X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fdpif-netdev.c;h=f78cda2616bafe614fa7ed526cde817897517f85;hb=8bfd0fdace852f2208b97430bd8de9aeadbf2fb0;hp=c7179ec19deb90fc748caaa5a80d5117450bd87f;hpb=4a38774146f0171a09d5cb1863595b3b546c766e;p=sliver-openvswitch.git diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c index c7179ec19..73eb99d16 100644 --- a/lib/dpif-netdev.c +++ b/lib/dpif-netdev.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2009, 2010 Nicira Networks. + * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -17,7 +17,6 @@ #include #include "dpif.h" -#include #include #include #include @@ -25,123 +24,320 @@ #include #include #include +#include #include #include #include #include #include +#include "classifier.h" #include "csum.h" +#include "dpif.h" #include "dpif-provider.h" +#include "dummy.h" +#include "dynamic-string.h" #include "flow.h" #include "hmap.h" +#include "latch.h" #include "list.h" +#include "meta-flow.h" #include "netdev.h" +#include "netdev-vport.h" +#include "netlink.h" +#include "odp-execute.h" #include "odp-util.h" #include "ofp-print.h" #include "ofpbuf.h" #include "packets.h" #include "poll-loop.h" -#include "queue.h" +#include "random.h" +#include "seq.h" +#include "shash.h" +#include "sset.h" #include "timeval.h" +#include "unixctl.h" #include "util.h" #include "vlog.h" VLOG_DEFINE_THIS_MODULE(dpif_netdev); +/* By default, choose a priority in the middle. */ +#define NETDEV_RULE_PRIORITY 0x8000 + /* Configuration parameters. */ -enum { N_QUEUES = 2 }; /* Number of queues for dpif_recv(). */ -enum { MAX_QUEUE_LEN = 100 }; /* Maximum number of packets per queue. */ -enum { MAX_PORTS = 256 }; /* Maximum number of ports. */ 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 }; -/* Datapath based on the network device interface from netdev.h. */ +/* 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)); + +/* Protects against changes to 'dp_netdevs'. */ +static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER; + +/* Contains all 'struct dp_netdev's. */ +static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex) + = SHASH_INITIALIZER(&dp_netdevs); + +struct dp_netdev_upcall { + struct dpif_upcall upcall; /* Queued upcall information. */ + struct ofpbuf buf; /* ofpbuf instance for upcall.packet. */ +}; + +/* A queue passing packets from a struct dp_netdev to its clients. + * + * + * Thread-safety + * ============= + * + * Any access at all requires the owning 'dp_netdev''s queue_mutex. */ +struct dp_netdev_queue { + struct dp_netdev_upcall upcalls[MAX_QUEUE_LEN] OVS_GUARDED; + unsigned int head OVS_GUARDED; + unsigned int tail OVS_GUARDED; +}; + +/* Datapath based on the network device interface from netdev.h. + * + * + * Thread-safety + * ============= + * + * Some members, marked 'const', are immutable. Accessing other members + * requires synchronization, as noted in more detail below. + * + * Acquisition order is, from outermost to innermost: + * + * dp_netdev_mutex (global) + * port_rwlock + * flow_mutex + * cls.rwlock + * queue_mutex + */ struct dp_netdev { - struct list node; - int dp_idx; - int open_cnt; - bool destroyed; - - bool drop_frags; /* Drop all IP fragments, if true. */ - struct ovs_queue queues[N_QUEUES]; /* Messages queued for dpif_recv(). */ - struct hmap flow_table; /* Flow table. */ - - /* Statistics. */ - long long int n_frags; /* Number of dropped IP fragments. */ - long long int n_hit; /* Number of flow table matches. */ - long long int n_missed; /* Number of flow table misses. */ - long long int n_lost; /* Number of misses not passed to client. */ - - /* Ports. */ - int n_ports; - struct dp_netdev_port *ports[MAX_PORTS]; - struct list port_list; - unsigned int serial; + const struct dpif_class *const class; + const char *const name; + struct ovs_refcount ref_cnt; + atomic_flag destroyed; + + /* Flows. + * + * Readers of 'cls' and 'flow_table' must take a 'cls->rwlock' read lock. + * + * Writers of 'cls' and 'flow_table' must take the 'flow_mutex' and then + * the 'cls->rwlock' write lock. (The outer 'flow_mutex' allows writers to + * atomically perform multiple operations on 'cls' and 'flow_table'.) + */ + struct ovs_mutex flow_mutex; + struct classifier cls; /* Classifier. Protected by cls.rwlock. */ + struct hmap flow_table OVS_GUARDED; /* Flow table. */ + + /* 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. */ + + /* 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. */ + + /* Ports. + * + * Any lookup into 'ports' or any access to the dp_netdev_ports found + * through 'ports' requires taking 'port_rwlock'. */ + struct ovs_rwlock port_rwlock; + struct hmap ports OVS_GUARDED; + struct seq *port_seq; /* Incremented whenever a port changes. */ + + /* Forwarding threads. */ + struct latch exit_latch; + struct dp_forwarder *forwarders; + size_t n_forwarders; }; +static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp, + odp_port_t) + OVS_REQ_RDLOCK(dp->port_rwlock); + /* A port in a netdev-based datapath. */ struct dp_netdev_port { - int port_no; /* Index into dp_netdev's 'ports'. */ - struct list node; /* Element in dp_netdev's 'port_list'. */ + struct hmap_node node; /* Node in dp_netdev's 'ports'. */ + odp_port_t port_no; struct netdev *netdev; - bool internal; /* Internal port (as ODP_PORT_INTERNAL)? */ + struct netdev_saved_flags *sf; + struct netdev_rx *rx; + char *type; /* Port type as requested by user. */ }; -/* A flow in dp_netdev's 'flow_table'. */ +/* A flow in dp_netdev's 'flow_table'. + * + * + * Thread-safety + * ============= + * + * Except near the beginning or ending of its lifespan, rule 'rule' belongs to + * its dp_netdev's classifier. The text below calls this classifier 'cls'. + * + * Motivation + * ---------- + * + * The thread safety rules described here for "struct dp_netdev_flow" are + * motivated by two goals: + * + * - Prevent threads that read members of "struct dp_netdev_flow" from + * reading bad data due to changes by some thread concurrently modifying + * those members. + * + * - Prevent two threads making changes to members of a given "struct + * dp_netdev_flow" from interfering with each other. + * + * + * Rules + * ----- + * + * A flow 'flow' may be accessed without a risk of being freed by code that + * holds a read-lock or write-lock on 'cls->rwlock' or that owns a reference to + * 'flow->ref_cnt' (or both). Code that needs to hold onto a flow for a while + * should take 'cls->rwlock', find the flow it needs, increment 'flow->ref_cnt' + * with dpif_netdev_flow_ref(), and drop 'cls->rwlock'. + * + * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the + * flow from being deleted from 'cls' (that's 'cls->rwlock') and it doesn't + * protect members of 'flow' from modification (that's 'flow->mutex'). + * + * 'flow->mutex' protects the members of 'flow' from modification. It doesn't + * protect the flow from being deleted from 'cls' (that's 'cls->rwlock') and it + * doesn't prevent the flow from being freed (that's 'flow->ref_cnt'). + * + * Some members, marked 'const', are immutable. Accessing other members + * requires synchronization, as noted in more detail below. + */ struct dp_netdev_flow { - struct hmap_node node; /* Element in dp_netdev's 'flow_table'. */ - struct flow key; + /* Packet classification. */ + const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */ + + /* Hash table index by unmasked 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. */ + struct ovs_mutex mutex OVS_ACQ_AFTER(dp_netdev_mutex); + + /* 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. */ + + /* 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; +}; + +static struct dp_netdev_flow *dp_netdev_flow_ref( + const struct dp_netdev_flow *); +static void dp_netdev_flow_unref(struct dp_netdev_flow *); + +/* A set of datapath actions within a "struct dp_netdev_flow". + * + * + * Thread-safety + * ============= + * + * A struct dp_netdev_actions 'actions' may be accessed without a risk of being + * freed by code that holds a read-lock or write-lock on 'flow->mutex' (where + * '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. */ + unsigned int size; /* Size of 'actions', in bytes. */ +}; - /* Statistics. */ - struct timespec used; /* Last used time. */ - long long int packet_count; /* Number of packets matched. */ - long long int byte_count; /* Number of bytes matched. */ - uint16_t tcp_ctl; /* Bitwise-OR of seen tcp_ctl values. */ +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 *); - /* Actions. */ - union odp_action *actions; - unsigned int n_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 { + struct dp_netdev *dp; + pthread_t thread; + char *name; + uint32_t min_hash, max_hash; }; /* Interface to netdev-based datapath. */ struct dpif_netdev { struct dpif dpif; struct dp_netdev *dp; - int listen_mask; - unsigned int dp_serial; + uint64_t last_port_seq; }; -/* All netdev-based datapaths. */ -static struct dp_netdev *dp_netdevs[256]; -struct list dp_netdev_list = LIST_INITIALIZER(&dp_netdev_list); -enum { N_DP_NETDEVS = ARRAY_SIZE(dp_netdevs) }; - -/* Maximum port MTU seen so far. */ -static int max_mtu = ETH_PAYLOAD_MAX; - -static int get_port_by_number(struct dp_netdev *, uint16_t port_no, - struct dp_netdev_port **portp); -static int get_port_by_name(struct dp_netdev *, const char *devname, - struct dp_netdev_port **portp); -static void dp_netdev_free(struct dp_netdev *); +static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no, + struct dp_netdev_port **portp) + OVS_REQ_RDLOCK(dp->port_rwlock); +static int get_port_by_name(struct dp_netdev *dp, const char *devname, + struct dp_netdev_port **portp) + OVS_REQ_RDLOCK(dp->port_rwlock); +static void dp_netdev_free(struct dp_netdev *) + OVS_REQUIRES(dp_netdev_mutex); static void dp_netdev_flow_flush(struct dp_netdev *); -static int do_add_port(struct dp_netdev *, const char *devname, uint16_t flags, - uint16_t port_no); -static int do_del_port(struct dp_netdev *, uint16_t port_no); -static int dp_netdev_output_control(struct dp_netdev *, const struct ofpbuf *, - int queue_no, int port_no, uint32_t arg); -static int dp_netdev_execute_actions(struct dp_netdev *, - struct ofpbuf *, struct flow *, - const union odp_action *, int n); +static int do_add_port(struct dp_netdev *dp, const char *devname, + const char *type, odp_port_t port_no) + 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 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); +static void dp_netdev_execute_actions(struct dp_netdev *dp, + const struct flow *, struct ofpbuf *, + struct pkt_metadata *, + const struct nlattr *actions, + size_t actions_len) + OVS_REQ_RDLOCK(dp->port_rwlock); +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); static struct dpif_netdev * dpif_netdev_cast(const struct dpif *dpif) { - dpif_assert_class(dpif, &dpif_netdev_class); + ovs_assert(dpif->dpif_class->open == dpif_netdev_open); return CONTAINER_OF(dpif, struct dpif_netdev, dpif); } @@ -152,154 +348,243 @@ get_dp_netdev(const struct dpif *dpif) } static int -name_to_dp_idx(const char *name) +dpif_netdev_enumerate(struct sset *all_dps) { - if (!strncmp(name, "dp", 2) && isdigit((unsigned char)name[2])) { - int dp_idx = atoi(name + 2); - if (dp_idx >= 0 && dp_idx < N_DP_NETDEVS) { - return dp_idx; - } + struct shash_node *node; + + ovs_mutex_lock(&dp_netdev_mutex); + SHASH_FOR_EACH(node, &dp_netdevs) { + sset_add(all_dps, node->name); } - return -1; + ovs_mutex_unlock(&dp_netdev_mutex); + + return 0; } -static struct dp_netdev * -find_dp_netdev(const char *name) +static bool +dpif_netdev_class_is_dummy(const struct dpif_class *class) { - int dp_idx; - size_t i; - - dp_idx = name_to_dp_idx(name); - if (dp_idx >= 0) { - return dp_netdevs[dp_idx]; - } + return class != &dpif_netdev_class; +} - for (i = 0; i < N_DP_NETDEVS; i++) { - struct dp_netdev *dp = dp_netdevs[i]; - if (dp) { - struct dp_netdev_port *port; - if (!get_port_by_name(dp, name, &port)) { - return dp; - } - } - } - return NULL; +static const char * +dpif_netdev_port_open_type(const struct dpif_class *class, const char *type) +{ + return strcmp(type, "internal") ? type + : dpif_netdev_class_is_dummy(class) ? "dummy" + : "tap"; } static struct dpif * create_dpif_netdev(struct dp_netdev *dp) { + uint16_t netflow_id = hash_string(dp->name, 0); struct dpif_netdev *dpif; - char *dpname; - dp->open_cnt++; + ovs_refcount_ref(&dp->ref_cnt); - dpname = xasprintf("dp%d", dp->dp_idx); dpif = xmalloc(sizeof *dpif); - dpif_init(&dpif->dpif, &dpif_netdev_class, dpname, dp->dp_idx, dp->dp_idx); + dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id); dpif->dp = dp; - dpif->listen_mask = 0; - dpif->dp_serial = dp->serial; - free(dpname); + dpif->last_port_seq = seq_read(dp->port_seq); return &dpif->dpif; } +/* Choose an unused, non-zero port number and return it on success. + * Return ODPP_NONE on failure. */ +static odp_port_t +choose_port(struct dp_netdev *dp, const char *name) + OVS_REQ_RDLOCK(dp->port_rwlock) +{ + uint32_t port_no; + + if (dp->class != &dpif_netdev_class) { + const char *p; + int start_no = 0; + + /* If the port name begins with "br", start the number search at + * 100 to make writing tests easier. */ + if (!strncmp(name, "br", 2)) { + start_no = 100; + } + + /* If the port name contains a number, try to assign that port number. + * This can make writing unit tests easier because port numbers are + * predictable. */ + for (p = name; *p != '\0'; p++) { + if (isdigit((unsigned char) *p)) { + port_no = start_no + strtol(p, NULL, 10); + if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE) + && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) { + return u32_to_odp(port_no); + } + break; + } + } + } + + for (port_no = 1; port_no <= UINT16_MAX; port_no++) { + if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) { + return u32_to_odp(port_no); + } + } + + return ODPP_NONE; +} + static int -create_dp_netdev(const char *name, int dp_idx, struct dpif **dpifp) +create_dp_netdev(const char *name, const struct dpif_class *class, + struct dp_netdev **dpp) + OVS_REQUIRES(dp_netdev_mutex) { struct dp_netdev *dp; int error; int i; - if (dp_netdevs[dp_idx]) { - return EBUSY; - } + dp = xzalloc(sizeof *dp); + shash_add(&dp_netdevs, name, dp); + + *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); + + ovs_mutex_init(&dp->flow_mutex); + classifier_init(&dp->cls, NULL); + hmap_init(&dp->flow_table); - /* Create datapath. */ - dp_netdevs[dp_idx] = dp = xzalloc(sizeof *dp); - list_push_back(&dp_netdev_list, &dp->node); - dp->dp_idx = dp_idx; - dp->open_cnt = 0; - dp->drop_frags = false; + ovs_mutex_init(&dp->queue_mutex); + ovs_mutex_lock(&dp->queue_mutex); for (i = 0; i < N_QUEUES; i++) { - queue_init(&dp->queues[i]); + dp->queues[i].head = dp->queues[i].tail = 0; } - hmap_init(&dp->flow_table); - list_init(&dp->port_list); - error = do_add_port(dp, name, ODP_PORT_INTERNAL, ODPP_LOCAL); + ovs_mutex_unlock(&dp->queue_mutex); + dp->queue_seq = seq_create(); + + dp->n_hit = ovsthread_counter_create(); + dp->n_missed = ovsthread_counter_create(); + dp->n_lost = ovsthread_counter_create(); + + ovs_rwlock_init(&dp->port_rwlock); + hmap_init(&dp->ports); + dp->port_seq = seq_create(); + latch_init(&dp->exit_latch); + + ovs_rwlock_wrlock(&dp->port_rwlock); + error = do_add_port(dp, name, "internal", ODPP_LOCAL); + ovs_rwlock_unlock(&dp->port_rwlock); if (error) { dp_netdev_free(dp); - return ENODEV; + return error; } + dp_netdev_set_threads(dp, 2); - *dpifp = create_dpif_netdev(dp); + *dpp = dp; return 0; } static int -dpif_netdev_open(const struct dpif_class *class OVS_UNUSED, const char *name, +dpif_netdev_open(const struct dpif_class *class, const char *name, bool create, struct dpif **dpifp) { - if (create) { - if (find_dp_netdev(name)) { - return EEXIST; - } else { - int dp_idx = name_to_dp_idx(name); - if (dp_idx >= 0) { - return create_dp_netdev(name, dp_idx, dpifp); - } else { - /* Scan for unused dp_idx number. */ - for (dp_idx = 0; dp_idx < N_DP_NETDEVS; dp_idx++) { - int error = create_dp_netdev(name, dp_idx, dpifp); - if (error != EBUSY) { - return error; - } - } + struct dp_netdev *dp; + int error; - /* All datapath numbers in use. */ - return ENOBUFS; - } - } + ovs_mutex_lock(&dp_netdev_mutex); + dp = shash_find_data(&dp_netdevs, name); + if (!dp) { + error = create ? create_dp_netdev(name, class, &dp) : ENODEV; } else { - struct dp_netdev *dp = find_dp_netdev(name); - if (dp) { - *dpifp = create_dpif_netdev(dp); - return 0; - } else { - return ENODEV; + error = (dp->class != class ? EINVAL + : create ? EEXIST + : 0); + } + if (!error) { + *dpifp = create_dpif_netdev(dp); + } + ovs_mutex_unlock(&dp_netdev_mutex); + + return error; +} + +static void +dp_netdev_purge_queues(struct dp_netdev *dp) +{ + int i; + + ovs_mutex_lock(&dp->queue_mutex); + for (i = 0; i < N_QUEUES; i++) { + struct dp_netdev_queue *q = &dp->queues[i]; + + 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(&dp->queue_mutex); } +/* Requires dp_netdev_mutex so that we can't get a new reference to 'dp' + * through the 'dp_netdevs' shash while freeing 'dp'. */ static void dp_netdev_free(struct dp_netdev *dp) + OVS_REQUIRES(dp_netdev_mutex) { - int i; + struct dp_netdev_port *port, *next; + + shash_find_and_delete(&dp_netdevs, dp->name); + + dp_netdev_set_threads(dp, 0); + free(dp->forwarders); dp_netdev_flow_flush(dp); - while (dp->n_ports > 0) { - struct dp_netdev_port *port = CONTAINER_OF( - dp->port_list.next, struct dp_netdev_port, node); + ovs_rwlock_wrlock(&dp->port_rwlock); + HMAP_FOR_EACH_SAFE (port, next, node, &dp->ports) { do_del_port(dp, port->port_no); } - for (i = 0; i < N_QUEUES; i++) { - queue_destroy(&dp->queues[i]); - } + 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); + + classifier_destroy(&dp->cls); hmap_destroy(&dp->flow_table); - dp_netdevs[dp->dp_idx] = NULL; - list_remove(&dp->node); + 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); } +static void +dp_netdev_unref(struct dp_netdev *dp) +{ + if (dp) { + /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't + * get a new reference to 'dp' through the 'dp_netdevs' shash. */ + ovs_mutex_lock(&dp_netdev_mutex); + if (ovs_refcount_unref(&dp->ref_cnt) == 1) { + dp_netdev_free(dp); + } + ovs_mutex_unlock(&dp_netdev_mutex); + } +} + static void dpif_netdev_close(struct dpif *dpif) { struct dp_netdev *dp = get_dp_netdev(dpif); - assert(dp->open_cnt > 0); - if (--dp->open_cnt == 0 && dp->destroyed) { - dp_netdev_free(dp); - } + + dp_netdev_unref(dp); free(dpif); } @@ -307,75 +592,77 @@ static int dpif_netdev_destroy(struct dpif *dpif) { struct dp_netdev *dp = get_dp_netdev(dpif); - dp->destroyed = true; + + if (!atomic_flag_test_and_set(&dp->destroyed)) { + if (ovs_refcount_unref(&dp->ref_cnt) == 1) { + /* Can't happen: 'dpif' still owns a reference to 'dp'. */ + OVS_NOT_REACHED(); + } + } + return 0; } static int -dpif_netdev_get_stats(const struct dpif *dpif, struct odp_stats *stats) +dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats) { struct dp_netdev *dp = get_dp_netdev(dpif); - memset(stats, 0, sizeof *stats); + + fat_rwlock_rdlock(&dp->cls.rwlock); stats->n_flows = hmap_count(&dp->flow_table); - stats->cur_capacity = hmap_capacity(&dp->flow_table); - stats->max_capacity = MAX_FLOWS; - stats->n_ports = dp->n_ports; - stats->max_ports = MAX_PORTS; - stats->n_frags = dp->n_frags; - stats->n_hit = dp->n_hit; - stats->n_missed = dp->n_missed; - stats->n_lost = dp->n_lost; - stats->max_miss_queue = MAX_QUEUE_LEN; - stats->max_action_queue = MAX_QUEUE_LEN; - return 0; -} + fat_rwlock_unlock(&dp->cls.rwlock); -static int -dpif_netdev_get_drop_frags(const struct dpif *dpif, bool *drop_fragsp) -{ - struct dp_netdev *dp = get_dp_netdev(dpif); - *drop_fragsp = dp->drop_frags; - return 0; -} + 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_masks = UINT32_MAX; + stats->n_mask_hit = UINT64_MAX; -static int -dpif_netdev_set_drop_frags(struct dpif *dpif, bool drop_frags) -{ - struct dp_netdev *dp = get_dp_netdev(dpif); - dp->drop_frags = drop_frags; return 0; } static int -do_add_port(struct dp_netdev *dp, const char *devname, uint16_t flags, - uint16_t port_no) +do_add_port(struct dp_netdev *dp, const char *devname, const char *type, + odp_port_t port_no) + OVS_REQ_WRLOCK(dp->port_rwlock) { - bool internal = (flags & ODP_PORT_INTERNAL) != 0; + struct netdev_saved_flags *sf; struct dp_netdev_port *port; - struct netdev_options netdev_options; struct netdev *netdev; - int mtu; + struct netdev_rx *rx; + enum netdev_flags flags; + const char *open_type; int error; /* XXX reject devices already in some dp_netdev. */ /* Open and validate network device. */ - memset(&netdev_options, 0, sizeof netdev_options); - netdev_options.name = devname; - netdev_options.ethertype = NETDEV_ETH_TYPE_ANY; - if (internal) { - netdev_options.type = "tap"; - } - - error = netdev_open(&netdev_options, &netdev); + open_type = dpif_netdev_port_open_type(dp->class, type); + error = netdev_open(devname, open_type, &netdev); if (error) { return error; } - /* XXX reject loopback devices */ /* XXX reject non-Ethernet devices */ - error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, false); + netdev_get_flags(netdev, &flags); + if (flags & NETDEV_LOOPBACK) { + VLOG_ERR("%s: cannot add a loopback device", devname); + netdev_close(netdev); + 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; + } + + error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf); if (error) { + netdev_rx_close(rx); netdev_close(netdev); return error; } @@ -383,59 +670,88 @@ do_add_port(struct dp_netdev *dp, const char *devname, uint16_t flags, port = xmalloc(sizeof *port); port->port_no = port_no; port->netdev = netdev; - port->internal = internal; - - netdev_get_mtu(netdev, &mtu); - if (mtu > max_mtu) { - max_mtu = mtu; - } + port->sf = sf; + port->rx = rx; + port->type = xstrdup(type); - list_push_back(&dp->port_list, &port->node); - dp->ports[port_no] = port; - dp->n_ports++; - dp->serial++; + hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0)); + seq_change(dp->port_seq); return 0; } static int -dpif_netdev_port_add(struct dpif *dpif, const char *devname, uint16_t flags, - uint16_t *port_nop) +dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev, + odp_port_t *port_nop) { struct dp_netdev *dp = get_dp_netdev(dpif); - int port_no; + char namebuf[NETDEV_VPORT_NAME_BUFSIZE]; + const char *dpif_port; + odp_port_t port_no; + int error; - for (port_no = 0; port_no < MAX_PORTS; port_no++) { - if (!dp->ports[port_no]) { - *port_nop = port_no; - return do_add_port(dp, devname, flags, port_no); - } + ovs_rwlock_wrlock(&dp->port_rwlock); + dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf); + if (*port_nop != ODPP_NONE) { + port_no = *port_nop; + error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0; + } else { + port_no = choose_port(dp, dpif_port); + error = port_no == ODPP_NONE ? EFBIG : 0; + } + if (!error) { + *port_nop = port_no; + error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no); } - return EFBIG; + ovs_rwlock_unlock(&dp->port_rwlock); + + return error; } static int -dpif_netdev_port_del(struct dpif *dpif, uint16_t port_no) +dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no) { struct dp_netdev *dp = get_dp_netdev(dpif); - return port_no == ODPP_LOCAL ? EINVAL : do_del_port(dp, port_no); + int error; + + ovs_rwlock_wrlock(&dp->port_rwlock); + error = port_no == ODPP_LOCAL ? EINVAL : do_del_port(dp, port_no); + ovs_rwlock_unlock(&dp->port_rwlock); + + return error; } static bool -is_valid_port_number(uint16_t port_no) +is_valid_port_number(odp_port_t port_no) +{ + return port_no != ODPP_NONE; +} + +static struct dp_netdev_port * +dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no) + OVS_REQ_RDLOCK(dp->port_rwlock) { - return port_no < MAX_PORTS; + struct dp_netdev_port *port; + + HMAP_FOR_EACH_IN_BUCKET (port, node, hash_int(odp_to_u32(port_no), 0), + &dp->ports) { + if (port->port_no == port_no) { + return port; + } + } + return NULL; } static int get_port_by_number(struct dp_netdev *dp, - uint16_t port_no, struct dp_netdev_port **portp) + odp_port_t port_no, struct dp_netdev_port **portp) + OVS_REQ_RDLOCK(dp->port_rwlock) { if (!is_valid_port_number(port_no)) { *portp = NULL; return EINVAL; } else { - *portp = dp->ports[port_no]; + *portp = dp_netdev_lookup_port(dp, port_no); return *portp ? 0 : ENOENT; } } @@ -443,10 +759,11 @@ get_port_by_number(struct dp_netdev *dp, static int get_port_by_name(struct dp_netdev *dp, const char *devname, struct dp_netdev_port **portp) + OVS_REQ_RDLOCK(dp->port_rwlock) { struct dp_netdev_port *port; - LIST_FOR_EACH (port, node, &dp->port_list) { + HMAP_FOR_EACH (port, node, &dp->ports) { if (!strcmp(netdev_get_name(port->netdev), devname)) { *portp = port; return 0; @@ -456,10 +773,10 @@ get_port_by_name(struct dp_netdev *dp, } static int -do_del_port(struct dp_netdev *dp, uint16_t port_no) +do_del_port(struct dp_netdev *dp, odp_port_t port_no) + OVS_REQ_WRLOCK(dp->port_rwlock) { struct dp_netdev_port *port; - char *name; int error; error = get_port_by_number(dp, port_no, &port); @@ -467,431 +784,673 @@ do_del_port(struct dp_netdev *dp, uint16_t port_no) return error; } - list_remove(&port->node); - dp->ports[port->port_no] = NULL; - dp->n_ports--; - dp->serial++; + hmap_remove(&dp->ports, &port->node); + seq_change(dp->port_seq); - name = xstrdup(netdev_get_name(port->netdev)); netdev_close(port->netdev); - - free(name); + netdev_restore_flags(port->sf); + netdev_rx_close(port->rx); + free(port->type); free(port); return 0; } static void -answer_port_query(const struct dp_netdev_port *port, struct odp_port *odp_port) +answer_port_query(const struct dp_netdev_port *port, + struct dpif_port *dpif_port) { - memset(odp_port, 0, sizeof *odp_port); - ovs_strlcpy(odp_port->devname, netdev_get_name(port->netdev), - sizeof odp_port->devname); - odp_port->port = port->port_no; - odp_port->flags = port->internal ? ODP_PORT_INTERNAL : 0; + dpif_port->name = xstrdup(netdev_get_name(port->netdev)); + dpif_port->type = xstrdup(port->type); + dpif_port->port_no = port->port_no; } static int -dpif_netdev_port_query_by_number(const struct dpif *dpif, uint16_t port_no, - struct odp_port *odp_port) +dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no, + struct dpif_port *dpif_port) { struct dp_netdev *dp = get_dp_netdev(dpif); struct dp_netdev_port *port; int error; + ovs_rwlock_rdlock(&dp->port_rwlock); error = get_port_by_number(dp, port_no, &port); - if (!error) { - answer_port_query(port, odp_port); + if (!error && dpif_port) { + answer_port_query(port, dpif_port); } + ovs_rwlock_unlock(&dp->port_rwlock); + return error; } static int dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname, - struct odp_port *odp_port) + struct dpif_port *dpif_port) { struct dp_netdev *dp = get_dp_netdev(dpif); struct dp_netdev_port *port; int error; + ovs_rwlock_rdlock(&dp->port_rwlock); error = get_port_by_name(dp, devname, &port); - if (!error) { - answer_port_query(port, odp_port); + if (!error && dpif_port) { + answer_port_query(port, dpif_port); } + ovs_rwlock_unlock(&dp->port_rwlock); + return error; } static void -dp_netdev_free_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow) +dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow) + OVS_REQ_WRLOCK(dp->cls.rwlock) + OVS_REQUIRES(dp->flow_mutex) +{ + struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr); + struct hmap_node *node = CONST_CAST(struct hmap_node *, &flow->node); + + 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) { - hmap_remove(&dp->flow_table, &flow->node); - free(flow->actions); - free(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); + } } static void dp_netdev_flow_flush(struct dp_netdev *dp) { - struct dp_netdev_flow *flow, *next; + struct dp_netdev_flow *netdev_flow, *next; - HMAP_FOR_EACH_SAFE (flow, next, node, &dp->flow_table) { - dp_netdev_free_flow(dp, flow); + ovs_mutex_lock(&dp->flow_mutex); + fat_rwlock_wrlock(&dp->cls.rwlock); + HMAP_FOR_EACH_SAFE (netdev_flow, next, node, &dp->flow_table) { + dp_netdev_remove_flow(dp, netdev_flow); } + fat_rwlock_unlock(&dp->cls.rwlock); + ovs_mutex_unlock(&dp->flow_mutex); } static int dpif_netdev_flow_flush(struct dpif *dpif) { struct dp_netdev *dp = get_dp_netdev(dpif); + dp_netdev_flow_flush(dp); return 0; } +struct dp_netdev_port_state { + uint32_t bucket; + uint32_t offset; + char *name; +}; + +static int +dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep) +{ + *statep = xzalloc(sizeof(struct dp_netdev_port_state)); + return 0; +} + static int -dpif_netdev_port_list(const struct dpif *dpif, struct odp_port *ports, int n) +dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_, + struct dpif_port *dpif_port) { + struct dp_netdev_port_state *state = state_; struct dp_netdev *dp = get_dp_netdev(dpif); - struct dp_netdev_port *port; - int i; + struct hmap_node *node; + int retval; - i = 0; - LIST_FOR_EACH (port, node, &dp->port_list) { - struct odp_port *odp_port = &ports[i]; - if (i >= n) { - break; - } - answer_port_query(port, odp_port); - i++; + ovs_rwlock_rdlock(&dp->port_rwlock); + node = hmap_at_position(&dp->ports, &state->bucket, &state->offset); + if (node) { + struct dp_netdev_port *port; + + port = CONTAINER_OF(node, struct dp_netdev_port, node); + + free(state->name); + state->name = xstrdup(netdev_get_name(port->netdev)); + dpif_port->name = state->name; + dpif_port->type = port->type; + dpif_port->port_no = port->port_no; + + retval = 0; + } else { + retval = EOF; } - return dp->n_ports; + ovs_rwlock_unlock(&dp->port_rwlock); + + return retval; +} + +static int +dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_) +{ + struct dp_netdev_port_state *state = state_; + free(state->name); + free(state); + return 0; } static int dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED) { struct dpif_netdev *dpif = dpif_netdev_cast(dpif_); - if (dpif->dp_serial != dpif->dp->serial) { - dpif->dp_serial = dpif->dp->serial; - return ENOBUFS; + uint64_t new_port_seq; + int error; + + new_port_seq = seq_read(dpif->dp->port_seq); + if (dpif->last_port_seq != new_port_seq) { + dpif->last_port_seq = new_port_seq; + error = ENOBUFS; } else { - return EAGAIN; + error = EAGAIN; } + + return error; } static void dpif_netdev_port_poll_wait(const struct dpif *dpif_) { struct dpif_netdev *dpif = dpif_netdev_cast(dpif_); - if (dpif->dp_serial != dpif->dp->serial) { - poll_immediate_wake(); - } + + seq_wait(dpif->dp->port_seq, dpif->last_port_seq); } static struct dp_netdev_flow * -dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct flow *key) +dp_netdev_flow_cast(const struct cls_rule *cr) { - struct dp_netdev_flow *flow; - - HMAP_FOR_EACH_WITH_HASH (flow, node, flow_hash(key, 0), &dp->flow_table) { - if (flow_equal(&flow->key, key)) { - return flow; - } - } - return NULL; + return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL; } -/* The caller must fill in odp_flow->key itself. */ -static void -answer_flow_query(struct dp_netdev_flow *flow, uint32_t query_flags, - struct odp_flow *odp_flow) +static struct dp_netdev_flow * +dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct flow *flow) + OVS_EXCLUDED(dp->cls.rwlock) { - if (flow) { - odp_flow->stats.n_packets = flow->packet_count; - odp_flow->stats.n_bytes = flow->byte_count; - odp_flow->stats.used_sec = flow->used.tv_sec; - odp_flow->stats.used_nsec = flow->used.tv_nsec; - odp_flow->stats.tcp_flags = TCP_FLAGS(flow->tcp_ctl); - odp_flow->stats.reserved = 0; - odp_flow->stats.error = 0; - if (odp_flow->n_actions > 0) { - unsigned int n = MIN(odp_flow->n_actions, flow->n_actions); - memcpy(odp_flow->actions, flow->actions, - n * sizeof *odp_flow->actions); - odp_flow->n_actions = flow->n_actions; - } + struct dp_netdev_flow *netdev_flow; - if (query_flags & ODPFF_ZERO_TCP_FLAGS) { - flow->tcp_ctl = 0; - } + fat_rwlock_rdlock(&dp->cls.rwlock); + netdev_flow = dp_netdev_flow_cast(classifier_lookup(&dp->cls, flow, NULL)); + dp_netdev_flow_ref(netdev_flow); + fat_rwlock_unlock(&dp->cls.rwlock); - } else { - odp_flow->stats.error = ENOENT; - } + return netdev_flow; } -static int -dpif_netdev_flow_get(const struct dpif *dpif, struct odp_flow flows[], int n) +static struct dp_netdev_flow * +dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow) + OVS_REQ_RDLOCK(dp->cls.rwlock) { - struct dp_netdev *dp = get_dp_netdev(dpif); - int i; - - for (i = 0; i < n; i++) { - struct odp_flow *odp_flow = &flows[i]; - struct flow key; + struct dp_netdev_flow *netdev_flow; - odp_flow_key_to_flow(&odp_flow->key, &key); - answer_flow_query(dp_netdev_lookup_flow(dp, &key), - odp_flow->flags, odp_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 0; -} -static int -dpif_netdev_validate_actions(const union odp_action *actions, int n_actions, - bool *mutates) -{ - unsigned int i; - - *mutates = false; - for (i = 0; i < n_actions; i++) { - const union odp_action *a = &actions[i]; - switch (a->type) { - case ODPAT_OUTPUT: - if (a->output.port >= MAX_PORTS) { - return EINVAL; - } - break; + return NULL; +} - case ODPAT_CONTROLLER: - break; +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; +} - case ODPAT_SET_DL_TCI: - *mutates = true; - if (a->dl_tci.tci & htons(VLAN_CFI)) { - return EINVAL; +static int +dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len, + const struct nlattr *mask_key, + uint32_t mask_key_len, const struct flow *flow, + struct flow *mask) +{ + if (mask_key_len) { + 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 + * as an error, with enough details to enable debugging. */ + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); + + if (!VLOG_DROP_ERR(&rl)) { + struct ds s; + + 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 (%s)", + ds_cstr(&s), odp_key_fitness_to_string(fitness)); + ds_destroy(&s); } - break; - case ODPAT_SET_NW_TOS: - *mutates = true; - if (a->nw_tos.nw_tos & IP_ECN_MASK) { - return EINVAL; + 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 + * prerequisities are not met. */ + memset(mask, 0x0, sizeof *mask); + + for (id = 0; id < MFF_N_IDS; ++id) { + /* Skip registers and metadata. */ + if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS) + && id != MFF_METADATA) { + const struct mf_field *mf = mf_from_id(id); + if (mf_are_prereqs_ok(mf, flow)) { + mf_mask_field(mf, mask); + } } - break; - - case ODPAT_STRIP_VLAN: - case ODPAT_SET_DL_SRC: - case ODPAT_SET_DL_DST: - case ODPAT_SET_NW_SRC: - case ODPAT_SET_NW_DST: - case ODPAT_SET_TP_SRC: - case ODPAT_SET_TP_DST: - *mutates = true; - break; - - default: - return EOPNOTSUPP; } } + return 0; } static int -set_flow_actions(struct dp_netdev_flow *flow, struct odp_flow *odp_flow) +dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len, + struct flow *flow) { - size_t n_bytes; - bool mutates; - int error; + odp_port_t in_port; + + if (odp_flow_key_to_flow(key, key_len, flow)) { + /* This should not happen: it indicates that odp_flow_key_from_flow() + * and odp_flow_key_to_flow() disagree on the acceptable form of a + * flow. Log the problem as an error, with enough details to enable + * debugging. */ + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); + + if (!VLOG_DROP_ERR(&rl)) { + struct ds s; + + ds_init(&s); + odp_flow_format(key, key_len, NULL, 0, NULL, &s, true); + VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s)); + ds_destroy(&s); + } - if (odp_flow->n_actions >= 4096 / sizeof *odp_flow->actions) { return EINVAL; } - error = dpif_netdev_validate_actions(odp_flow->actions, - odp_flow->n_actions, &mutates); - if (error) { - return error; + + in_port = flow->in_port.odp_port; + if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) { + return EINVAL; } - n_bytes = odp_flow->n_actions * sizeof *flow->actions; - flow->actions = xrealloc(flow->actions, n_bytes); - flow->n_actions = odp_flow->n_actions; - memcpy(flow->actions, odp_flow->actions, n_bytes); return 0; } static int -add_flow(struct dpif *dpif, struct odp_flow *odp_flow) +dpif_netdev_flow_get(const struct dpif *dpif, + const struct nlattr *nl_key, size_t nl_key_len, + struct ofpbuf **actionsp, struct dpif_flow_stats *stats) { struct dp_netdev *dp = get_dp_netdev(dpif); - struct dp_netdev_flow *flow; + struct dp_netdev_flow *netdev_flow; + struct flow key; int error; - flow = xzalloc(sizeof *flow); - odp_flow_key_to_flow(&odp_flow->key, &flow->key); - - error = set_flow_actions(flow, odp_flow); + error = dpif_netdev_flow_from_nlattrs(nl_key, nl_key_len, &key); if (error) { - free(flow); return error; } - hmap_insert(&dp->flow_table, &flow->node, flow_hash(&flow->key, 0)); + fat_rwlock_rdlock(&dp->cls.rwlock); + netdev_flow = dp_netdev_find_flow(dp, &key); + 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) { + *actionsp = ofpbuf_clone_data(actions->actions, actions->size); + dp_netdev_actions_unref(actions); + } + } else { + error = ENOENT; + } + + return error; +} + +static int +dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow, + const struct flow_wildcards *wc, + const struct nlattr *actions, + size_t actions_len) + OVS_REQUIRES(dp->flow_mutex) +{ + struct dp_netdev_flow *netdev_flow; + struct match match; + + 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); + + match_init(&match, flow, wc); + cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr), + &match, NETDEV_RULE_PRIORITY); + 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)); + fat_rwlock_unlock(&dp->cls.rwlock); + + ovs_mutex_unlock(&netdev_flow->mutex); + return 0; } static void -clear_stats(struct dp_netdev_flow *flow) +clear_stats(struct dp_netdev_flow *netdev_flow) + OVS_REQUIRES(netdev_flow->mutex) { - flow->used.tv_sec = 0; - flow->used.tv_nsec = 0; - flow->packet_count = 0; - flow->byte_count = 0; - flow->tcp_ctl = 0; + netdev_flow->used = 0; + netdev_flow->packet_count = 0; + netdev_flow->byte_count = 0; + netdev_flow->tcp_flags = 0; } static int -dpif_netdev_flow_put(struct dpif *dpif, struct odp_flow_put *put) +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 *flow; - struct flow key; + struct dp_netdev_flow *netdev_flow; + struct flow flow; + struct flow_wildcards wc; + int error; - odp_flow_key_to_flow(&put->flow.key, &key); - flow = dp_netdev_lookup_flow(dp, &key); - if (!flow) { - if (put->flags & ODPPF_CREATE) { + error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &flow); + if (error) { + return error; + } + error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len, + put->mask, put->mask_len, + &flow, &wc.masks); + if (error) { + return error; + } + + ovs_mutex_lock(&dp->flow_mutex); + netdev_flow = dp_netdev_lookup_flow(dp, &flow); + if (!netdev_flow) { + if (put->flags & DPIF_FP_CREATE) { if (hmap_count(&dp->flow_table) < MAX_FLOWS) { - return add_flow(dpif, &put->flow); + if (put->stats) { + memset(put->stats, 0, sizeof *put->stats); + } + error = dp_netdev_flow_add(dp, &flow, &wc, put->actions, + put->actions_len); } else { - return EFBIG; + error = EFBIG; } } else { - return ENOENT; + error = ENOENT; } } else { - if (put->flags & ODPPF_MODIFY) { - int error = set_flow_actions(flow, &put->flow); - if (!error && put->flags & ODPPF_ZERO_STATS) { - clear_stats(flow); + if (put->flags & DPIF_FP_MODIFY + && flow_equal(&flow, &netdev_flow->flow)) { + struct dp_netdev_actions *new_actions; + struct dp_netdev_actions *old_actions; + + 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; + if (put->stats) { + get_dpif_flow_stats(netdev_flow, put->stats); } - return error; + if (put->flags & DPIF_FP_ZERO_STATS) { + clear_stats(netdev_flow); + } + ovs_mutex_unlock(&netdev_flow->mutex); + + dp_netdev_actions_unref(old_actions); + } else if (put->flags & DPIF_FP_CREATE) { + error = EEXIST; } else { - return EEXIST; + /* Overlapping flow. */ + error = EINVAL; } + dp_netdev_flow_unref(netdev_flow); } -} + ovs_mutex_unlock(&dp->flow_mutex); + return error; +} static int -dpif_netdev_flow_del(struct dpif *dpif, struct odp_flow *odp_flow) +dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del) { struct dp_netdev *dp = get_dp_netdev(dpif); - struct dp_netdev_flow *flow; + struct dp_netdev_flow *netdev_flow; struct flow key; + int error; - odp_flow_key_to_flow(&odp_flow->key, &key); - flow = dp_netdev_lookup_flow(dp, &key); - if (flow) { - answer_flow_query(flow, 0, odp_flow); - dp_netdev_free_flow(dp, flow); - return 0; + error = dpif_netdev_flow_from_nlattrs(del->key, del->key_len, &key); + if (error) { + return error; + } + + ovs_mutex_lock(&dp->flow_mutex); + 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 { - return ENOENT; + error = ENOENT; } + 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_state *state; + + *statep = state = xmalloc(sizeof *state); + state->bucket = 0; + state->offset = 0; + state->actions = NULL; + return 0; } static int -dpif_netdev_flow_list(const struct dpif *dpif, struct odp_flow flows[], int n) +dpif_netdev_flow_dump_next(const struct dpif *dpif, 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_state *state = state_; struct dp_netdev *dp = get_dp_netdev(dpif); - struct dp_netdev_flow *flow; - int i; + struct dp_netdev_flow *netdev_flow; + struct hmap_node *node; + + fat_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); + } + fat_rwlock_unlock(&dp->cls.rwlock); + if (!node) { + return EOF; + } - i = 0; - HMAP_FOR_EACH (flow, node, &dp->flow_table) { - if (i >= n) { - break; - } + if (key) { + struct ofpbuf buf; + + ofpbuf_use_stack(&buf, &state->keybuf, sizeof state->keybuf); + odp_flow_key_from_flow(&buf, &netdev_flow->flow, + netdev_flow->flow.in_port.odp_port); + + *key = buf.data; + *key_len = buf.size; + } + + 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), + SIZE_MAX); - odp_flow_key_from_flow(&flows[i].key, &flow->key); - answer_flow_query(flow, 0, &flows[i]); - i++; + *mask = buf.data; + *mask_len = buf.size; } - return hmap_count(&dp->flow_table); + + 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); + *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_execute(struct dpif *dpif, - const union odp_action actions[], int n_actions, - const struct ofpbuf *packet) +dpif_netdev_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_) +{ + struct dp_netdev_flow_state *state = state_; + + dp_netdev_actions_unref(state->actions); + free(state); + return 0; +} + +static int +dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute) { struct dp_netdev *dp = get_dp_netdev(dpif); - struct ofpbuf copy; - bool mutates; + struct pkt_metadata *md = &execute->md; struct flow key; - int error; - if (packet->size < ETH_HEADER_LEN || packet->size > UINT16_MAX) { + if (execute->packet->size < ETH_HEADER_LEN || + execute->packet->size > UINT16_MAX) { return EINVAL; } - error = dpif_netdev_validate_actions(actions, n_actions, &mutates); - if (error) { - return error; - } + /* Extract flow key. */ + flow_extract(execute->packet, md->skb_priority, md->pkt_mark, &md->tunnel, + (union flow_in_port *)&md->in_port, &key); - if (mutates) { - /* We need a deep copy of 'packet' since we're going to modify its - * data. */ - ofpbuf_init(©, DP_NETDEV_HEADROOM + packet->size); - copy.data = (char*)copy.base + DP_NETDEV_HEADROOM; - ofpbuf_put(©, packet->data, packet->size); - } else { - /* We still need a shallow copy of 'packet', even though we won't - * modify its data, because flow_extract() modifies packet->l2, etc. - * We could probably get away with modifying those but it's more polite - * if we don't. */ - copy = *packet; - } - flow_extract(©, 0, -1, &key); - error = dp_netdev_execute_actions(dp, ©, &key, actions, n_actions); - if (mutates) { - ofpbuf_uninit(©); - } - return error; + ovs_rwlock_rdlock(&dp->port_rwlock); + dp_netdev_execute_actions(dp, &key, execute->packet, md, execute->actions, + execute->actions_len); + ovs_rwlock_unlock(&dp->port_rwlock); + + return 0; } static int -dpif_netdev_recv_get_mask(const struct dpif *dpif, int *listen_mask) +dpif_netdev_recv_set(struct dpif *dpif OVS_UNUSED, bool enable OVS_UNUSED) { - struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif); - *listen_mask = dpif_netdev->listen_mask; return 0; } static int -dpif_netdev_recv_set_mask(struct dpif *dpif, int listen_mask) +dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED, + uint32_t queue_id, uint32_t *priority) { - struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif); - if (!(listen_mask & ~ODPL_ALL)) { - dpif_netdev->listen_mask = listen_mask; - return 0; - } else { - return EINVAL; - } + *priority = queue_id; + return 0; } -static struct ovs_queue * -find_nonempty_queue(struct dpif *dpif) +static struct dp_netdev_queue * +find_nonempty_queue(struct dp_netdev *dp) + OVS_REQUIRES(dp->queue_mutex) { - struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif); - struct dp_netdev *dp = get_dp_netdev(dpif); - int mask = dpif_netdev->listen_mask; int i; for (i = 0; i < N_QUEUES; i++) { - struct ovs_queue *q = &dp->queues[i]; - if (q->n && mask & (1u << i)) { + struct dp_netdev_queue *q = &dp->queues[i]; + if (q->head != q->tail) { return q; } } @@ -899,400 +1458,485 @@ find_nonempty_queue(struct dpif *dpif) } static int -dpif_netdev_recv(struct dpif *dpif, struct ofpbuf **bufp) +dpif_netdev_recv(struct dpif *dpif, struct dpif_upcall *upcall, + struct ofpbuf *buf) { - struct ovs_queue *q = find_nonempty_queue(dpif); + struct dp_netdev *dp = get_dp_netdev(dpif); + struct dp_netdev_queue *q; + int error; + + ovs_mutex_lock(&dp->queue_mutex); + q = find_nonempty_queue(dp); if (q) { - *bufp = queue_pop_head(q); - return 0; + struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK]; + + *upcall = u->upcall; + + ofpbuf_uninit(buf); + *buf = u->buf; + + error = 0; } else { - return EAGAIN; + error = EAGAIN; } + ovs_mutex_unlock(&dp->queue_mutex); + + return error; } static void dpif_netdev_recv_wait(struct dpif *dpif) { - struct ovs_queue *q = find_nonempty_queue(dpif); - if (q) { + struct dp_netdev *dp = get_dp_netdev(dpif); + uint64_t seq; + + ovs_mutex_lock(&dp->queue_mutex); + seq = seq_read(dp->queue_seq); + if (find_nonempty_queue(dp)) { poll_immediate_wake(); } else { - /* No messages ready to be received, and dp_wait() will ensure that we - * wake up to queue new messages, so there is nothing to do. */ + seq_wait(dp->queue_seq, seq); } + ovs_mutex_unlock(&dp->queue_mutex); } - + static void -dp_netdev_flow_used(struct dp_netdev_flow *flow, struct flow *key, - const struct ofpbuf *packet) +dpif_netdev_recv_purge(struct dpif *dpif) { - time_timespec(&flow->used); - flow->packet_count++; - flow->byte_count += packet->size; - if (key->dl_type == htons(ETH_TYPE_IP) && key->nw_proto == IPPROTO_TCP) { - struct tcp_header *th = packet->l4; - flow->tcp_ctl |= th->tcp_ctl; - } + struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif); + + dp_netdev_purge_queues(dpif_netdev->dp); } + +/* Creates and returns a new 'struct dp_netdev_actions', with a reference count + * of 1, whose actions are a copy of from the 'ofpacts_len' bytes of + * 'ofpacts'. */ +struct dp_netdev_actions * +dp_netdev_actions_create(const struct nlattr *actions, size_t size) +{ + struct dp_netdev_actions *netdev_actions; -static void -dp_netdev_port_input(struct dp_netdev *dp, struct dp_netdev_port *port, - struct ofpbuf *packet) + 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_) { - struct dp_netdev_flow *flow; - struct flow key; + struct dp_netdev_actions *actions; - if (packet->size < ETH_HEADER_LEN) { - return; - } - if (flow_extract(packet, 0, port->port_no, &key) && dp->drop_frags) { - dp->n_frags++; - return; + actions = CONST_CAST(struct dp_netdev_actions *, actions_); + if (actions) { + ovs_refcount_ref(&actions->ref_cnt); } + return actions; +} - flow = dp_netdev_lookup_flow(dp, &key); - if (flow) { - dp_netdev_flow_used(flow, &key, packet); - dp_netdev_execute_actions(dp, packet, &key, - flow->actions, flow->n_actions); - dp->n_hit++; - } else { - dp->n_missed++; - dp_netdev_output_control(dp, packet, _ODPL_MISS_NR, port->port_no, 0); +/* Decrements 'actions''s refcount and frees 'actions' if the refcount reaches + * 0. */ +void +dp_netdev_actions_unref(struct dp_netdev_actions *actions) +{ + if (actions && ovs_refcount_unref(&actions->ref_cnt) == 1) { + free(actions->actions); + free(actions); } } - -static void -dp_netdev_run(void) + +static void * +dp_forwarder_main(void *f_) { + struct dp_forwarder *f = f_; + struct dp_netdev *dp = f->dp; struct ofpbuf packet; - struct dp_netdev *dp; - ofpbuf_init(&packet, DP_NETDEV_HEADROOM + max_mtu); - LIST_FOR_EACH (dp, node, &dp_netdev_list) { - struct dp_netdev_port *port; + f->name = xasprintf("forwarder_%u", ovsthread_id_self()); + set_subprogram_name("%s", f->name); - LIST_FOR_EACH (port, node, &dp->port_list) { - int error; + ofpbuf_init(&packet, 0); + while (!latch_is_set(&dp->exit_latch)) { + bool received_anything; + int i; - /* Reset packet contents. */ - packet.data = (char*)packet.base + DP_NETDEV_HEADROOM; - packet.size = 0; + ovs_rwlock_rdlock(&dp->port_rwlock); + for (i = 0; i < 50; i++) { + struct dp_netdev_port *port; - error = netdev_recv(port->netdev, &packet); - if (!error) { - dp_netdev_port_input(dp, port, &packet); - } else if (error != EAGAIN) { - 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), strerror(error)); + 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)); + } + } + } + + if (!received_anything) { + break; } } - } - ofpbuf_uninit(&packet); -} -static void -dp_netdev_wait(void) -{ - struct dp_netdev *dp; + if (received_anything) { + poll_immediate_wake(); + } else { + struct dp_netdev_port *port; - LIST_FOR_EACH (dp, node, &dp_netdev_list) { - struct dp_netdev_port *port; - LIST_FOR_EACH (port, node, &dp->port_list) { - netdev_recv_wait(port->netdev); + 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); } + ovs_rwlock_unlock(&dp->port_rwlock); + + poll_block(); } -} + ofpbuf_uninit(&packet); + free(f->name); + + return NULL; +} -/* Modify the TCI field of 'packet'. If a VLAN tag is present, its TCI field - * is replaced by 'tci'. If a VLAN tag is not present, one is added with the - * TCI field set to 'tci'. - */ static void -dp_netdev_set_dl_tci(struct ofpbuf *packet, uint16_t tci) +dp_netdev_set_threads(struct dp_netdev *dp, int n) { - struct vlan_eth_header *veh; - struct eth_header *eh; - - eh = packet->l2; - if (packet->size >= sizeof(struct vlan_eth_header) - && eh->eth_type == htons(ETH_TYPE_VLAN)) { - veh = packet->l2; - veh->veth_tci = tci; - } else { - /* Insert new 802.1Q header. */ - struct vlan_eth_header tmp; - memcpy(tmp.veth_dst, eh->eth_dst, ETH_ADDR_LEN); - memcpy(tmp.veth_src, eh->eth_src, ETH_ADDR_LEN); - tmp.veth_type = htons(ETH_TYPE_VLAN); - tmp.veth_tci = tci; - tmp.veth_next_type = eh->eth_type; + int i; - veh = ofpbuf_push_uninit(packet, VLAN_HEADER_LEN); - memcpy(veh, &tmp, sizeof tmp); - packet->l2 = (char*)packet->l2 - VLAN_HEADER_LEN; + if (n == dp->n_forwarders) { + return; } -} -static void -dp_netdev_strip_vlan(struct ofpbuf *packet) -{ - struct vlan_eth_header *veh = packet->l2; - if (packet->size >= sizeof *veh - && veh->veth_type == htons(ETH_TYPE_VLAN)) { - struct eth_header tmp; + /* Stop existing threads. */ + latch_set(&dp->exit_latch); + for (i = 0; i < dp->n_forwarders; i++) { + struct dp_forwarder *f = &dp->forwarders[i]; + + xpthread_join(f->thread, NULL); + } + latch_poll(&dp->exit_latch); + free(dp->forwarders); - memcpy(tmp.eth_dst, veh->veth_dst, ETH_ADDR_LEN); - memcpy(tmp.eth_src, veh->veth_src, ETH_ADDR_LEN); - tmp.eth_type = veh->veth_next_type; + /* Start new threads. */ + dp->forwarders = xmalloc(n * sizeof *dp->forwarders); + dp->n_forwarders = n; + for (i = 0; i < n; i++) { + struct dp_forwarder *f = &dp->forwarders[i]; - packet->size -= VLAN_HEADER_LEN; - packet->data = (char*)packet->data + VLAN_HEADER_LEN; - packet->l2 = (char*)packet->l2 + VLAN_HEADER_LEN; - memcpy(packet->data, &tmp, sizeof tmp); + 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); } } - + static void -dp_netdev_set_dl_src(struct ofpbuf *packet, const uint8_t dl_addr[ETH_ADDR_LEN]) +dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, + const struct ofpbuf *packet) + OVS_REQUIRES(netdev_flow->mutex) { - struct eth_header *eh = packet->l2; - memcpy(eh->eth_src, dl_addr, sizeof eh->eth_src); + 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); } static void -dp_netdev_set_dl_dst(struct ofpbuf *packet, const uint8_t dl_addr[ETH_ADDR_LEN]) -{ - struct eth_header *eh = packet->l2; - memcpy(eh->eth_dst, dl_addr, sizeof eh->eth_dst); -} - -static bool -is_ip(const struct ofpbuf *packet, const struct flow *key) +dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet, + struct pkt_metadata *md) + OVS_REQ_RDLOCK(dp->port_rwlock) { - return key->dl_type == htons(ETH_TYPE_IP) && packet->l4; -} + struct dp_netdev_flow *netdev_flow; + struct flow key; -static void -dp_netdev_set_nw_addr(struct ofpbuf *packet, struct flow *key, - const struct odp_action_nw_addr *a) -{ - if (is_ip(packet, key)) { - struct ip_header *nh = packet->l3; - uint32_t *field; - - field = a->type == ODPAT_SET_NW_SRC ? &nh->ip_src : &nh->ip_dst; - if (key->nw_proto == IP_TYPE_TCP && packet->l7) { - struct tcp_header *th = packet->l4; - th->tcp_csum = recalc_csum32(th->tcp_csum, *field, a->nw_addr); - } else if (key->nw_proto == IP_TYPE_UDP && packet->l7) { - struct udp_header *uh = packet->l4; - if (uh->udp_csum) { - uh->udp_csum = recalc_csum32(uh->udp_csum, *field, a->nw_addr); - if (!uh->udp_csum) { - uh->udp_csum = 0xffff; - } - } - } - nh->ip_csum = recalc_csum32(nh->ip_csum, *field, a->nw_addr); - *field = a->nw_addr; + if (packet->size < ETH_HEADER_LEN) { + 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); + 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_execute_actions(dp, &key, packet, 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); } } -static void -dp_netdev_set_nw_tos(struct ofpbuf *packet, struct flow *key, - const struct odp_action_nw_tos *a) +static int +dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet, + int queue_no, const struct flow *flow, + const struct nlattr *userdata) + OVS_EXCLUDED(dp->queue_mutex) { - if (is_ip(packet, key)) { - struct ip_header *nh = packet->l3; - uint8_t *field = &nh->ip_tos; + struct dp_netdev_queue *q = &dp->queues[queue_no]; + int error; - /* Set the DSCP bits and preserve the ECN bits. */ - uint8_t new = a->nw_tos | (nh->ip_tos & IP_ECN_MASK); + ovs_mutex_lock(&dp->queue_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; - nh->ip_csum = recalc_csum16(nh->ip_csum, htons((uint16_t)*field), - htons((uint16_t)a->nw_tos)); - *field = new; - } -} + upcall->type = queue_no; -static void -dp_netdev_set_tp_port(struct ofpbuf *packet, struct flow *key, - const struct odp_action_tp_port *a) -{ - if (is_ip(packet, key)) { - uint16_t *field; - if (key->nw_proto == IPPROTO_TCP && packet->l7) { - struct tcp_header *th = packet->l4; - field = a->type == ODPAT_SET_TP_SRC ? &th->tcp_src : &th->tcp_dst; - th->tcp_csum = recalc_csum16(th->tcp_csum, *field, a->tp_port); - *field = a->tp_port; - } else if (key->nw_proto == IPPROTO_UDP && packet->l7) { - struct udp_header *uh = packet->l4; - field = a->type == ODPAT_SET_TP_SRC ? &uh->udp_src : &uh->udp_dst; - uh->udp_csum = recalc_csum16(uh->udp_csum, *field, a->tp_port); - *field = a->tp_port; - } else { - return; + /* Allocate buffer big enough for everything. */ + buf_size = ODPUTIL_FLOW_KEY_BYTES; + if (userdata) { + buf_size += NLA_ALIGN(userdata->nla_len); } - } -} + ofpbuf_init(buf, buf_size); -static void -dp_netdev_output_port(struct dp_netdev *dp, struct ofpbuf *packet, - uint16_t out_port) -{ - struct dp_netdev_port *p = dp->ports[out_port]; - if (p) { - netdev_send(p->netdev, packet); - } -} + /* Put ODP flow. */ + odp_flow_key_from_flow(buf, flow, flow->in_port.odp_port); + upcall->key = buf->data; + upcall->key_len = buf->size; -static int -dp_netdev_output_control(struct dp_netdev *dp, const struct ofpbuf *packet, - int queue_no, int port_no, uint32_t arg) -{ - struct ovs_queue *q = &dp->queues[queue_no]; - struct odp_msg *header; - struct ofpbuf *msg; - size_t msg_size; + /* Put userdata. */ + if (userdata) { + upcall->userdata = ofpbuf_put(buf, userdata, + NLA_ALIGN(userdata->nla_len)); + } - if (q->n >= MAX_QUEUE_LEN) { - dp->n_lost++; - return ENOBUFS; - } + /* Steal packet data. */ + ovs_assert(packet->source == OFPBUF_MALLOC); + upcall->packet = *packet; + ofpbuf_use(packet, NULL, 0); - msg_size = sizeof *header + packet->size; - msg = ofpbuf_new_with_headroom(msg_size, DPIF_RECV_MSG_PADDING); - header = ofpbuf_put_uninit(msg, sizeof *header); - header->type = queue_no; - header->length = msg_size; - header->port = port_no; - header->arg = arg; - ofpbuf_put(msg, packet->data, packet->size); - queue_push_tail(q, msg); + seq_change(dp->queue_seq); - return 0; + error = 0; + } else { + ovsthread_counter_inc(dp->n_lost, 1); + error = ENOBUFS; + } + ovs_mutex_unlock(&dp->queue_mutex); + + return error; } -/* Returns true if 'packet' is an invalid Ethernet+IPv4 ARP packet: one with - * screwy or truncated header fields or one whose inner and outer Ethernet - * address differ. */ -static bool -dp_netdev_is_spoofed_arp(struct ofpbuf *packet, const struct flow *key) +struct dp_netdev_execute_aux { + struct dp_netdev *dp; + const struct flow *key; +}; + +static void +dp_execute_cb(void *aux_, struct ofpbuf *packet, + const struct pkt_metadata *md OVS_UNUSED, + const struct nlattr *a, bool may_steal) + OVS_NO_THREAD_SAFETY_ANALYSIS { - struct arp_eth_header *arp; - struct eth_header *eth; - ptrdiff_t l3_size; + struct dp_netdev_execute_aux *aux = aux_; + int type = nl_attr_type(a); + struct dp_netdev_port *p; + + 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); + } + break; - if (key->dl_type != htons(ETH_TYPE_ARP)) { - return false; - } + case OVS_ACTION_ATTR_USERSPACE: { + const struct nlattr *userdata; - l3_size = (char *) ofpbuf_end(packet) - (char *) packet->l3; - if (l3_size < sizeof(struct arp_eth_header)) { - return true; - } + userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA); - eth = packet->l2; - arp = packet->l3; - return (arp->ar_hrd != htons(ARP_HRD_ETHERNET) - || arp->ar_pro != htons(ARP_PRO_IP) - || arp->ar_hln != ETH_HEADER_LEN - || arp->ar_pln != 4 - || !eth_addr_equals(arp->ar_sha, eth->eth_src)); + /* 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, + userdata); + if (!may_steal) { + ofpbuf_uninit(packet); + } + break; + } + case OVS_ACTION_ATTR_PUSH_VLAN: + case OVS_ACTION_ATTR_POP_VLAN: + case OVS_ACTION_ATTR_PUSH_MPLS: + case OVS_ACTION_ATTR_POP_MPLS: + case OVS_ACTION_ATTR_SET: + case OVS_ACTION_ATTR_SAMPLE: + case OVS_ACTION_ATTR_UNSPEC: + case __OVS_ACTION_ATTR_MAX: + OVS_NOT_REACHED(); + } } -static int -dp_netdev_execute_actions(struct dp_netdev *dp, - struct ofpbuf *packet, struct flow *key, - const union odp_action *actions, int n_actions) +static void +dp_netdev_execute_actions(struct dp_netdev *dp, const struct flow *key, + struct ofpbuf *packet, struct pkt_metadata *md, + const struct nlattr *actions, size_t actions_len) + OVS_REQ_RDLOCK(dp->port_rwlock) { - int i; - for (i = 0; i < n_actions; i++) { - const union odp_action *a = &actions[i]; - - switch (a->type) { - case ODPAT_OUTPUT: - dp_netdev_output_port(dp, packet, a->output.port); - break; - - case ODPAT_CONTROLLER: - dp_netdev_output_control(dp, packet, _ODPL_ACTION_NR, - key->in_port, a->controller.arg); - break; - - case ODPAT_SET_DL_TCI: - dp_netdev_set_dl_tci(packet, a->dl_tci.tci); - break; - - case ODPAT_STRIP_VLAN: - dp_netdev_strip_vlan(packet); - break; - - case ODPAT_SET_DL_SRC: - dp_netdev_set_dl_src(packet, a->dl_addr.dl_addr); - break; - - case ODPAT_SET_DL_DST: - dp_netdev_set_dl_dst(packet, a->dl_addr.dl_addr); - break; - - case ODPAT_SET_NW_SRC: - case ODPAT_SET_NW_DST: - dp_netdev_set_nw_addr(packet, key, &a->nw_addr); - break; - - case ODPAT_SET_NW_TOS: - dp_netdev_set_nw_tos(packet, key, &a->nw_tos); - break; - - case ODPAT_SET_TP_SRC: - case ODPAT_SET_TP_DST: - dp_netdev_set_tp_port(packet, key, &a->tp_port); - break; - - case ODPAT_DROP_SPOOFED_ARP: - if (dp_netdev_is_spoofed_arp(packet, key)) { - return 0; - } - } - } - return 0; + struct dp_netdev_execute_aux aux = {dp, key}; + + odp_execute_actions(&aux, packet, md, actions, actions_len, dp_execute_cb); } const struct dpif_class dpif_netdev_class = { "netdev", - dp_netdev_run, - dp_netdev_wait, - NULL, /* enumerate */ + dpif_netdev_enumerate, + dpif_netdev_port_open_type, dpif_netdev_open, dpif_netdev_close, - NULL, /* get_all_names */ dpif_netdev_destroy, + NULL, /* run */ + NULL, /* wait */ dpif_netdev_get_stats, - dpif_netdev_get_drop_frags, - dpif_netdev_set_drop_frags, dpif_netdev_port_add, dpif_netdev_port_del, dpif_netdev_port_query_by_number, dpif_netdev_port_query_by_name, - dpif_netdev_port_list, + NULL, /* port_get_pid */ + dpif_netdev_port_dump_start, + dpif_netdev_port_dump_next, + dpif_netdev_port_dump_done, dpif_netdev_port_poll, dpif_netdev_port_poll_wait, dpif_netdev_flow_get, dpif_netdev_flow_put, dpif_netdev_flow_del, dpif_netdev_flow_flush, - dpif_netdev_flow_list, + dpif_netdev_flow_dump_start, + dpif_netdev_flow_dump_next, + dpif_netdev_flow_dump_done, dpif_netdev_execute, - dpif_netdev_recv_get_mask, - dpif_netdev_recv_set_mask, - NULL, /* get_sflow_probability */ - NULL, /* set_sflow_probability */ - NULL, /* queue_to_priority */ + NULL, /* operate */ + dpif_netdev_recv_set, + dpif_netdev_queue_to_priority, dpif_netdev_recv, dpif_netdev_recv_wait, + dpif_netdev_recv_purge, }; + +static void +dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED, + const char *argv[], void *aux OVS_UNUSED) +{ + struct dp_netdev_port *port; + struct dp_netdev *dp; + odp_port_t port_no; + + ovs_mutex_lock(&dp_netdev_mutex); + dp = shash_find_data(&dp_netdevs, argv[1]); + if (!dp || !dpif_netdev_class_is_dummy(dp->class)) { + ovs_mutex_unlock(&dp_netdev_mutex); + unixctl_command_reply_error(conn, "unknown datapath or not a dummy"); + return; + } + ovs_refcount_ref(&dp->ref_cnt); + ovs_mutex_unlock(&dp_netdev_mutex); + + ovs_rwlock_wrlock(&dp->port_rwlock); + if (get_port_by_name(dp, argv[2], &port)) { + unixctl_command_reply_error(conn, "unknown port"); + goto exit; + } + + port_no = u32_to_odp(atoi(argv[3])); + if (!port_no || port_no == ODPP_NONE) { + unixctl_command_reply_error(conn, "bad port number"); + goto exit; + } + if (dp_netdev_lookup_port(dp, port_no)) { + unixctl_command_reply_error(conn, "port number already in use"); + goto exit; + } + hmap_remove(&dp->ports, &port->node); + port->port_no = port_no; + hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0)); + seq_change(dp->port_seq); + unixctl_command_reply(conn, NULL); + +exit: + ovs_rwlock_unlock(&dp->port_rwlock); + dp_netdev_unref(dp); +} + +static void +dpif_dummy_register__(const char *type) +{ + struct dpif_class *class; + + class = xmalloc(sizeof *class); + *class = dpif_netdev_class; + class->type = xstrdup(type); + dp_register_provider(class); +} + +void +dpif_dummy_register(bool override) +{ + if (override) { + struct sset types; + const char *type; + + sset_init(&types); + dp_enumerate_types(&types); + SSET_FOR_EACH (type, &types) { + if (!dp_unregister_provider(type)) { + dpif_dummy_register__(type); + } + } + sset_destroy(&types); + } + + dpif_dummy_register__("dummy"); + + unixctl_command_register("dpif-dummy/change-port-number", + "DP PORT NEW-NUMBER", + 3, 3, dpif_dummy_change_port_number, NULL); +}