X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fdpif-netdev.c;h=55712dde1674beb117e1ea70db81d3aec82a2a07;hb=27bbe15dec4e1862396b5c4d265f0ced71b49930;hp=fdea0a77e16420ae25c44d12af11be77c6593b00;hpb=6c3eee823e95d13c613935f0252c8a523bc6ef20;p=sliver-openvswitch.git diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c index fdea0a77e..55712dde1 100644 --- a/lib/dpif-netdev.c +++ b/lib/dpif-netdev.c @@ -43,12 +43,14 @@ #include "list.h" #include "meta-flow.h" #include "netdev.h" +#include "netdev-dpdk.h" #include "netdev-vport.h" #include "netlink.h" #include "odp-execute.h" #include "odp-util.h" #include "ofp-print.h" #include "ofpbuf.h" +#include "ovs-rcu.h" #include "packets.h" #include "poll-loop.h" #include "random.h" @@ -65,15 +67,15 @@ VLOG_DEFINE_THIS_MODULE(dpif_netdev); /* By default, choose a priority in the middle. */ #define NETDEV_RULE_PRIORITY 0x8000 +#define NR_THREADS 1 +/* Use per thread recirc_depth to prevent recirculation loop. */ +#define MAX_RECIRC_DEPTH 5 +DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0) + /* Configuration parameters. */ enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */ -/* Enough headroom to add a vlan tag, plus an extra 2 bytes to allow IP - * headers to be aligned on a 4-byte boundary. */ -enum { DP_NETDEV_HEADROOM = 2 + VLAN_HEADER_LEN }; - /* Queues. */ -enum { N_QUEUES = 2 }; /* Number of queues for dpif_recv(). */ enum { MAX_QUEUE_LEN = 128 }; /* Maximum number of packets per queue. */ enum { QUEUE_MASK = MAX_QUEUE_LEN - 1 }; BUILD_ASSERT_DECL(IS_POW2(MAX_QUEUE_LEN)); @@ -90,14 +92,17 @@ struct dp_netdev_upcall { struct ofpbuf buf; /* ofpbuf instance for upcall.packet. */ }; -/* A queue passing packets from a struct dp_netdev to its clients. +/* A queue passing packets from a struct dp_netdev to its clients (handlers). * * * Thread-safety * ============= * - * Any access at all requires the owning 'dp_netdev''s queue_mutex. */ + * Any access at all requires the owning 'dp_netdev''s queue_rwlock and + * its own mutex. */ struct dp_netdev_queue { + struct ovs_mutex mutex; + struct seq *seq; /* Incremented whenever a packet is queued. */ struct dp_netdev_upcall upcalls[MAX_QUEUE_LEN] OVS_GUARDED; unsigned int head OVS_GUARDED; unsigned int tail OVS_GUARDED; @@ -118,7 +123,7 @@ struct dp_netdev_queue { * port_rwlock * flow_mutex * cls.rwlock - * queue_mutex + * queue_rwlock */ struct dp_netdev { const struct dpif_class *const class; @@ -140,17 +145,17 @@ struct dp_netdev { /* Queues. * - * Everything in 'queues' is protected by 'queue_mutex'. */ - struct ovs_mutex queue_mutex; - struct dp_netdev_queue queues[N_QUEUES]; - struct seq *queue_seq; /* Incremented whenever a packet is queued. */ + * 'queue_rwlock' protects the modification of 'handler_queues' and + * 'n_handlers'. The queue elements are protected by its + * 'handler_queues''s mutex. */ + struct fat_rwlock queue_rwlock; + struct dp_netdev_queue *handler_queues; + uint32_t n_handlers; /* Statistics. * - * ovsthread_counter is internally synchronized. */ - struct ovsthread_counter *n_hit; /* Number of flow table matches. */ - struct ovsthread_counter *n_missed; /* Number of flow table misses. */ - struct ovsthread_counter *n_lost; /* Number of misses not passed up. */ + * ovsthread_stats is internally synchronized. */ + struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */ /* Ports. * @@ -162,21 +167,39 @@ struct dp_netdev { /* Forwarding threads. */ struct latch exit_latch; - struct dp_forwarder *forwarders; - size_t n_forwarders; + struct pmd_thread *pmd_threads; + size_t n_pmd_threads; + int pmd_count; }; static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t) OVS_REQ_RDLOCK(dp->port_rwlock); +enum dp_stat_type { + DP_STAT_HIT, /* Packets that matched in the flow table. */ + DP_STAT_MISS, /* Packets that did not match. */ + DP_STAT_LOST, /* Packets not passed up to the client. */ + DP_N_STATS +}; + +/* Contained by struct dp_netdev's 'stats' member. */ +struct dp_netdev_stats { + struct ovs_mutex mutex; /* Protects 'n'. */ + + /* Indexed by DP_STAT_*, protected by 'mutex'. */ + unsigned long long int n[DP_N_STATS] OVS_GUARDED; +}; + + /* A port in a netdev-based datapath. */ struct dp_netdev_port { struct hmap_node node; /* Node in dp_netdev's 'ports'. */ odp_port_t port_no; struct netdev *netdev; struct netdev_saved_flags *sf; - struct netdev_rx *rx; + struct netdev_rxq **rxq; + struct ovs_refcount ref_cnt; char *type; /* Port type as requested by user. */ }; @@ -231,12 +254,6 @@ struct dp_netdev_flow { const struct hmap_node node; /* In owning dp_netdev's 'flow_table'. */ const struct flow flow; /* The flow that created this entry. */ - /* Number of references. - * The classifier owns one reference. - * Any thread trying to keep a rule from being freed should hold its own - * reference. */ - struct ovs_refcount ref_cnt; - /* Protects members marked OVS_GUARDED. * * Acquire after datapath's flow_mutex. */ @@ -245,22 +262,27 @@ struct dp_netdev_flow { /* Statistics. * * Reading or writing these members requires 'mutex'. */ - long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */ - long long int packet_count OVS_GUARDED; /* Number of packets matched. */ - long long int byte_count OVS_GUARDED; /* Number of bytes matched. */ - uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */ + struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */ /* Actions. * * Reading 'actions' requires 'mutex'. * Writing 'actions' requires 'mutex' and (to allow for transactions) the * datapath's flow_mutex. */ - struct dp_netdev_actions *actions OVS_GUARDED; + OVSRCU_TYPE(struct dp_netdev_actions *) actions; }; -static struct dp_netdev_flow *dp_netdev_flow_ref( - const struct dp_netdev_flow *); -static void dp_netdev_flow_unref(struct dp_netdev_flow *); +static void dp_netdev_flow_free(struct dp_netdev_flow *); + +/* Contained by struct dp_netdev_flow's 'stats' member. */ +struct dp_netdev_flow_stats { + struct ovs_mutex mutex; /* Guards all the other members. */ + + long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */ + long long int packet_count OVS_GUARDED; /* Number of packets matched. */ + long long int byte_count OVS_GUARDED; /* Number of bytes matched. */ + uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */ +}; /* A set of datapath actions within a "struct dp_netdev_flow". * @@ -273,8 +295,6 @@ static void dp_netdev_flow_unref(struct dp_netdev_flow *); * 'flow' is the dp_netdev_flow for which 'flow->actions == actions') or that * owns a reference to 'actions->ref_cnt' (or both). */ struct dp_netdev_actions { - struct ovs_refcount ref_cnt; - /* These members are immutable: they do not change during the struct's * lifetime. */ struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */ @@ -283,17 +303,26 @@ struct dp_netdev_actions { struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *, size_t); -struct dp_netdev_actions *dp_netdev_actions_ref( - const struct dp_netdev_actions *); -void dp_netdev_actions_unref(struct dp_netdev_actions *); +struct dp_netdev_actions *dp_netdev_flow_get_actions( + const struct dp_netdev_flow *); +static void dp_netdev_actions_free(struct dp_netdev_actions *); -/* A thread that receives packets from some ports, looks them up in the flow - * table, and executes the actions it finds. */ -struct dp_forwarder { +/* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate + * the performance overhead of interrupt processing. Therefore netdev can + * not implement rx-wait for these devices. dpif-netdev needs to poll + * these device to check for recv buffer. pmd-thread does polling for + * devices assigned to itself thread. + * + * DPDK used PMD for accessing NIC. + * + * A thread that receives packets from PMD ports, looks them up in the flow + * table, and executes the actions it finds. + **/ +struct pmd_thread { struct dp_netdev *dp; pthread_t thread; - char *name; - uint32_t min_hash, max_hash; + int id; + atomic_uint change_seq; }; /* Interface to netdev-based datapath. */ @@ -317,22 +346,24 @@ static int do_add_port(struct dp_netdev *dp, const char *devname, OVS_REQ_WRLOCK(dp->port_rwlock); static int do_del_port(struct dp_netdev *dp, odp_port_t port_no) OVS_REQ_WRLOCK(dp->port_rwlock); +static void dp_netdev_destroy_all_queues(struct dp_netdev *dp) + OVS_REQ_WRLOCK(dp->queue_rwlock); static int dpif_netdev_open(const struct dpif_class *, const char *name, bool create, struct dpif **); static int dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *, - int queue_no, const struct flow *, - const struct nlattr *userdata) - OVS_EXCLUDED(dp->queue_mutex); + int queue_no, int type, + const struct miniflow *, + const struct nlattr *userdata); static void dp_netdev_execute_actions(struct dp_netdev *dp, - const struct flow *, struct ofpbuf *, + const struct miniflow *, + struct ofpbuf *, bool may_steal, struct pkt_metadata *, const struct nlattr *actions, - size_t actions_len) - OVS_REQ_RDLOCK(dp->port_rwlock); + size_t actions_len); static void dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet, - struct pkt_metadata *) - OVS_REQ_RDLOCK(dp->port_rwlock); -static void dp_netdev_set_threads(struct dp_netdev *, int n); + struct pkt_metadata *); + +static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n); static struct dpif_netdev * dpif_netdev_cast(const struct dpif *dpif) @@ -440,7 +471,6 @@ create_dp_netdev(const char *name, const struct dpif_class *class, { struct dp_netdev *dp; int error; - int i; dp = xzalloc(sizeof *dp); shash_add(&dp_netdevs, name, dp); @@ -448,23 +478,15 @@ create_dp_netdev(const char *name, const struct dpif_class *class, *CONST_CAST(const struct dpif_class **, &dp->class) = class; *CONST_CAST(const char **, &dp->name) = xstrdup(name); ovs_refcount_init(&dp->ref_cnt); - atomic_flag_init(&dp->destroyed); + atomic_flag_clear(&dp->destroyed); ovs_mutex_init(&dp->flow_mutex); classifier_init(&dp->cls, NULL); hmap_init(&dp->flow_table); - ovs_mutex_init(&dp->queue_mutex); - ovs_mutex_lock(&dp->queue_mutex); - for (i = 0; i < N_QUEUES; i++) { - dp->queues[i].head = dp->queues[i].tail = 0; - } - ovs_mutex_unlock(&dp->queue_mutex); - dp->queue_seq = seq_create(); + fat_rwlock_init(&dp->queue_rwlock); - dp->n_hit = ovsthread_counter_create(); - dp->n_missed = ovsthread_counter_create(); - dp->n_lost = ovsthread_counter_create(); + ovsthread_stats_init(&dp->stats); ovs_rwlock_init(&dp->port_rwlock); hmap_init(&dp->ports); @@ -478,7 +500,6 @@ create_dp_netdev(const char *name, const struct dpif_class *class, dp_netdev_free(dp); return error; } - dp_netdev_set_threads(dp, 2); *dpp = dp; return 0; @@ -510,20 +531,21 @@ dpif_netdev_open(const struct dpif_class *class, const char *name, static void dp_netdev_purge_queues(struct dp_netdev *dp) + OVS_REQ_WRLOCK(dp->queue_rwlock) { int i; - ovs_mutex_lock(&dp->queue_mutex); - for (i = 0; i < N_QUEUES; i++) { - struct dp_netdev_queue *q = &dp->queues[i]; + for (i = 0; i < dp->n_handlers; i++) { + struct dp_netdev_queue *q = &dp->handler_queues[i]; + ovs_mutex_lock(&q->mutex); while (q->tail != q->head) { struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK]; ofpbuf_uninit(&u->upcall.packet); ofpbuf_uninit(&u->buf); } + ovs_mutex_unlock(&q->mutex); } - ovs_mutex_unlock(&dp->queue_mutex); } /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp' @@ -533,11 +555,13 @@ dp_netdev_free(struct dp_netdev *dp) OVS_REQUIRES(dp_netdev_mutex) { struct dp_netdev_port *port, *next; + struct dp_netdev_stats *bucket; + int i; shash_find_and_delete(&dp_netdevs, dp->name); - dp_netdev_set_threads(dp, 0); - free(dp->forwarders); + dp_netdev_set_pmd_threads(dp, 0); + free(dp->pmd_threads); dp_netdev_flow_flush(dp); ovs_rwlock_wrlock(&dp->port_rwlock); @@ -545,21 +569,24 @@ dp_netdev_free(struct dp_netdev *dp) do_del_port(dp, port->port_no); } ovs_rwlock_unlock(&dp->port_rwlock); - ovsthread_counter_destroy(dp->n_hit); - ovsthread_counter_destroy(dp->n_missed); - ovsthread_counter_destroy(dp->n_lost); - dp_netdev_purge_queues(dp); - seq_destroy(dp->queue_seq); - ovs_mutex_destroy(&dp->queue_mutex); + OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) { + ovs_mutex_destroy(&bucket->mutex); + free_cacheline(bucket); + } + ovsthread_stats_destroy(&dp->stats); + + fat_rwlock_wrlock(&dp->queue_rwlock); + dp_netdev_destroy_all_queues(dp); + fat_rwlock_unlock(&dp->queue_rwlock); + + fat_rwlock_destroy(&dp->queue_rwlock); classifier_destroy(&dp->cls); hmap_destroy(&dp->flow_table); ovs_mutex_destroy(&dp->flow_mutex); seq_destroy(dp->port_seq); hmap_destroy(&dp->ports); - atomic_flag_destroy(&dp->destroyed); - ovs_refcount_destroy(&dp->ref_cnt); latch_destroy(&dp->exit_latch); free(CONST_CAST(char *, dp->name)); free(dp); @@ -607,20 +634,40 @@ static int dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats) { struct dp_netdev *dp = get_dp_netdev(dpif); + struct dp_netdev_stats *bucket; + size_t i; - ovs_rwlock_rdlock(&dp->cls.rwlock); + fat_rwlock_rdlock(&dp->cls.rwlock); stats->n_flows = hmap_count(&dp->flow_table); - ovs_rwlock_unlock(&dp->cls.rwlock); + fat_rwlock_unlock(&dp->cls.rwlock); - stats->n_hit = ovsthread_counter_read(dp->n_hit); - stats->n_missed = ovsthread_counter_read(dp->n_missed); - stats->n_lost = ovsthread_counter_read(dp->n_lost); + stats->n_hit = stats->n_missed = stats->n_lost = 0; + OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) { + ovs_mutex_lock(&bucket->mutex); + stats->n_hit += bucket->n[DP_STAT_HIT]; + stats->n_missed += bucket->n[DP_STAT_MISS]; + stats->n_lost += bucket->n[DP_STAT_LOST]; + ovs_mutex_unlock(&bucket->mutex); + } stats->n_masks = UINT32_MAX; stats->n_mask_hit = UINT64_MAX; return 0; } +static void +dp_netdev_reload_pmd_threads(struct dp_netdev *dp) +{ + int i; + + for (i = 0; i < dp->n_pmd_threads; i++) { + struct pmd_thread *f = &dp->pmd_threads[i]; + int id; + + atomic_add(&f->change_seq, 1, &id); + } +} + static int do_add_port(struct dp_netdev *dp, const char *devname, const char *type, odp_port_t port_no) @@ -629,10 +676,10 @@ do_add_port(struct dp_netdev *dp, const char *devname, const char *type, struct netdev_saved_flags *sf; struct dp_netdev_port *port; struct netdev *netdev; - struct netdev_rx *rx; enum netdev_flags flags; const char *open_type; int error; + int i; /* XXX reject devices already in some dp_netdev. */ @@ -651,28 +698,40 @@ do_add_port(struct dp_netdev *dp, const char *devname, const char *type, return EINVAL; } - error = netdev_rx_open(netdev, &rx); - if (error - && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) { - VLOG_ERR("%s: cannot receive packets on this network device (%s)", - devname, ovs_strerror(errno)); - netdev_close(netdev); - return error; + port = xzalloc(sizeof *port); + port->port_no = port_no; + port->netdev = netdev; + port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev)); + port->type = xstrdup(type); + for (i = 0; i < netdev_n_rxq(netdev); i++) { + error = netdev_rxq_open(netdev, &port->rxq[i], i); + if (error + && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) { + VLOG_ERR("%s: cannot receive packets on this network device (%s)", + devname, ovs_strerror(errno)); + netdev_close(netdev); + return error; + } } error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf); if (error) { - netdev_rx_close(rx); + for (i = 0; i < netdev_n_rxq(netdev); i++) { + netdev_rxq_close(port->rxq[i]); + } netdev_close(netdev); + free(port->rxq); + free(port); return error; } - - port = xmalloc(sizeof *port); - port->port_no = port_no; - port->netdev = netdev; port->sf = sf; - port->rx = rx; - port->type = xstrdup(type); + + if (netdev_is_pmd(netdev)) { + dp->pmd_count++; + dp_netdev_set_pmd_threads(dp, NR_THREADS); + dp_netdev_reload_pmd_threads(dp); + } + ovs_refcount_init(&port->ref_cnt); hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0)); seq_change(dp->port_seq); @@ -756,6 +815,31 @@ get_port_by_number(struct dp_netdev *dp, } } +static void +port_ref(struct dp_netdev_port *port) +{ + if (port) { + ovs_refcount_ref(&port->ref_cnt); + } +} + +static void +port_unref(struct dp_netdev_port *port) +{ + if (port && ovs_refcount_unref(&port->ref_cnt) == 1) { + int i; + + netdev_close(port->netdev); + netdev_restore_flags(port->sf); + + for (i = 0; i < netdev_n_rxq(port->netdev); i++) { + netdev_rxq_close(port->rxq[i]); + } + free(port->type); + free(port); + } +} + static int get_port_by_name(struct dp_netdev *dp, const char *devname, struct dp_netdev_port **portp) @@ -786,13 +870,11 @@ do_del_port(struct dp_netdev *dp, odp_port_t port_no) hmap_remove(&dp->ports, &port->node); seq_change(dp->port_seq); + if (netdev_is_pmd(port->netdev)) { + dp_netdev_reload_pmd_threads(dp); + } - netdev_close(port->netdev); - netdev_restore_flags(port->sf); - netdev_rx_close(port->rx); - free(port->type); - free(port); - + port_unref(port); return 0; } @@ -841,6 +923,24 @@ dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname, return error; } +static void +dp_netdev_flow_free(struct dp_netdev_flow *flow) +{ + struct dp_netdev_flow_stats *bucket; + size_t i; + + OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) { + ovs_mutex_destroy(&bucket->mutex); + free_cacheline(bucket); + } + ovsthread_stats_destroy(&flow->stats); + + cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr)); + dp_netdev_actions_free(dp_netdev_flow_get_actions(flow)); + ovs_mutex_destroy(&flow->mutex); + free(flow); +} + static void dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow) OVS_REQ_WRLOCK(dp->cls.rwlock) @@ -851,30 +951,7 @@ dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow) classifier_remove(&dp->cls, cr); hmap_remove(&dp->flow_table, node); - dp_netdev_flow_unref(flow); -} - -static struct dp_netdev_flow * -dp_netdev_flow_ref(const struct dp_netdev_flow *flow_) -{ - struct dp_netdev_flow *flow = CONST_CAST(struct dp_netdev_flow *, flow_); - if (flow) { - ovs_refcount_ref(&flow->ref_cnt); - } - return flow; -} - -static void -dp_netdev_flow_unref(struct dp_netdev_flow *flow) -{ - if (flow && ovs_refcount_unref(&flow->ref_cnt) == 1) { - cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr)); - ovs_mutex_lock(&flow->mutex); - dp_netdev_actions_unref(flow->actions); - ovs_mutex_unlock(&flow->mutex); - ovs_mutex_destroy(&flow->mutex); - free(flow); - } + ovsrcu_postpone(dp_netdev_flow_free, flow); } static void @@ -883,11 +960,11 @@ dp_netdev_flow_flush(struct dp_netdev *dp) struct dp_netdev_flow *netdev_flow, *next; ovs_mutex_lock(&dp->flow_mutex); - ovs_rwlock_wrlock(&dp->cls.rwlock); + fat_rwlock_wrlock(&dp->cls.rwlock); HMAP_FOR_EACH_SAFE (netdev_flow, next, node, &dp->flow_table) { dp_netdev_remove_flow(dp, netdev_flow); } - ovs_rwlock_unlock(&dp->cls.rwlock); + fat_rwlock_unlock(&dp->cls.rwlock); ovs_mutex_unlock(&dp->flow_mutex); } @@ -986,15 +1063,16 @@ dp_netdev_flow_cast(const struct cls_rule *cr) } static struct dp_netdev_flow * -dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct flow *flow) +dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key) OVS_EXCLUDED(dp->cls.rwlock) { struct dp_netdev_flow *netdev_flow; + struct cls_rule *rule; - ovs_rwlock_rdlock(&dp->cls.rwlock); - netdev_flow = dp_netdev_flow_cast(classifier_lookup(&dp->cls, flow, NULL)); - dp_netdev_flow_ref(netdev_flow); - ovs_rwlock_unlock(&dp->cls.rwlock); + fat_rwlock_rdlock(&dp->cls.rwlock); + rule = classifier_lookup_miniflow_first(&dp->cls, key); + netdev_flow = dp_netdev_flow_cast(rule); + fat_rwlock_unlock(&dp->cls.rwlock); return netdev_flow; } @@ -1008,7 +1086,7 @@ dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow) HMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0), &dp->flow_table) { if (flow_equal(&netdev_flow->flow, flow)) { - return dp_netdev_flow_ref(netdev_flow); + return netdev_flow; } } @@ -1018,12 +1096,19 @@ dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow) static void get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow, struct dpif_flow_stats *stats) - OVS_REQ_RDLOCK(netdev_flow->mutex) { - stats->n_packets = netdev_flow->packet_count; - stats->n_bytes = netdev_flow->byte_count; - stats->used = netdev_flow->used; - stats->tcp_flags = netdev_flow->tcp_flags; + struct dp_netdev_flow_stats *bucket; + size_t i; + + memset(stats, 0, sizeof *stats); + OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) { + ovs_mutex_lock(&bucket->mutex); + stats->n_packets += bucket->packet_count; + stats->n_bytes += bucket->byte_count; + stats->used = MAX(stats->used, bucket->used); + stats->tcp_flags |= bucket->tcp_flags; + ovs_mutex_unlock(&bucket->mutex); + } } static int @@ -1033,7 +1118,10 @@ dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len, struct flow *mask) { if (mask_key_len) { - if (odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow)) { + enum odp_key_fitness fitness; + + fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow); + if (fitness) { /* This should not happen: it indicates that * odp_flow_key_from_mask() and odp_flow_key_to_mask() * disagree on the acceptable form of a mask. Log the problem @@ -1046,14 +1134,13 @@ dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len, ds_init(&s); odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s, true); - VLOG_ERR("internal error parsing flow mask %s", ds_cstr(&s)); + VLOG_ERR("internal error parsing flow mask %s (%s)", + ds_cstr(&s), odp_key_fitness_to_string(fitness)); ds_destroy(&s); } return EINVAL; } - /* Force unwildcard the in_port. */ - mask->in_port.odp_port = u32_to_odp(UINT32_MAX); } else { enum mf_field_id id; /* No mask key, unwildcard everything except fields whose @@ -1072,6 +1159,14 @@ dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len, } } + /* Force unwildcard the in_port. + * + * We need to do this even in the case where we unwildcard "everything" + * above because "everything" only includes the 16-bit OpenFlow port number + * mask->in_port.ofp_port, which only covers half of the 32-bit datapath + * port number mask->in_port.odp_port. */ + mask->in_port.odp_port = u32_to_odp(UINT32_MAX); + return 0; } @@ -1123,29 +1218,22 @@ dpif_netdev_flow_get(const struct dpif *dpif, return error; } - ovs_rwlock_rdlock(&dp->cls.rwlock); + fat_rwlock_rdlock(&dp->cls.rwlock); netdev_flow = dp_netdev_find_flow(dp, &key); - ovs_rwlock_unlock(&dp->cls.rwlock); + fat_rwlock_unlock(&dp->cls.rwlock); if (netdev_flow) { - struct dp_netdev_actions *actions = NULL; - - ovs_mutex_lock(&netdev_flow->mutex); if (stats) { get_dpif_flow_stats(netdev_flow, stats); } - if (actionsp) { - actions = dp_netdev_actions_ref(netdev_flow->actions); - } - ovs_mutex_unlock(&netdev_flow->mutex); - - dp_netdev_flow_unref(netdev_flow); if (actionsp) { + struct dp_netdev_actions *actions; + + actions = dp_netdev_flow_get_actions(netdev_flow); *actionsp = ofpbuf_clone_data(actions->actions, actions->size); - dp_netdev_actions_unref(actions); } - } else { + } else { error = ENOENT; } @@ -1164,37 +1252,42 @@ dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow, netdev_flow = xzalloc(sizeof *netdev_flow); *CONST_CAST(struct flow *, &netdev_flow->flow) = *flow; - ovs_refcount_init(&netdev_flow->ref_cnt); ovs_mutex_init(&netdev_flow->mutex); - ovs_mutex_lock(&netdev_flow->mutex); - netdev_flow->actions = dp_netdev_actions_create(actions, actions_len); + ovsthread_stats_init(&netdev_flow->stats); + + ovsrcu_set(&netdev_flow->actions, + dp_netdev_actions_create(actions, actions_len)); match_init(&match, flow, wc); cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr), &match, NETDEV_RULE_PRIORITY); - ovs_rwlock_wrlock(&dp->cls.rwlock); + fat_rwlock_wrlock(&dp->cls.rwlock); classifier_insert(&dp->cls, CONST_CAST(struct cls_rule *, &netdev_flow->cr)); hmap_insert(&dp->flow_table, CONST_CAST(struct hmap_node *, &netdev_flow->node), flow_hash(flow, 0)); - ovs_rwlock_unlock(&dp->cls.rwlock); - - ovs_mutex_unlock(&netdev_flow->mutex); + fat_rwlock_unlock(&dp->cls.rwlock); return 0; } static void clear_stats(struct dp_netdev_flow *netdev_flow) - OVS_REQUIRES(netdev_flow->mutex) { - netdev_flow->used = 0; - netdev_flow->packet_count = 0; - netdev_flow->byte_count = 0; - netdev_flow->tcp_flags = 0; + struct dp_netdev_flow_stats *bucket; + size_t i; + + OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) { + ovs_mutex_lock(&bucket->mutex); + bucket->used = 0; + bucket->packet_count = 0; + bucket->byte_count = 0; + bucket->tcp_flags = 0; + ovs_mutex_unlock(&bucket->mutex); + } } static int @@ -1203,6 +1296,7 @@ dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put) struct dp_netdev *dp = get_dp_netdev(dpif); struct dp_netdev_flow *netdev_flow; struct flow flow; + struct miniflow miniflow; struct flow_wildcards wc; int error; @@ -1216,9 +1310,10 @@ dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put) if (error) { return error; } + miniflow_init(&miniflow, &flow); ovs_mutex_lock(&dp->flow_mutex); - netdev_flow = dp_netdev_lookup_flow(dp, &flow); + netdev_flow = dp_netdev_lookup_flow(dp, &miniflow); if (!netdev_flow) { if (put->flags & DPIF_FP_CREATE) { if (hmap_count(&dp->flow_table) < MAX_FLOWS) { @@ -1242,25 +1337,23 @@ dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put) new_actions = dp_netdev_actions_create(put->actions, put->actions_len); - ovs_mutex_lock(&netdev_flow->mutex); - old_actions = netdev_flow->actions; - netdev_flow->actions = new_actions; + old_actions = dp_netdev_flow_get_actions(netdev_flow); + ovsrcu_set(&netdev_flow->actions, new_actions); + if (put->stats) { get_dpif_flow_stats(netdev_flow, put->stats); } if (put->flags & DPIF_FP_ZERO_STATS) { clear_stats(netdev_flow); } - ovs_mutex_unlock(&netdev_flow->mutex); - dp_netdev_actions_unref(old_actions); + ovsrcu_postpone(dp_netdev_actions_free, old_actions); } else if (put->flags & DPIF_FP_CREATE) { error = EEXIST; } else { /* Overlapping flow. */ error = EINVAL; } - dp_netdev_flow_unref(netdev_flow); } ovs_mutex_unlock(&dp->flow_mutex); @@ -1281,121 +1374,151 @@ dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del) } ovs_mutex_lock(&dp->flow_mutex); - ovs_rwlock_wrlock(&dp->cls.rwlock); + fat_rwlock_wrlock(&dp->cls.rwlock); netdev_flow = dp_netdev_find_flow(dp, &key); if (netdev_flow) { if (del->stats) { - ovs_mutex_lock(&netdev_flow->mutex); get_dpif_flow_stats(netdev_flow, del->stats); - ovs_mutex_unlock(&netdev_flow->mutex); } dp_netdev_remove_flow(dp, netdev_flow); } else { error = ENOENT; } - ovs_rwlock_unlock(&dp->cls.rwlock); + fat_rwlock_unlock(&dp->cls.rwlock); ovs_mutex_unlock(&dp->flow_mutex); return error; } struct dp_netdev_flow_state { - uint32_t bucket; - uint32_t offset; struct dp_netdev_actions *actions; struct odputil_keybuf keybuf; struct odputil_keybuf maskbuf; struct dpif_flow_stats stats; }; -static int -dpif_netdev_flow_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep) +struct dp_netdev_flow_iter { + uint32_t bucket; + uint32_t offset; + int status; + struct ovs_mutex mutex; +}; + +static void +dpif_netdev_flow_dump_state_init(void **statep) { struct dp_netdev_flow_state *state; *statep = state = xmalloc(sizeof *state); - state->bucket = 0; - state->offset = 0; state->actions = NULL; +} + +static void +dpif_netdev_flow_dump_state_uninit(void *state_) +{ + struct dp_netdev_flow_state *state = state_; + + free(state); +} + +static int +dpif_netdev_flow_dump_start(const struct dpif *dpif OVS_UNUSED, void **iterp) +{ + struct dp_netdev_flow_iter *iter; + + *iterp = iter = xmalloc(sizeof *iter); + iter->bucket = 0; + iter->offset = 0; + iter->status = 0; + ovs_mutex_init(&iter->mutex); return 0; } +/* XXX the caller must use 'actions' without quiescing */ static int -dpif_netdev_flow_dump_next(const struct dpif *dpif, void *state_, +dpif_netdev_flow_dump_next(const struct dpif *dpif, void *iter_, void *state_, const struct nlattr **key, size_t *key_len, const struct nlattr **mask, size_t *mask_len, const struct nlattr **actions, size_t *actions_len, const struct dpif_flow_stats **stats) { + struct dp_netdev_flow_iter *iter = iter_; struct dp_netdev_flow_state *state = state_; struct dp_netdev *dp = get_dp_netdev(dpif); struct dp_netdev_flow *netdev_flow; - struct hmap_node *node; + struct flow_wildcards wc; + int error; - ovs_rwlock_rdlock(&dp->cls.rwlock); - node = hmap_at_position(&dp->flow_table, &state->bucket, &state->offset); - if (node) { - netdev_flow = CONTAINER_OF(node, struct dp_netdev_flow, node); - dp_netdev_flow_ref(netdev_flow); + ovs_mutex_lock(&iter->mutex); + error = iter->status; + if (!error) { + struct hmap_node *node; + + fat_rwlock_rdlock(&dp->cls.rwlock); + node = hmap_at_position(&dp->flow_table, &iter->bucket, &iter->offset); + if (node) { + netdev_flow = CONTAINER_OF(node, struct dp_netdev_flow, node); + } + fat_rwlock_unlock(&dp->cls.rwlock); + if (!node) { + iter->status = error = EOF; + } } - ovs_rwlock_unlock(&dp->cls.rwlock); - if (!node) { - return EOF; + ovs_mutex_unlock(&iter->mutex); + if (error) { + return error; } + minimask_expand(&netdev_flow->cr.match.mask, &wc); + if (key) { struct ofpbuf buf; ofpbuf_use_stack(&buf, &state->keybuf, sizeof state->keybuf); - odp_flow_key_from_flow(&buf, &netdev_flow->flow, + odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks, netdev_flow->flow.in_port.odp_port); - *key = buf.data; - *key_len = buf.size; + *key = ofpbuf_data(&buf); + *key_len = ofpbuf_size(&buf); } if (key && mask) { struct ofpbuf buf; - struct flow_wildcards wc; ofpbuf_use_stack(&buf, &state->maskbuf, sizeof state->maskbuf); - minimask_expand(&netdev_flow->cr.match.mask, &wc); odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow, - odp_to_u32(wc.masks.in_port.odp_port)); + odp_to_u32(wc.masks.in_port.odp_port), + SIZE_MAX); - *mask = buf.data; - *mask_len = buf.size; + *mask = ofpbuf_data(&buf); + *mask_len = ofpbuf_size(&buf); } if (actions || stats) { - dp_netdev_actions_unref(state->actions); state->actions = NULL; - ovs_mutex_lock(&netdev_flow->mutex); if (actions) { - state->actions = dp_netdev_actions_ref(netdev_flow->actions); + state->actions = dp_netdev_flow_get_actions(netdev_flow); *actions = state->actions->actions; *actions_len = state->actions->size; } + if (stats) { get_dpif_flow_stats(netdev_flow, &state->stats); *stats = &state->stats; } - ovs_mutex_unlock(&netdev_flow->mutex); } - dp_netdev_flow_unref(netdev_flow); - return 0; } static int -dpif_netdev_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_) +dpif_netdev_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *iter_) { - struct dp_netdev_flow_state *state = state_; + struct dp_netdev_flow_iter *iter = iter_; - dp_netdev_actions_unref(state->actions); - free(state); + ovs_mutex_destroy(&iter->mutex); + free(iter); return 0; } @@ -1404,28 +1527,99 @@ dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute) { struct dp_netdev *dp = get_dp_netdev(dpif); struct pkt_metadata *md = &execute->md; - struct flow key; + struct { + struct miniflow flow; + uint32_t buf[FLOW_U32S]; + } key; - if (execute->packet->size < ETH_HEADER_LEN || - execute->packet->size > UINT16_MAX) { + if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN || + ofpbuf_size(execute->packet) > UINT16_MAX) { return EINVAL; } /* Extract flow key. */ - flow_extract(execute->packet, md->skb_priority, md->pkt_mark, &md->tunnel, - (union flow_in_port *)&md->in_port, &key); + miniflow_initialize(&key.flow, key.buf); + miniflow_extract(execute->packet, md, &key.flow); ovs_rwlock_rdlock(&dp->port_rwlock); - dp_netdev_execute_actions(dp, &key, execute->packet, md, execute->actions, - execute->actions_len); + dp_netdev_execute_actions(dp, &key.flow, execute->packet, false, md, + execute->actions, execute->actions_len); ovs_rwlock_unlock(&dp->port_rwlock); return 0; } +static void +dp_netdev_destroy_all_queues(struct dp_netdev *dp) + OVS_REQ_WRLOCK(dp->queue_rwlock) +{ + size_t i; + + dp_netdev_purge_queues(dp); + + for (i = 0; i < dp->n_handlers; i++) { + struct dp_netdev_queue *q = &dp->handler_queues[i]; + + ovs_mutex_destroy(&q->mutex); + seq_destroy(q->seq); + } + free(dp->handler_queues); + dp->handler_queues = NULL; + dp->n_handlers = 0; +} + +static void +dp_netdev_refresh_queues(struct dp_netdev *dp, uint32_t n_handlers) + OVS_REQ_WRLOCK(dp->queue_rwlock) +{ + if (dp->n_handlers != n_handlers) { + size_t i; + + dp_netdev_destroy_all_queues(dp); + + dp->n_handlers = n_handlers; + dp->handler_queues = xzalloc(n_handlers * sizeof *dp->handler_queues); + + for (i = 0; i < n_handlers; i++) { + struct dp_netdev_queue *q = &dp->handler_queues[i]; + + ovs_mutex_init(&q->mutex); + q->seq = seq_create(); + } + } +} + +static int +dpif_netdev_recv_set(struct dpif *dpif, bool enable) +{ + struct dp_netdev *dp = get_dp_netdev(dpif); + + if ((dp->handler_queues != NULL) == enable) { + return 0; + } + + fat_rwlock_wrlock(&dp->queue_rwlock); + if (!enable) { + dp_netdev_destroy_all_queues(dp); + } else { + dp_netdev_refresh_queues(dp, 1); + } + fat_rwlock_unlock(&dp->queue_rwlock); + + return 0; +} + static int -dpif_netdev_recv_set(struct dpif *dpif OVS_UNUSED, bool enable OVS_UNUSED) +dpif_netdev_handlers_set(struct dpif *dpif, uint32_t n_handlers) { + struct dp_netdev *dp = get_dp_netdev(dpif); + + fat_rwlock_wrlock(&dp->queue_rwlock); + if (dp->handler_queues) { + dp_netdev_refresh_queues(dp, n_handlers); + } + fat_rwlock_unlock(&dp->queue_rwlock); + return 0; } @@ -1437,62 +1631,86 @@ dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED, return 0; } -static struct dp_netdev_queue * -find_nonempty_queue(struct dp_netdev *dp) - OVS_REQUIRES(dp->queue_mutex) +static bool +dp_netdev_recv_check(const struct dp_netdev *dp, const uint32_t handler_id) + OVS_REQ_RDLOCK(dp->queue_rwlock) { - int i; + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); - for (i = 0; i < N_QUEUES; i++) { - struct dp_netdev_queue *q = &dp->queues[i]; - if (q->head != q->tail) { - return q; - } + if (!dp->handler_queues) { + VLOG_WARN_RL(&rl, "receiving upcall disabled"); + return false; } - return NULL; + + if (handler_id >= dp->n_handlers) { + VLOG_WARN_RL(&rl, "handler index out of bound"); + return false; + } + + return true; } static int -dpif_netdev_recv(struct dpif *dpif, struct dpif_upcall *upcall, - struct ofpbuf *buf) +dpif_netdev_recv(struct dpif *dpif, uint32_t handler_id, + struct dpif_upcall *upcall, struct ofpbuf *buf) { struct dp_netdev *dp = get_dp_netdev(dpif); struct dp_netdev_queue *q; - int error; + int error = 0; + + fat_rwlock_rdlock(&dp->queue_rwlock); + + if (!dp_netdev_recv_check(dp, handler_id)) { + error = EAGAIN; + goto out; + } - ovs_mutex_lock(&dp->queue_mutex); - q = find_nonempty_queue(dp); - if (q) { + q = &dp->handler_queues[handler_id]; + ovs_mutex_lock(&q->mutex); + if (q->head != q->tail) { struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK]; *upcall = u->upcall; ofpbuf_uninit(buf); *buf = u->buf; - - error = 0; } else { error = EAGAIN; } - ovs_mutex_unlock(&dp->queue_mutex); + ovs_mutex_unlock(&q->mutex); + +out: + fat_rwlock_unlock(&dp->queue_rwlock); return error; } static void -dpif_netdev_recv_wait(struct dpif *dpif) +dpif_netdev_recv_wait(struct dpif *dpif, uint32_t handler_id) { struct dp_netdev *dp = get_dp_netdev(dpif); + struct dp_netdev_queue *q; uint64_t seq; - ovs_mutex_lock(&dp->queue_mutex); - seq = seq_read(dp->queue_seq); - if (find_nonempty_queue(dp)) { + fat_rwlock_rdlock(&dp->queue_rwlock); + + if (!dp_netdev_recv_check(dp, handler_id)) { + goto out; + } + + q = &dp->handler_queues[handler_id]; + ovs_mutex_lock(&q->mutex); + seq = seq_read(q->seq); + if (q->head != q->tail) { poll_immediate_wake(); } else { - seq_wait(dp->queue_seq, seq); + seq_wait(q->seq, seq); } - ovs_mutex_unlock(&dp->queue_mutex); + + ovs_mutex_unlock(&q->mutex); + +out: + fat_rwlock_unlock(&dp->queue_rwlock); } static void @@ -1500,7 +1718,9 @@ dpif_netdev_recv_purge(struct dpif *dpif) { struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif); + fat_rwlock_wrlock(&dpif_netdev->dp->queue_rwlock); dp_netdev_purge_queues(dpif_netdev->dp); + fat_rwlock_unlock(&dpif_netdev->dp->queue_rwlock); } /* Creates and returns a new 'struct dp_netdev_actions', with a reference count @@ -1512,229 +1732,361 @@ dp_netdev_actions_create(const struct nlattr *actions, size_t size) struct dp_netdev_actions *netdev_actions; netdev_actions = xmalloc(sizeof *netdev_actions); - ovs_refcount_init(&netdev_actions->ref_cnt); netdev_actions->actions = xmemdup(actions, size); netdev_actions->size = size; return netdev_actions; } -/* Increments 'actions''s refcount. */ struct dp_netdev_actions * -dp_netdev_actions_ref(const struct dp_netdev_actions *actions_) +dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow) { - struct dp_netdev_actions *actions; + return ovsrcu_get(struct dp_netdev_actions *, &flow->actions); +} + +static void +dp_netdev_actions_free(struct dp_netdev_actions *actions) +{ + free(actions->actions); + free(actions); +} + - actions = CONST_CAST(struct dp_netdev_actions *, actions_); - if (actions) { - ovs_refcount_ref(&actions->ref_cnt); +static void +dp_netdev_process_rxq_port(struct dp_netdev *dp, + struct dp_netdev_port *port, + struct netdev_rxq *rxq) +{ + struct ofpbuf *packet[NETDEV_MAX_RX_BATCH]; + int error, c; + + error = netdev_rxq_recv(rxq, packet, &c); + if (!error) { + struct pkt_metadata md = PKT_METADATA_INITIALIZER(port->port_no); + int i; + + for (i = 0; i < c; i++) { + dp_netdev_port_input(dp, packet[i], &md); + } + } else if (error != EAGAIN && error != EOPNOTSUPP) { + static struct vlog_rate_limit rl + = VLOG_RATE_LIMIT_INIT(1, 5); + + VLOG_ERR_RL(&rl, "error receiving data from %s: %s", + netdev_get_name(port->netdev), + ovs_strerror(error)); } - return actions; } -/* Decrements 'actions''s refcount and frees 'actions' if the refcount reaches - * 0. */ -void -dp_netdev_actions_unref(struct dp_netdev_actions *actions) +static void +dpif_netdev_run(struct dpif *dpif) { - if (actions && ovs_refcount_unref(&actions->ref_cnt) == 1) { - free(actions->actions); - free(actions); + struct dp_netdev_port *port; + struct dp_netdev *dp = get_dp_netdev(dpif); + + ovs_rwlock_rdlock(&dp->port_rwlock); + + HMAP_FOR_EACH (port, node, &dp->ports) { + if (!netdev_is_pmd(port->netdev)) { + int i; + + for (i = 0; i < netdev_n_rxq(port->netdev); i++) { + dp_netdev_process_rxq_port(dp, port, port->rxq[i]); + } + } } + + ovs_rwlock_unlock(&dp->port_rwlock); } - -static void * -dp_forwarder_main(void *f_) + +static void +dpif_netdev_wait(struct dpif *dpif) +{ + struct dp_netdev_port *port; + struct dp_netdev *dp = get_dp_netdev(dpif); + + ovs_rwlock_rdlock(&dp->port_rwlock); + + HMAP_FOR_EACH (port, node, &dp->ports) { + if (!netdev_is_pmd(port->netdev)) { + int i; + + for (i = 0; i < netdev_n_rxq(port->netdev); i++) { + netdev_rxq_wait(port->rxq[i]); + } + } + } + ovs_rwlock_unlock(&dp->port_rwlock); +} + +struct rxq_poll { + struct dp_netdev_port *port; + struct netdev_rxq *rx; +}; + +static int +pmd_load_queues(struct pmd_thread *f, + struct rxq_poll **ppoll_list, int poll_cnt) { - struct dp_forwarder *f = f_; struct dp_netdev *dp = f->dp; - struct ofpbuf packet; + struct rxq_poll *poll_list = *ppoll_list; + struct dp_netdev_port *port; + int id = f->id; + int index; + int i; - f->name = xasprintf("forwarder_%u", ovsthread_id_self()); - set_subprogram_name("%s", f->name); + /* Simple scheduler for netdev rx polling. */ + ovs_rwlock_rdlock(&dp->port_rwlock); + for (i = 0; i < poll_cnt; i++) { + port_unref(poll_list[i].port); + } - ofpbuf_init(&packet, 0); - while (!latch_is_set(&dp->exit_latch)) { - bool received_anything; - int i; + poll_cnt = 0; + index = 0; - ovs_rwlock_rdlock(&dp->port_rwlock); - for (i = 0; i < 50; i++) { - struct dp_netdev_port *port; - - received_anything = false; - HMAP_FOR_EACH (port, node, &f->dp->ports) { - if (port->rx - && port->node.hash >= f->min_hash - && port->node.hash <= f->max_hash) { - int buf_size; - int error; - int mtu; - - if (netdev_get_mtu(port->netdev, &mtu)) { - mtu = ETH_PAYLOAD_MAX; - } - buf_size = DP_NETDEV_HEADROOM + VLAN_ETH_HEADER_LEN + mtu; - - ofpbuf_clear(&packet); - ofpbuf_reserve_with_tailroom(&packet, DP_NETDEV_HEADROOM, - buf_size); - - error = netdev_rx_recv(port->rx, &packet); - if (!error) { - struct pkt_metadata md - = PKT_METADATA_INITIALIZER(port->port_no); - dp_netdev_port_input(dp, &packet, &md); - - received_anything = true; - } else if (error != EAGAIN && error != EOPNOTSUPP) { - static struct vlog_rate_limit rl - = VLOG_RATE_LIMIT_INIT(1, 5); - - VLOG_ERR_RL(&rl, "error receiving data from %s: %s", - netdev_get_name(port->netdev), - ovs_strerror(error)); - } + HMAP_FOR_EACH (port, node, &f->dp->ports) { + if (netdev_is_pmd(port->netdev)) { + int i; + + for (i = 0; i < netdev_n_rxq(port->netdev); i++) { + if ((index % dp->n_pmd_threads) == id) { + poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1)); + + port_ref(port); + poll_list[poll_cnt].port = port; + poll_list[poll_cnt].rx = port->rxq[i]; + poll_cnt++; } + index++; } + } + } - if (!received_anything) { - break; - } + ovs_rwlock_unlock(&dp->port_rwlock); + *ppoll_list = poll_list; + return poll_cnt; +} + +static void * +pmd_thread_main(void *f_) +{ + struct pmd_thread *f = f_; + struct dp_netdev *dp = f->dp; + unsigned int lc = 0; + struct rxq_poll *poll_list; + unsigned int port_seq; + int poll_cnt; + int i; + + poll_cnt = 0; + poll_list = NULL; + + pmd_thread_setaffinity_cpu(f->id); +reload: + poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt); + atomic_read(&f->change_seq, &port_seq); + + for (;;) { + unsigned int c_port_seq; + int i; + + for (i = 0; i < poll_cnt; i++) { + dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx); } - if (received_anything) { - poll_immediate_wake(); - } else { - struct dp_netdev_port *port; + if (lc++ > 1024) { + ovsrcu_quiesce(); - HMAP_FOR_EACH (port, node, &f->dp->ports) - if (port->rx - && port->node.hash >= f->min_hash - && port->node.hash <= f->max_hash) { - netdev_rx_wait(port->rx); - } - seq_wait(dp->port_seq, seq_read(dp->port_seq)); - latch_wait(&dp->exit_latch); + /* TODO: need completely userspace based signaling method. + * to keep this thread entirely in userspace. + * For now using atomic counter. */ + lc = 0; + atomic_read_explicit(&f->change_seq, &c_port_seq, memory_order_consume); + if (c_port_seq != port_seq) { + break; + } } - ovs_rwlock_unlock(&dp->port_rwlock); + } - poll_block(); + if (!latch_is_set(&f->dp->exit_latch)){ + goto reload; } - ofpbuf_uninit(&packet); - free(f->name); + for (i = 0; i < poll_cnt; i++) { + port_unref(poll_list[i].port); + } + free(poll_list); return NULL; } static void -dp_netdev_set_threads(struct dp_netdev *dp, int n) +dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n) { int i; - if (n == dp->n_forwarders) { + if (n == dp->n_pmd_threads) { return; } /* Stop existing threads. */ latch_set(&dp->exit_latch); - for (i = 0; i < dp->n_forwarders; i++) { - struct dp_forwarder *f = &dp->forwarders[i]; + dp_netdev_reload_pmd_threads(dp); + for (i = 0; i < dp->n_pmd_threads; i++) { + struct pmd_thread *f = &dp->pmd_threads[i]; xpthread_join(f->thread, NULL); } latch_poll(&dp->exit_latch); - free(dp->forwarders); + free(dp->pmd_threads); /* Start new threads. */ - dp->forwarders = xmalloc(n * sizeof *dp->forwarders); - dp->n_forwarders = n; + dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads); + dp->n_pmd_threads = n; + for (i = 0; i < n; i++) { - struct dp_forwarder *f = &dp->forwarders[i]; + struct pmd_thread *f = &dp->pmd_threads[i]; f->dp = dp; - f->min_hash = UINT32_MAX / n * i; - f->max_hash = UINT32_MAX / n * (i + 1) - 1; - if (i == n - 1) { - f->max_hash = UINT32_MAX; - } - xpthread_create(&f->thread, NULL, dp_forwarder_main, f); + f->id = i; + atomic_store(&f->change_seq, 1); + + /* Each thread will distribute all devices rx-queues among + * themselves. */ + f->thread = ovs_thread_create("pmd", pmd_thread_main, f); } } + +static void * +dp_netdev_flow_stats_new_cb(void) +{ + struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket); + ovs_mutex_init(&bucket->mutex); + return bucket; +} + static void dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, - const struct ofpbuf *packet) - OVS_REQUIRES(netdev_flow->mutex) + const struct ofpbuf *packet, + const struct miniflow *key) { - netdev_flow->used = time_msec(); - netdev_flow->packet_count++; - netdev_flow->byte_count += packet->size; - netdev_flow->tcp_flags |= packet_get_tcp_flags(packet, &netdev_flow->flow); + uint16_t tcp_flags = miniflow_get_tcp_flags(key); + long long int now = time_msec(); + struct dp_netdev_flow_stats *bucket; + + bucket = ovsthread_stats_bucket_get(&netdev_flow->stats, + dp_netdev_flow_stats_new_cb); + + ovs_mutex_lock(&bucket->mutex); + bucket->used = MAX(now, bucket->used); + bucket->packet_count++; + bucket->byte_count += ofpbuf_size(packet); + bucket->tcp_flags |= tcp_flags; + ovs_mutex_unlock(&bucket->mutex); +} + +static void * +dp_netdev_stats_new_cb(void) +{ + struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket); + ovs_mutex_init(&bucket->mutex); + return bucket; } static void -dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet, - struct pkt_metadata *md) +dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type) +{ + struct dp_netdev_stats *bucket; + + bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb); + ovs_mutex_lock(&bucket->mutex); + bucket->n[type]++; + ovs_mutex_unlock(&bucket->mutex); +} + +static void +dp_netdev_input(struct dp_netdev *dp, struct ofpbuf *packet, + struct pkt_metadata *md) OVS_REQ_RDLOCK(dp->port_rwlock) { struct dp_netdev_flow *netdev_flow; - struct flow key; + struct { + struct miniflow flow; + uint32_t buf[FLOW_U32S]; + } key; - if (packet->size < ETH_HEADER_LEN) { + if (ofpbuf_size(packet) < ETH_HEADER_LEN) { + ofpbuf_delete(packet); return; } - flow_extract(packet, md->skb_priority, md->pkt_mark, &md->tunnel, - (union flow_in_port *)&md->in_port, &key); - netdev_flow = dp_netdev_lookup_flow(dp, &key); + miniflow_initialize(&key.flow, key.buf); + miniflow_extract(packet, md, &key.flow); + + netdev_flow = dp_netdev_lookup_flow(dp, &key.flow); if (netdev_flow) { struct dp_netdev_actions *actions; - ovs_mutex_lock(&netdev_flow->mutex); - dp_netdev_flow_used(netdev_flow, packet); - actions = dp_netdev_actions_ref(netdev_flow->actions); - ovs_mutex_unlock(&netdev_flow->mutex); + dp_netdev_flow_used(netdev_flow, packet, &key.flow); - dp_netdev_execute_actions(dp, &key, packet, md, + actions = dp_netdev_flow_get_actions(netdev_flow); + dp_netdev_execute_actions(dp, &key.flow, packet, true, md, actions->actions, actions->size); - dp_netdev_actions_unref(actions); - ovsthread_counter_inc(dp->n_hit, 1); - } else { - ovsthread_counter_inc(dp->n_missed, 1); - dp_netdev_output_userspace(dp, packet, DPIF_UC_MISS, &key, NULL); + dp_netdev_count_packet(dp, DP_STAT_HIT); + } else if (dp->handler_queues) { + dp_netdev_count_packet(dp, DP_STAT_MISS); + dp_netdev_output_userspace(dp, packet, + miniflow_hash_5tuple(&key.flow, 0) + % dp->n_handlers, + DPIF_UC_MISS, &key.flow, NULL); + ofpbuf_delete(packet); } } +static void +dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet, + struct pkt_metadata *md) + OVS_REQ_RDLOCK(dp->port_rwlock) +{ + uint32_t *recirc_depth = recirc_depth_get(); + + *recirc_depth = 0; + dp_netdev_input(dp, packet, md); +} + static int dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet, - int queue_no, const struct flow *flow, + int queue_no, int type, const struct miniflow *key, const struct nlattr *userdata) - OVS_EXCLUDED(dp->queue_mutex) { - struct dp_netdev_queue *q = &dp->queues[queue_no]; + struct dp_netdev_queue *q; int error; - ovs_mutex_lock(&dp->queue_mutex); + fat_rwlock_rdlock(&dp->queue_rwlock); + q = &dp->handler_queues[queue_no]; + ovs_mutex_lock(&q->mutex); if (q->head - q->tail < MAX_QUEUE_LEN) { struct dp_netdev_upcall *u = &q->upcalls[q->head++ & QUEUE_MASK]; struct dpif_upcall *upcall = &u->upcall; struct ofpbuf *buf = &u->buf; size_t buf_size; + struct flow flow; - upcall->type = queue_no; + upcall->type = type; /* Allocate buffer big enough for everything. */ buf_size = ODPUTIL_FLOW_KEY_BYTES; if (userdata) { buf_size += NLA_ALIGN(userdata->nla_len); } + buf_size += ofpbuf_size(packet); ofpbuf_init(buf, buf_size); /* Put ODP flow. */ - odp_flow_key_from_flow(buf, flow, flow->in_port.odp_port); - upcall->key = buf->data; - upcall->key_len = buf->size; + miniflow_expand(key, &flow); + odp_flow_key_from_flow(buf, &flow, NULL, flow.in_port.odp_port); + upcall->key = ofpbuf_data(buf); + upcall->key_len = ofpbuf_size(buf); /* Put userdata. */ if (userdata) { @@ -1742,43 +2094,44 @@ dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet, NLA_ALIGN(userdata->nla_len)); } - /* Steal packet data. */ - ovs_assert(packet->source == OFPBUF_MALLOC); - upcall->packet = *packet; - ofpbuf_use(packet, NULL, 0); + ofpbuf_set_data(&upcall->packet, + ofpbuf_put(buf, ofpbuf_data(packet), ofpbuf_size(packet))); + ofpbuf_set_size(&upcall->packet, ofpbuf_size(packet)); - seq_change(dp->queue_seq); + seq_change(q->seq); error = 0; } else { - ovsthread_counter_inc(dp->n_lost, 1); + dp_netdev_count_packet(dp, DP_STAT_LOST); error = ENOBUFS; } - ovs_mutex_unlock(&dp->queue_mutex); + ovs_mutex_unlock(&q->mutex); + fat_rwlock_unlock(&dp->queue_rwlock); return error; } struct dp_netdev_execute_aux { struct dp_netdev *dp; - const struct flow *key; + const struct miniflow *key; }; static void dp_execute_cb(void *aux_, struct ofpbuf *packet, - const struct pkt_metadata *md OVS_UNUSED, + struct pkt_metadata *md, const struct nlattr *a, bool may_steal) OVS_NO_THREAD_SAFETY_ANALYSIS { struct dp_netdev_execute_aux *aux = aux_; int type = nl_attr_type(a); struct dp_netdev_port *p; + uint32_t *depth = recirc_depth_get(); switch ((enum ovs_action_attr)type) { case OVS_ACTION_ATTR_OUTPUT: p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a))); if (p) { - netdev_send(p->netdev, packet); + netdev_send(p->netdev, packet, may_steal); } break; @@ -1787,17 +2140,58 @@ dp_execute_cb(void *aux_, struct ofpbuf *packet, userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA); - /* Make a copy if we are not allowed to steal the packet's data. */ - if (!may_steal) { - packet = ofpbuf_clone_with_headroom(packet, DP_NETDEV_HEADROOM); - } - dp_netdev_output_userspace(aux->dp, packet, DPIF_UC_ACTION, aux->key, + dp_netdev_output_userspace(aux->dp, packet, + miniflow_hash_5tuple(aux->key, 0) + % aux->dp->n_handlers, + DPIF_UC_ACTION, aux->key, userdata); - if (!may_steal) { - ofpbuf_uninit(packet); + + if (may_steal) { + ofpbuf_delete(packet); + } + break; + } + + case OVS_ACTION_ATTR_HASH: { + const struct ovs_action_hash *hash_act; + uint32_t hash; + + hash_act = nl_attr_get(a); + if (hash_act->hash_alg == OVS_HASH_ALG_L4) { + /* Hash need not be symmetric, nor does it need to include + * L2 fields. */ + hash = miniflow_hash_5tuple(aux->key, hash_act->hash_basis); + if (!hash) { + hash = 1; /* 0 is not valid */ + } + + } else { + VLOG_WARN("Unknown hash algorithm specified for the hash action."); + hash = 2; } + + md->dp_hash = hash; break; } + + case OVS_ACTION_ATTR_RECIRC: + if (*depth < MAX_RECIRC_DEPTH) { + struct pkt_metadata recirc_md = *md; + struct ofpbuf *recirc_packet; + + recirc_packet = may_steal ? packet : ofpbuf_clone(packet); + recirc_md.recirc_id = nl_attr_get_u32(a); + + (*depth)++; + dp_netdev_input(aux->dp, recirc_packet, &recirc_md); + (*depth)--; + + break; + } else { + VLOG_WARN("Packet dropped. Max recirculation depth exceeded."); + } + break; + case OVS_ACTION_ATTR_PUSH_VLAN: case OVS_ACTION_ATTR_POP_VLAN: case OVS_ACTION_ATTR_PUSH_MPLS: @@ -1811,14 +2205,15 @@ dp_execute_cb(void *aux_, struct ofpbuf *packet, } static void -dp_netdev_execute_actions(struct dp_netdev *dp, const struct flow *key, - struct ofpbuf *packet, struct pkt_metadata *md, +dp_netdev_execute_actions(struct dp_netdev *dp, const struct miniflow *key, + struct ofpbuf *packet, bool may_steal, + struct pkt_metadata *md, const struct nlattr *actions, size_t actions_len) - OVS_REQ_RDLOCK(dp->port_rwlock) { struct dp_netdev_execute_aux aux = {dp, key}; - odp_execute_actions(&aux, packet, md, actions, actions_len, dp_execute_cb); + odp_execute_actions(&aux, packet, may_steal, md, + actions, actions_len, dp_execute_cb); } const struct dpif_class dpif_netdev_class = { @@ -1828,8 +2223,8 @@ const struct dpif_class dpif_netdev_class = { dpif_netdev_open, dpif_netdev_close, dpif_netdev_destroy, - NULL, /* run */ - NULL, /* wait */ + dpif_netdev_run, + dpif_netdev_wait, dpif_netdev_get_stats, dpif_netdev_port_add, dpif_netdev_port_del, @@ -1845,12 +2240,16 @@ const struct dpif_class dpif_netdev_class = { dpif_netdev_flow_put, dpif_netdev_flow_del, dpif_netdev_flow_flush, + dpif_netdev_flow_dump_state_init, dpif_netdev_flow_dump_start, dpif_netdev_flow_dump_next, + NULL, dpif_netdev_flow_dump_done, + dpif_netdev_flow_dump_state_uninit, dpif_netdev_execute, NULL, /* operate */ dpif_netdev_recv_set, + dpif_netdev_handlers_set, dpif_netdev_queue_to_priority, dpif_netdev_recv, dpif_netdev_recv_wait,