_present, + &flow->ufid, flow->pmd_id, &actions, dump_seq, + reval_seq, flow->stats.used, 0, NULL); + + return 0; +} + +/* Attempts to insert a ukey into the shared ukey maps. * - * Returns true on success, setting *result to the matching ukey and returning - * it in a locked state. Otherwise, returns false and clears *result. */ + * On success, returns true, installs the ukey and returns it in a locked + * state. Otherwise, returns false. */ 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) +ukey_install_start(struct udpif *udpif, struct udpif_key *new_ukey) + OVS_TRY_LOCK(true, new_ukey->mutex) { - struct udpif_key *ukey; - uint32_t hash, idx; + struct umap *umap; + struct udpif_key *old_ukey; + uint32_t 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)) { + idx = new_ukey->hash % N_UMAPS; + umap = &udpif->ukeys[idx]; + ovs_mutex_lock(&umap->mutex); + old_ukey = ukey_lookup(udpif, &new_ukey->ufid); + if (old_ukey) { + /* Uncommon case: A ukey is already installed with the same UFID. */ + if (old_ukey->key_len == new_ukey->key_len + && !memcmp(old_ukey->key, new_ukey->key, new_ukey->key_len)) { + COVERAGE_INC(handler_duplicate_upcall); + } else { + struct ds ds = DS_EMPTY_INITIALIZER; + + odp_format_ufid(&old_ukey->ufid, &ds); + ds_put_cstr(&ds, " "); + odp_flow_key_format(old_ukey->key, old_ukey->key_len, &ds); + ds_put_cstr(&ds, "\n"); + odp_format_ufid(&new_ukey->ufid, &ds); + ds_put_cstr(&ds, " "); + odp_flow_key_format(new_ukey->key, new_ukey->key_len, &ds); + + VLOG_WARN_RL(&rl, "Conflicting ukey for flows:\n%s", ds_cstr(&ds)); + ds_destroy(&ds); + } + } else { + ovs_mutex_lock(&new_ukey->mutex); + cmap_insert(&umap->cmap, &new_ukey->cmap_node, new_ukey->hash); locked = true; } - ovs_mutex_unlock(&udpif->ukeys[idx].mutex); + ovs_mutex_unlock(&umap->mutex); - if (locked) { - *result = ukey; + return locked; +} + +static void +ukey_install_finish__(struct udpif_key *ukey) OVS_REQUIRES(ukey->mutex) +{ + ukey->flow_exists = true; +} + +static bool +ukey_install_finish(struct udpif_key *ukey, int error) + OVS_RELEASES(ukey->mutex) +{ + if (!error) { + ukey_install_finish__(ukey); + } + ovs_mutex_unlock(&ukey->mutex); + + return !error; +} + +static bool +ukey_install(struct udpif *udpif, struct udpif_key *ukey) +{ + /* The usual way to keep 'ukey->flow_exists' in sync with the datapath is + * to call ukey_install_start(), install the corresponding datapath flow, + * then call ukey_install_finish(). The netdev interface using upcall_cb() + * doesn't provide a function to separately finish the flow installation, + * so we perform the operations together here. + * + * This is fine currently, as revalidator threads will only delete this + * ukey during revalidator_sweep() and only if the dump_seq is mismatched. + * It is unlikely for a revalidator thread to advance dump_seq and reach + * the next GC phase between ukey creation and flow installation. */ + return ukey_install_start(udpif, ukey) && ukey_install_finish(ukey, 0); +} + +/* Searches for a ukey in 'udpif->ukeys' that matches 'flow' and attempts to + * lock the ukey. If the ukey does not exist, create it. + * + * Returns 0 on success, setting *result to the matching ukey and returning it + * in a locked state. Otherwise, returns an errno and clears *result. EBUSY + * indicates that another thread is handling this flow. Other errors indicate + * an unexpected condition creating a new ukey. + * + * *error is an output parameter provided to appease the threadsafety analyser, + * and its value matches the return value. */ +static int +ukey_acquire(struct udpif *udpif, const struct dpif_flow *flow, + struct udpif_key **result, int *error) + OVS_TRY_LOCK(0, (*result)->mutex) +{ + struct udpif_key *ukey; + int retval; + + ukey = ukey_lookup(udpif, &flow->ufid); + if (ukey) { + retval = ovs_mutex_trylock(&ukey->mutex); } else { + /* Usually we try to avoid installing flows from revalidator threads, + * because locking on a umap may cause handler threads to block. + * However there are certain cases, like when ovs-vswitchd is + * restarted, where it is desirable to handle flows that exist in the + * datapath gracefully (ie, don't just clear the datapath). */ + bool install; + + retval = ukey_create_from_dpif_flow(udpif, flow, &ukey); + if (retval) { + goto done; + } + install = ukey_install_start(udpif, ukey); + if (install) { + ukey_install_finish__(ukey); + retval = 0; + } else { + ukey_delete__(ukey); + retval = EBUSY; + } + } + +done: + *error = retval; + if (retval) { *result = NULL; + } else { + *result = ukey; } - return locked; + return retval; } static void -ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey) +ukey_delete__(struct udpif_key *ukey) OVS_NO_THREAD_SAFETY_ANALYSIS { - if (revalidator) { - hmap_remove(revalidator->ukeys, &ukey->hmap_node); + if (ukey) { + if (ukey->key_recirc_id) { + recirc_free_id(ukey->key_recirc_id); + } + recirc_refs_unref(&ukey->recircs); + xlate_cache_delete(ukey->xcache); + ofpbuf_delete(ovsrcu_get(struct ofpbuf *, &ukey->actions)); + ovs_mutex_destroy(&ukey->mutex); + free(ukey); } - xlate_cache_delete(ukey->xcache); - ovs_mutex_destroy(&ukey->mutex); - free(ukey); +} + +static void +ukey_delete(struct umap *umap, struct udpif_key *ukey) + OVS_REQUIRES(umap->mutex) +{ + cmap_remove(&umap->cmap, &ukey->cmap_node, ukey->hash); + ovsrcu_postpone(ukey_delete__, ukey); } static bool -should_revalidate(uint64_t packets, long long int used) +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. @@ -1095,172 +1759,236 @@ should_revalidate(uint64_t packets, long long int used) duration = now - used; metric = duration / packets; - if (metric > 200) { - return false; + if (metric < 200) { + /* The flow is receiving more than ~5pps, so keep it. */ + return true; } - return true; + return false; } -static bool +/* Verifies that the datapath actions of 'ukey' are still correct, and pushes + * 'stats' for it. + * + * Returns a recommended action for 'ukey', options include: + * UKEY_DELETE The ukey should be deleted. + * UKEY_KEEP The ukey is fine as is. + * UKEY_MODIFY The ukey's actions should be changed but is otherwise + * fine. Callers should change the actions to those found + * in the caller supplied 'odp_actions' buffer. The + * recirculation references can be found in 'recircs' and + * must be handled by the caller. + * + * If the result is UKEY_MODIFY, then references to all recirc_ids used by the + * new flow will be held within 'recircs' (which may be none). + * + * The caller is responsible for both initializing 'recircs' prior this call, + * and ensuring any references are eventually freed. + */ +static enum reval_result revalidate_ukey(struct udpif *udpif, struct udpif_key *ukey, - const struct dpif_flow *f) + const struct dpif_flow_stats *stats, + struct ofpbuf *odp_actions, uint64_t reval_seq, + struct recirc_refs *recircs) OVS_REQUIRES(ukey->mutex) { - uint64_t slow_path_buf[128 / 8]; struct xlate_out xout, *xoutp; struct netflow *netflow; struct ofproto_dpif *ofproto; struct dpif_flow_stats push; - struct ofpbuf xout_actions; - struct flow flow, dp_mask; - uint32_t *dp32, *xout32; - odp_port_t odp_in_port; + struct flow flow; + struct flow_wildcards dp_mask, wc; + enum reval_result result; + ofp_port_t ofp_in_port; struct xlate_in xin; long long int last_used; int error; - size_t i; - bool may_learn, ok; + bool need_revalidate; - ok = false; + result = UKEY_DELETE; xoutp = NULL; netflow = NULL; + ofpbuf_clear(odp_actions); + need_revalidate = (ukey->reval_seq != reval_seq); 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 + push.used = stats->used; + push.tcp_flags = stats->tcp_flags; + push.n_packets = (stats->n_packets > ukey->stats.n_packets + ? 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 + push.n_bytes = (stats->n_bytes > ukey->stats.n_bytes + ? stats->n_bytes - ukey->stats.n_bytes : 0); - if (udpif->need_revalidate && last_used - && !should_revalidate(push.n_packets, last_used)) { - ok = false; + if (need_revalidate && last_used + && !should_revalidate(udpif, push.n_packets, last_used)) { goto exit; } /* We will push the stats, so update the ukey stats cache. */ - ukey->stats = f->stats; - if (!push.n_packets && !udpif->need_revalidate) { - ok = true; + ukey->stats = *stats; + if (!push.n_packets && !need_revalidate) { + result = UKEY_KEEP; goto exit; } - may_learn = push.n_packets > 0; - if (ukey->xcache && !udpif->need_revalidate) { - xlate_push_stats(ukey->xcache, may_learn, &push); - ok = true; + if (ukey->xcache && !need_revalidate) { + xlate_push_stats(ukey->xcache, &push); + result = UKEY_KEEP; + goto exit; + } + + if (odp_flow_key_to_flow(ukey->key, ukey->key_len, &flow) + == ODP_FIT_ERROR) { goto exit; } - error = xlate_receive(udpif->backer, NULL, ukey->key, ukey->key_len, &flow, - &ofproto, NULL, NULL, &netflow, &odp_in_port); + error = xlate_lookup(udpif->backer, &flow, &ofproto, NULL, NULL, &netflow, + &ofp_in_port); if (error) { goto exit; } - if (udpif->need_revalidate) { + if (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; + xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL, push.tcp_flags, + NULL, need_revalidate ? &wc : NULL, odp_actions); + if (push.n_packets) { + xin.resubmit_stats = &push; + xin.may_learn = true; + } xin.xcache = ukey->xcache; - xin.may_learn = may_learn; - xin.skip_wildcards = !udpif->need_revalidate; xlate_actions(&xin, &xout); xoutp = &xout; - if (!udpif->need_revalidate) { - ok = true; + if (!need_revalidate) { + result = UKEY_KEEP; goto exit; } - if (!xout.slow) { - 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, &flow, odp_in_port, &xout_actions); + if (xout.slow) { + ofpbuf_clear(odp_actions); + compose_slow_path(udpif, &xout, &flow, flow.in_port.odp_port, + odp_actions); } - if (f->actions_len != ofpbuf_size(&xout_actions) - || memcmp(ofpbuf_data(&xout_actions), f->actions, f->actions_len)) { + if (odp_flow_key_to_mask(ukey->mask, ukey->mask_len, ukey->key, + ukey->key_len, &dp_mask.masks, &flow) + == ODP_FIT_ERROR) { goto exit; } - if (odp_flow_key_to_mask(f->mask, f->mask_len, &dp_mask, &flow) - == ODP_FIT_ERROR) { + /* Do not modify if any bit is wildcarded by the installed datapath flow, + * but not the newly revalidated wildcard mask (wc), i.e., if revalidation + * tells that the datapath flow is now too generic and must be narrowed + * down. Note that we do not know if the datapath has ignored any of the + * wildcarded bits, so we may be overtly conservative here. */ + if (flow_wildcards_has_extra(&dp_mask, &wc)) { goto exit; } - /* Since the kernel is free to ignore wildcarded bits in the mask, we can't - * directly check that the masks are the same. Instead we check that the - * 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. */ - dp32 = (uint32_t *) &dp_mask; - xout32 = (uint32_t *) &xout.wc.masks; - for (i = 0; i < FLOW_U32S; i++) { - if ((dp32[i] | xout32[i]) != dp32[i]) { - goto exit; - } + if (!ofpbuf_equal(odp_actions, + ovsrcu_get(struct ofpbuf *, &ukey->actions))) { + /* The datapath mask was OK, but the actions seem to have changed. + * Let's modify it in place. */ + result = UKEY_MODIFY; + /* Transfer recirc action ID references to the caller. */ + recirc_refs_swap(recircs, &xoutp->recircs); + goto exit; } - ok = true; + + result = UKEY_KEEP; exit: - if (netflow) { - if (!ok) { - netflow_expire(netflow, &flow); - netflow_flow_clear(netflow, &flow); - } - netflow_unref(netflow); + if (result != UKEY_DELETE) { + ukey->reval_seq = reval_seq; + } + if (netflow && result == UKEY_DELETE) { + netflow_flow_clear(netflow, &flow); } xlate_out_uninit(xoutp); - return ok; + return result; } -struct dump_op { - struct udpif_key *ukey; - struct dpif_flow_stats stats; /* Stats for 'op'. */ - struct dpif_op op; /* Flow del operation. */ -}; +static void +delete_op_init__(struct udpif *udpif, struct ukey_op *op, + const struct dpif_flow *flow) +{ + op->ukey = NULL; + op->dop.type = DPIF_OP_FLOW_DEL; + op->dop.u.flow_del.key = flow->key; + op->dop.u.flow_del.key_len = flow->key_len; + op->dop.u.flow_del.ufid = flow->ufid_present ? &flow->ufid : NULL; + op->dop.u.flow_del.pmd_id = flow->pmd_id; + op->dop.u.flow_del.stats = &op->stats; + op->dop.u.flow_del.terse = udpif_use_ufid(udpif); +} + +static void +delete_op_init(struct udpif *udpif, struct ukey_op *op, struct udpif_key *ukey) +{ + op->ukey = ukey; + op->dop.type = DPIF_OP_FLOW_DEL; + op->dop.u.flow_del.key = ukey->key; + op->dop.u.flow_del.key_len = ukey->key_len; + op->dop.u.flow_del.ufid = ukey->ufid_present ? &ukey->ufid : NULL; + op->dop.u.flow_del.pmd_id = ukey->pmd_id; + op->dop.u.flow_del.stats = &op->stats; + op->dop.u.flow_del.terse = udpif_use_ufid(udpif); +} static void -dump_op_init(struct dump_op *op, const struct nlattr *key, size_t key_len, - struct udpif_key *ukey) +modify_op_init(struct ukey_op *op, struct udpif_key *ukey) { 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; + op->dop.type = DPIF_OP_FLOW_PUT; + op->dop.u.flow_put.flags = DPIF_FP_MODIFY; + op->dop.u.flow_put.key = ukey->key; + op->dop.u.flow_put.key_len = ukey->key_len; + op->dop.u.flow_put.mask = ukey->mask; + op->dop.u.flow_put.mask_len = ukey->mask_len; + op->dop.u.flow_put.ufid = &ukey->ufid; + op->dop.u.flow_put.pmd_id = ukey->pmd_id; + op->dop.u.flow_put.stats = NULL; + ukey_get_actions(ukey, &op->dop.u.flow_put.actions, + &op->dop.u.flow_put.actions_len); } static void -push_dump_ops__(struct udpif *udpif, struct dump_op *ops, size_t n_ops) +push_ukey_ops__(struct udpif *udpif, struct ukey_op *ops, size_t n_ops) { struct dpif_op *opsp[REVALIDATE_MAX_BATCH]; size_t i; ovs_assert(n_ops <= REVALIDATE_MAX_BATCH); for (i = 0; i < n_ops; i++) { - opsp[i] = &ops[i].op; + opsp[i] = &ops[i].dop; } dpif_operate(udpif->dpif, opsp, n_ops); for (i = 0; i < n_ops; i++) { - struct dump_op *op = &ops[i]; + struct ukey_op *op = &ops[i]; struct dpif_flow_stats *push, *stats, push_buf; - stats = op->op.u.flow_del.stats; + stats = op->dop.u.flow_del.stats; + push = &push_buf; + + if (op->dop.type != DPIF_OP_FLOW_DEL) { + /* Only deleted flows need their stats pushed. */ + continue; + } + + if (op->dop.error) { + /* flow_del error, 'stats' is unusable. */ + continue; + } + if (op->ukey) { - push = &push_buf; 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; @@ -1272,38 +2000,44 @@ push_dump_ops__(struct udpif *udpif, struct dump_op *ops, size_t n_ops) } if (push->n_packets || netflow_exists()) { + const struct nlattr *key = op->dop.u.flow_del.key; + size_t key_len = op->dop.u.flow_del.key_len; struct ofproto_dpif *ofproto; struct netflow *netflow; + ofp_port_t ofp_in_port; struct flow flow; - bool may_learn; + int error; - may_learn = push->n_packets > 0; if (op->ukey) { ovs_mutex_lock(&op->ukey->mutex); if (op->ukey->xcache) { - xlate_push_stats(op->ukey->xcache, may_learn, push); + xlate_push_stats(op->ukey->xcache, push); ovs_mutex_unlock(&op->ukey->mutex); continue; } ovs_mutex_unlock(&op->ukey->mutex); + key = op->ukey->key; + key_len = op->ukey->key_len; } - 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)) { + if (odp_flow_key_to_flow(key, key_len, &flow) + == ODP_FIT_ERROR) { + continue; + } + + error = xlate_lookup(udpif->backer, &flow, &ofproto, NULL, NULL, + &netflow, &ofp_in_port); + if (!error) { struct xlate_in xin; - xlate_in_init(&xin, ofproto, &flow, NULL, push->tcp_flags, - NULL); + xlate_in_init(&xin, ofproto, &flow, ofp_in_port, NULL, + push->tcp_flags, NULL, NULL, NULL); xin.resubmit_stats = push->n_packets ? push : NULL; - xin.may_learn = may_learn; - xin.skip_wildcards = true; + xin.may_learn = push->n_packets > 0; xlate_actions_for_side_effects(&xin); if (netflow) { - netflow_expire(netflow, &flow); netflow_flow_clear(netflow, &flow); - netflow_unref(netflow); } } } @@ -1311,28 +2045,67 @@ push_dump_ops__(struct udpif *udpif, struct dump_op *ops, size_t n_ops) } static void -push_dump_ops(struct revalidator *revalidator, - struct dump_op *ops, size_t n_ops) +push_ukey_ops(struct udpif *udpif, struct umap *umap, + struct ukey_op *ops, size_t n_ops) { int i; - push_dump_ops__(revalidator->udpif, ops, n_ops); + push_ukey_ops__(udpif, ops, n_ops); + ovs_mutex_lock(&umap->mutex); for (i = 0; i < n_ops; i++) { - ukey_delete(revalidator, ops[i].ukey); + ukey_delete(umap, ops[i].ukey); + } + ovs_mutex_unlock(&umap->mutex); +} + +static void +log_unexpected_flow(const struct dpif_flow *flow, int error) +{ + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(10, 60); + struct ds ds = DS_EMPTY_INITIALIZER; + + ds_put_format(&ds, "Failed to acquire udpif_key corresponding to " + "unexpected flow (%s): ", ovs_strerror(error)); + odp_format_ufid(&flow->ufid, &ds); + VLOG_WARN_RL(&rl, "%s", ds_cstr(&ds)); +} + +static void +reval_op_init(struct ukey_op *op, enum reval_result result, + struct udpif *udpif, struct udpif_key *ukey, + struct recirc_refs *recircs, struct ofpbuf *odp_actions) +{ + if (result == UKEY_DELETE) { + delete_op_init(udpif, op, ukey); + } else if (result == UKEY_MODIFY) { + /* Store the new recircs. */ + recirc_refs_swap(&ukey->recircs, recircs); + /* Release old recircs. */ + recirc_refs_unref(recircs); + /* ukey->key_recirc_id remains, as the key is the same as before. */ + + ukey_set_actions(ukey, odp_actions); + modify_op_init(op, ukey); } } static void revalidate(struct revalidator *revalidator) { + uint64_t odp_actions_stub[1024 / 8]; + struct ofpbuf odp_actions = OFPBUF_STUB_INITIALIZER(odp_actions_stub); + struct udpif *udpif = revalidator->udpif; struct dpif_flow_dump_thread *dump_thread; + uint64_t dump_seq, reval_seq; unsigned int flow_limit; - atomic_read(&udpif->flow_limit, &flow_limit); + dump_seq = seq_read(udpif->dump_seq); + reval_seq = seq_read(udpif->reval_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]; + struct ukey_op ops[REVALIDATE_MAX_BATCH]; int n_ops = 0; struct dpif_flow flows[REVALIDATE_MAX_BATCH]; @@ -1369,22 +2142,35 @@ revalidate(struct revalidator *revalidator) for (f = flows; f < &flows[n_dumped]; f++) { long long int used = f->stats.used; + struct recirc_refs recircs = RECIRC_REFS_EMPTY_INITIALIZER; + enum reval_result result; struct udpif_key *ukey; - bool already_dumped, mark; - - 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(upcall_duplicate_flow); + bool already_dumped; + int error; + + if (ukey_acquire(udpif, f, &ukey, &error)) { + if (error == EBUSY) { + /* Another thread is processing this flow, so don't bother + * processing it.*/ + COVERAGE_INC(upcall_ukey_contention); + } else { + log_unexpected_flow(f, error); + if (error != ENOENT) { + delete_op_init__(udpif, &ops[n_ops++], f); + } + } continue; } - already_dumped = ukey->mark || !ukey->flow_exists; + already_dumped = ukey->dump_seq == dump_seq; if (already_dumped) { - /* The flow has already been dumped and handled by another - * revalidator during this flow dump operation. Skip it. */ - COVERAGE_INC(upcall_duplicate_flow); + /* 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; } @@ -1393,58 +2179,114 @@ revalidate(struct revalidator *revalidator) used = ukey->created; } if (kill_them_all || (used && used < now - max_idle)) { - mark = false; + result = UKEY_DELETE; } else { - mark = revalidate_ukey(udpif, ukey, f); + result = revalidate_ukey(udpif, ukey, &f->stats, &odp_actions, + reval_seq, &recircs); } - ukey->mark = ukey->flow_exists = mark; + ukey->dump_seq = dump_seq; + ukey->flow_exists = result != UKEY_DELETE; - if (!mark) { - dump_op_init(&ops[n_ops++], f->key, f->key_len, ukey); + if (result != UKEY_KEEP) { + /* Takes ownership of 'recircs'. */ + reval_op_init(&ops[n_ops++], result, udpif, ukey, &recircs, + &odp_actions); } ovs_mutex_unlock(&ukey->mutex); } if (n_ops) { - push_dump_ops__(udpif, ops, n_ops); + push_ukey_ops__(udpif, ops, n_ops); } + ovsrcu_quiesce(); } dpif_flow_dump_thread_destroy(dump_thread); + ofpbuf_uninit(&odp_actions); } +/* Pauses the 'revalidator', can only proceed after main thread + * calls udpif_resume_revalidators(). */ static void -revalidator_sweep__(struct revalidator *revalidator, bool purge) - OVS_NO_THREAD_SAFETY_ANALYSIS +revalidator_pause(struct revalidator *revalidator) { - struct dump_op ops[REVALIDATE_MAX_BATCH]; - struct udpif_key *ukey, *next; - size_t n_ops; - - n_ops = 0; + /* The first block is for sync'ing the pause with main thread. */ + ovs_barrier_block(&revalidator->udpif->pause_barrier); + /* The second block is for pausing until main thread resumes. */ + ovs_barrier_block(&revalidator->udpif->pause_barrier); +} - /* 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 (!purge && ukey->mark) { - ukey->mark = false; - } else if (!ukey->flow_exists) { - ukey_delete(revalidator, ukey); - } else { - struct dump_op *op = &ops[n_ops++]; +static void +revalidator_sweep__(struct revalidator *revalidator, bool purge) +{ + struct udpif *udpif; + uint64_t dump_seq, reval_seq; + int slice; + + udpif = revalidator->udpif; + dump_seq = seq_read(udpif->dump_seq); + reval_seq = seq_read(udpif->reval_seq); + slice = revalidator - udpif->revalidators; + ovs_assert(slice < udpif->n_revalidators); + + for (int i = slice; i < N_UMAPS; i += udpif->n_revalidators) { + uint64_t odp_actions_stub[1024 / 8]; + struct ofpbuf odp_actions = OFPBUF_STUB_INITIALIZER(odp_actions_stub); + + struct ukey_op ops[REVALIDATE_MAX_BATCH]; + struct udpif_key *ukey; + struct umap *umap = &udpif->ukeys[i]; + size_t n_ops = 0; + + CMAP_FOR_EACH(ukey, cmap_node, &umap->cmap) { + bool flow_exists, seq_mismatch; + struct recirc_refs recircs = RECIRC_REFS_EMPTY_INITIALIZER; + enum reval_result result; + + /* Handler threads could be holding a ukey lock while it installs a + * new flow, so don't hang around waiting for access to it. */ + if (ovs_mutex_trylock(&ukey->mutex)) { + continue; + } + flow_exists = ukey->flow_exists; + seq_mismatch = (ukey->dump_seq != dump_seq + && ukey->reval_seq != reval_seq); + + if (purge) { + result = UKEY_DELETE; + } else if (!seq_mismatch) { + result = UKEY_KEEP; + } else { + struct dpif_flow_stats stats; + COVERAGE_INC(revalidate_missed_dp_flow); + memset(&stats, 0, sizeof stats); + result = revalidate_ukey(udpif, ukey, &stats, &odp_actions, + reval_seq, &recircs); + } + if (result != UKEY_KEEP) { + /* Takes ownership of 'recircs'. */ + reval_op_init(&ops[n_ops++], result, udpif, ukey, &recircs, + &odp_actions); + } + ovs_mutex_unlock(&ukey->mutex); - /* 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); if (n_ops == REVALIDATE_MAX_BATCH) { - push_dump_ops(revalidator, ops, n_ops); + push_ukey_ops(udpif, umap, ops, n_ops); n_ops = 0; } + + if (!flow_exists) { + ovs_mutex_lock(&umap->mutex); + ukey_delete(umap, ukey); + ovs_mutex_unlock(&umap->mutex); + } + } + + if (n_ops) { + push_ukey_ops(udpif, umap, ops, n_ops); } - } - if (n_ops) { - push_dump_ops(revalidator, ops, n_ops); + ofpbuf_uninit(&odp_actions); + ovsrcu_quiesce(); } } @@ -1459,6 +2301,39 @@ revalidator_purge(struct revalidator *revalidator) { revalidator_sweep__(revalidator, true); } + +/* In reaction to dpif purge, purges all 'ukey's with same 'pmd_id'. */ +static void +dp_purge_cb(void *aux, unsigned pmd_id) +{ + struct udpif *udpif = aux; + size_t i; + + udpif_pause_revalidators(udpif); + for (i = 0; i < N_UMAPS; i++) { + struct ukey_op ops[REVALIDATE_MAX_BATCH]; + struct udpif_key *ukey; + struct umap *umap = &udpif->ukeys[i]; + size_t n_ops = 0; + + CMAP_FOR_EACH(ukey, cmap_node, &umap->cmap) { + if (ukey->pmd_id == pmd_id) { + delete_op_init(udpif, &ops[n_ops++], ukey); + if (n_ops == REVALIDATE_MAX_BATCH) { + push_ukey_ops(udpif, umap, ops, n_ops); + n_ops = 0; + } + } + } + + if (n_ops) { + push_ukey_ops(udpif, umap, ops, n_ops); + } + + ovsrcu_quiesce(); + } + udpif_resume_revalidators(udpif); +} static void upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED, @@ -1469,24 +2344,33 @@ upcall_unixctl_show(struct unixctl_conn *conn, int argc OVS_UNUSED, LIST_FOR_EACH (udpif, list_node, &all_udpifs) { unsigned int flow_limit; + bool ufid_enabled; size_t i; - atomic_read(&udpif->flow_limit, &flow_limit); + atomic_read_relaxed(&udpif->flow_limit, &flow_limit); + ufid_enabled = udpif_use_ufid(udpif); ds_put_format(&ds, "%s:\n", dpif_name(udpif->dpif)); 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, "\tdump duration : %lldms\n", udpif->dump_duration); - + ds_put_format(&ds, "\tufid enabled : "); + if (ufid_enabled) { + ds_put_format(&ds, "true\n"); + } else { + ds_put_format(&ds, "false\n"); + } ds_put_char(&ds, '\n'); + for (i = 0; i < n_revalidators; i++) { struct revalidator *revalidator = &udpif->revalidators[i]; + int j, elements = 0; - 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); + for (j = i; j < N_UMAPS; j += n_revalidators) { + elements += cmap_count(&udpif->ukeys[j].cmap); + } + ds_put_format(&ds, "\t%u: (keys %d)\n", revalidator->id, elements); } } @@ -1504,7 +2388,7 @@ upcall_unixctl_disable_megaflows(struct unixctl_conn *conn, const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED) { - atomic_store(&enable_megaflows, false); + atomic_store_relaxed(&enable_megaflows, false); udpif_flush_all_datapaths(); unixctl_command_reply(conn, "megaflows disabled"); } @@ -1519,11 +2403,36 @@ upcall_unixctl_enable_megaflows(struct unixctl_conn *conn, const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED) { - atomic_store(&enable_megaflows, true); + atomic_store_relaxed(&enable_megaflows, true); udpif_flush_all_datapaths(); unixctl_command_reply(conn, "megaflows enabled"); } +/* Disable skipping flow attributes during flow dump. + * + * This command is only needed for advanced debugging, so it's not + * documented in the man page. */ +static void +upcall_unixctl_disable_ufid(struct unixctl_conn *conn, int argc OVS_UNUSED, + const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED) +{ + atomic_store_relaxed(&enable_ufid, false); + unixctl_command_reply(conn, "Datapath dumping tersely using UFID disabled"); +} + +/* Re-enable skipping flow attributes during flow dump. + * + * This command is only needed for advanced debugging, so it's not documented + * in the man page. */ +static void +upcall_unixctl_enable_ufid(struct unixctl_conn *conn, int argc OVS_UNUSED, + const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED) +{ + atomic_store_relaxed(&enable_ufid, true); + unixctl_command_reply(conn, "Datapath dumping tersely using UFID enabled " + "for supported datapaths"); +} + /* Set the flow limit. * * This command is only needed for advanced debugging, so it's not @@ -1539,9 +2448,45 @@ upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, unsigned int flow_limit = atoi(argv[1]); LIST_FOR_EACH (udpif, list_node, &all_udpifs) { - atomic_store(&udpif->flow_limit, flow_limit); + 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."); + } +} + +static void +upcall_unixctl_purge(struct unixctl_conn *conn, int argc OVS_UNUSED, + const char *argv[] OVS_UNUSED, void *aux OVS_UNUSED) +{ + struct udpif *udpif; + + LIST_FOR_EACH (udpif, list_node, &all_udpifs) { + int n; + + for (n = 0; n < udpif->n_revalidators; n++) { + revalidator_purge(&udpif->revalidators[n]); + } + } + unixctl_command_reply(conn, ""); +}