Open vSwitchのソースコードを読む(11) flow table のエントリ

static bool
handle_openflow(struct ofconn *ofconn, struct ofpbuf *ofp_msg)
{
    int error = handle_openflow__(ofconn, ofp_msg);
    if (error && error != OFPROTO_POSTPONE) {
        ofconn_send_error(ofconn, ofp_msg->data, error);
    }
    COVERAGE_INC(ofproto_recv_openflow);
    return error != OFPROTO_POSTPONE;
}

static enum ofperr
handle_openflow__(struct ofconn *ofconn, const struct ofpbuf *msg)
{
    const struct ofp_header *oh = msg->data;
    enum ofptype type;
    enum ofperr error;

    error = ofptype_decode(&type, oh);
    if (error) {
        return error;
    }

    switch (type) {
        /* OpenFlow requests. */
    case OFPTYPE_ECHO_REQUEST:
        return handle_echo_request(ofconn, oh);

    case OFPTYPE_FEATURES_REQUEST:
        return handle_features_request(ofconn, oh);

    （略）

    case OFPTYPE_FLOW_MOD:
        return handle_flow_mod(ofconn, oh);

    case OFPTYPE_BARRIER_REQUEST:
        return handle_barrier_request(ofconn, oh);

    （略）

    default:
        return OFPERR_OFPBRC_BAD_TYPE;
    }
}

static enum ofperr
handle_flow_mod(struct ofconn *ofconn, const struct ofp_header *oh)
{
    struct ofproto *ofproto = ofconn_get_ofproto(ofconn);
    struct ofputil_flow_mod fm;
    uint64_t ofpacts_stub[1024 / 8];
    struct ofpbuf ofpacts;
    enum ofperr error;
    long long int now;

    error = reject_slave_controller(ofconn);
    if (error) {
        goto exit;
    }

    ofpbuf_use_stub(&ofpacts, ofpacts_stub, sizeof ofpacts_stub);
    error = ofputil_decode_flow_mod(&fm, oh, ofconn_get_protocol(ofconn),
                                    &ofpacts);
    if (!error) {
        error = ofpacts_check(fm.ofpacts, fm.ofpacts_len,
                              &fm.match.flow, ofproto->max_ports);
    }
    if (!error) {
        error = handle_flow_mod__(ofproto, ofconn, &fm, oh);
    }
    if (error) {
        goto exit_free_ofpacts;
    }

    /* Record the operation for logging a summary report. */
    （略）
}

static enum ofperr
handle_flow_mod__(struct ofproto *ofproto, struct ofconn *ofconn,
                  const struct ofputil_flow_mod *fm,
                  const struct ofp_header *oh)
{
    if (ofproto->n_pending >= 50) {
        ovs_assert(!list_is_empty(&ofproto->pending));
        return OFPROTO_POSTPONE;
    }

    switch (fm->command) {
    case OFPFC_ADD:
        return add_flow(ofproto, ofconn, fm, oh);

    case OFPFC_MODIFY:
        return modify_flows_loose(ofproto, ofconn, fm, oh);

    case OFPFC_MODIFY_STRICT:
        return modify_flow_strict(ofproto, ofconn, fm, oh);

    case OFPFC_DELETE:
        return delete_flows_loose(ofproto, ofconn, fm, oh);

    case OFPFC_DELETE_STRICT:
        return delete_flow_strict(ofproto, ofconn, fm, oh);

    default:
        if (fm->command > 0xff) {
            VLOG_WARN_RL(&rl, "%s: flow_mod has explicit table_id but "
                         "flow_mod_table_id extension is not enabled",
                         ofproto->name);
        }
        return OFPERR_OFPFMFC_BAD_COMMAND;
    }
}

/* Implements OFPFC_ADD and the cases for OFPFC_MODIFY and OFPFC_MODIFY_STRICT
 * in which no matching flow already exists in the flow table.
 *
 * Adds the flow specified by 'ofm', which is followed by 'n_actions'
 * ofp_actions, to the ofproto's flow table.  Returns 0 on success, an OpenFlow
 * error code on failure, or OFPROTO_POSTPONE if the operation cannot be
 * initiated now but may be retried later.
 *
 * Upon successful return, takes ownership of 'fm->ofpacts'.  On failure,
 * ownership remains with the caller.
 *
 * 'ofconn' is used to retrieve the packet buffer specified in ofm->buffer_id,
 * if any. */
static enum ofperr
add_flow(struct ofproto *ofproto, struct ofconn *ofconn,
         const struct ofputil_flow_mod *fm, const struct ofp_header *request)
{
    struct oftable *table;
    struct ofopgroup *group;
    struct rule *victim;
    struct cls_rule cr;
    struct rule *rule;
    int error;

    error = check_table_id(ofproto, fm->table_id);
    if (error) {
        return error;
    }

    /* Pick table. */
    if (fm->table_id == 0xff) {
        uint8_t table_id;
        if (ofproto->ofproto_class->rule_choose_table) {
            error = ofproto->ofproto_class->rule_choose_table(ofproto,
                                                              &fm->match,
                                                              &table_id);
            if (error) {
                return error;
            }
            ovs_assert(table_id < ofproto->n_tables);
            table = &ofproto->tables[table_id];
        } else {
            table = &ofproto->tables[0];
        }
    } else if (fm->table_id < ofproto->n_tables) {
        table = &ofproto->tables[fm->table_id];
    } else {
        return OFPERR_OFPBRC_BAD_TABLE_ID;
    }

if (table->flags & OFTABLE_READONLY) {
    return OFPERR_OFPBRC_EPERM;
}

/* Allocate new rule and initialize classifier rule. */
rule = ofproto->ofproto_class->rule_alloc();
if (!rule) {
    VLOG_WARN_RL(&rl, "%s: failed to create rule (%s)",
                 ofproto->name, strerror(error));
    return ENOMEM;
}
cls_rule_init(&rule->cr, &fm->match, fm->priority);

/* Serialize against pending deletion. */
if (is_flow_deletion_pending(ofproto, &cr, table - ofproto->tables)) {
    cls_rule_destroy(&rule->cr);
    ofproto->ofproto_class->rule_dealloc(rule);
    return OFPROTO_POSTPONE;
}

/* Check for overlap, if requested. */
if (fm->flags & OFPFF_CHECK_OVERLAP
    && classifier_rule_overlaps(&table->cls, &rule->cr)) {
    cls_rule_destroy(&rule->cr);
    ofproto->ofproto_class->rule_dealloc(rule);
    return OFPERR_OFPFMFC_OVERLAP;
}

/* FIXME: Implement OFPFF12_RESET_COUNTS */

rule->ofproto = ofproto;
rule->pending = NULL;
rule->flow_cookie = fm->new_cookie;
rule->created = rule->modified = rule->used = time_msec();
rule->idle_timeout = fm->idle_timeout;
rule->hard_timeout = fm->hard_timeout;
rule->table_id = table - ofproto->tables;
rule->send_flow_removed = (fm->flags & OFPFF_SEND_FLOW_REM) != 0;
/* FIXME: Implement OF 1.3 flags OFPFF13_NO_PKT_COUNTS
   and OFPFF13_NO_BYT_COUNTS */
rule->ofpacts = xmemdup(fm->ofpacts, fm->ofpacts_len);
rule->ofpacts_len = fm->ofpacts_len;
rule->evictable = true;
rule->eviction_group = NULL;
list_init(&rule->expirable);
rule->monitor_flags = 0;
rule->add_seqno = 0;
rule->modify_seqno = 0;

/* Insert new rule. */
victim = oftable_replace_rule(rule);
if (victim && !rule_is_modifiable(victim)) {
    error = OFPERR_OFPBRC_EPERM;
} else if (victim && victim->pending) {
    error = OFPROTO_POSTPONE;
} else {
    struct ofoperation *op;
    struct rule *evict;

    if (classifier_count(&table->cls) > table->max_flows) {
        bool was_evictable;

        was_evictable = rule->evictable;
        rule->evictable = false;
        evict = choose_rule_to_evict(table);
        rule->evictable = was_evictable;

        if (!evict) {
            error = OFPERR_OFPFMFC_TABLE_FULL;
            goto exit;
        } else if (evict->pending) {
            error = OFPROTO_POSTPONE;
            goto exit;
        }
    } else {
        evict = NULL;
    }

    group = ofopgroup_create(ofproto, ofconn, request, fm->buffer_id);
    op = ofoperation_create(group, rule, OFOPERATION_ADD, 0);
    op->victim = victim;

    error = ofproto->ofproto_class->rule_construct(rule);
    if (error) {
        op->group->n_running--;
        ofoperation_destroy(rule->pending);
    } else if (evict) {
        delete_flow__(evict, group);
    }
    ofopgroup_submit(group);
}

exit:
    /* Back out if an error occurred. */
    if (error) {
        oftable_substitute_rule(rule, victim);
        ofproto_rule_destroy__(rule);
    }
    return error;
}

/* A single OpenFlow request can execute any number of operations.  The
 * ofopgroup maintain OpenFlow state common to all of the operations, e.g. the
 * ofconn to which an error reply should be sent if necessary.
 *
 * ofproto initiates some operations internally.  These operations are still
 * assigned to groups but will not have an associated ofconn. */
struct ofopgroup {
    struct ofproto *ofproto;    /* Owning ofproto. */
    struct list ofproto_node;   /* In ofproto's "pending" list. */
    struct list ops;            /* List of "struct ofoperation"s. */
    int n_running;              /* Number of ops still pending. */

    /* Data needed to send OpenFlow reply on failure or to send a buffered
     * packet on success.
     *
     * If list_is_empty(ofconn_node) then this ofopgroup never had an
     * associated ofconn or its ofconn's connection dropped after it initiated
     * the operation.  In the latter case 'ofconn' is a wild pointer that
     * refers to freed memory, so the 'ofconn' member must be used only if
     * !list_is_empty(ofconn_node).
     */
    struct list ofconn_node;    /* In ofconn's list of pending opgroups. */
    struct ofconn *ofconn;      /* ofconn for reply (but see note above). */
    struct ofp_header *request; /* Original request (truncated at 64 bytes). */
    uint32_t buffer_id;         /* Buffer id from original request. */
};

/* Creates and returns a new ofopgroup for 'ofproto'.
 *
 * If 'ofconn' is NULL, the new ofopgroup is not associated with any OpenFlow
 * connection.  The 'request' and 'buffer_id' arguments are ignored.
 *
 * If 'ofconn' is nonnull, then the new ofopgroup is associated with 'ofconn'.
 * If the ofopgroup eventually fails, then the error reply will include
 * 'request'.  If the ofopgroup eventually succeeds, then the packet with
 * buffer id 'buffer_id' on 'ofconn' will be sent by 'ofconn''s ofproto.
 *
 * The caller should add operations to the returned group with
 * ofoperation_create() and then submit it with ofopgroup_submit(). */
static struct ofopgroup *
ofopgroup_create(struct ofproto *ofproto, struct ofconn *ofconn,
                 const struct ofp_header *request, uint32_t buffer_id)
{
    struct ofopgroup *group = ofopgroup_create_unattached(ofproto);
    if (ofconn) {
        size_t request_len = ntohs(request->length);

        ovs_assert(ofconn_get_ofproto(ofconn) == ofproto);

        ofconn_add_opgroup(ofconn, &group->ofconn_node);
        group->ofconn = ofconn;
        group->request = xmemdup(request, MIN(request_len, 64));
        group->buffer_id = buffer_id;
    }
    return group;
}

/* Submits 'group' for processing.
 *
 * If 'group' contains no operations (e.g. none were ever added, or all of the
 * ones that were added completed synchronously), then it is destroyed
 * immediately.  Otherwise it is added to the ofproto's list of pending
 * groups. */
static void
ofopgroup_submit(struct ofopgroup *group)
{
    if (!group->n_running) {
        ofopgroup_complete(group);
    } else {
        list_push_back(&group->ofproto->pending, &group->ofproto_node);
        group->ofproto->n_pending++;
    }
}

static void
ofopgroup_complete(struct ofopgroup *group)
{
    struct ofproto *ofproto = group->ofproto;

    struct ofconn *abbrev_ofconn;
    ovs_be32 abbrev_xid;

    struct ofoperation *op, *next_op;
    int error;

    ovs_assert(!group->n_running);

    error = 0;
    LIST_FOR_EACH (op, group_node, &group->ops) {
        if (op->error) {
            error = op->error;
            break;
        }
    }

    if (!error && group->ofconn && group->buffer_id != UINT32_MAX) {
        LIST_FOR_EACH (op, group_node, &group->ops) {
            if (op->type != OFOPERATION_DELETE) {
                struct ofpbuf *packet;
                uint16_t in_port;

                error = ofconn_pktbuf_retrieve(group->ofconn, group->buffer_id,
                                               &packet, &in_port);
                if (packet) {
                    ovs_assert(!error);
                    error = rule_execute(op->rule, in_port, packet);
                }
                break;
            }
        }
    }

    if (!error && !list_is_empty(&group->ofconn_node)) {
        abbrev_ofconn = group->ofconn;
        abbrev_xid = group->request->xid;
    } else {
        abbrev_ofconn = NULL;
        abbrev_xid = htonl(0);
    }
    LIST_FOR_EACH_SAFE (op, next_op, group_node, &group->ops) {
        struct rule *rule = op->rule;

        /* We generally want to report the change to active OpenFlow flow
           monitors (e.g. NXST_FLOW_MONITOR).  There are three exceptions:

              - The operation failed.

              - The affected rule is not visible to controllers.

              - The operation's only effect was to update rule->modified. */
        if (!(op->error
              || ofproto_rule_is_hidden(rule)
              || (op->type == OFOPERATION_MODIFY
                  && op->ofpacts
                  && rule->flow_cookie == op->flow_cookie))) {
            /* Check that we can just cast from ofoperation_type to
             * nx_flow_update_event. */
            BUILD_ASSERT_DECL((enum nx_flow_update_event) OFOPERATION_ADD
                              == NXFME_ADDED);
            BUILD_ASSERT_DECL((enum nx_flow_update_event) OFOPERATION_DELETE
                              == NXFME_DELETED);
            BUILD_ASSERT_DECL((enum nx_flow_update_event) OFOPERATION_MODIFY
                              == NXFME_MODIFIED);

            ofmonitor_report(ofproto->connmgr, rule,
                             (enum nx_flow_update_event) op->type,
                             op->reason, abbrev_ofconn, abbrev_xid);
        }

        rule->pending = NULL;

        switch (op->type) {
        case OFOPERATION_ADD:
            if (!op->error) {
                uint16_t vid_mask;

                ofproto_rule_destroy__(op->victim);
                vid_mask = minimask_get_vid_mask(&rule->cr.match.mask);
                if (vid_mask == VLAN_VID_MASK) {
                    if (ofproto->vlan_bitmap) {
                        uint16_t vid = miniflow_get_vid(&rule->cr.match.flow);
                        if (!bitmap_is_set(ofproto->vlan_bitmap, vid)) {
                            bitmap_set1(ofproto->vlan_bitmap, vid);
                            ofproto->vlans_changed = true;
                        }
                    } else {
                        ofproto->vlans_changed = true;
                    }
                }
            } else {
                oftable_substitute_rule(rule, op->victim);
                ofproto_rule_destroy__(rule);
            }
            break;

        case OFOPERATION_DELETE:
            ovs_assert(!op->error);
            ofproto_rule_destroy__(rule);
            op->rule = NULL;
            break;

        case OFOPERATION_MODIFY:
            if (!op->error) {
                rule->modified = time_msec();
            } else {
                rule->flow_cookie = op->flow_cookie;
                if (op->ofpacts) {
                    free(rule->ofpacts);
                    rule->ofpacts = op->ofpacts;
                    rule->ofpacts_len = op->ofpacts_len;
                    op->ofpacts = NULL;
                    op->ofpacts_len = 0;
                }
            }
            break;

        default:
            NOT_REACHED();
        }

        ofoperation_destroy(op);
    }

    ofmonitor_flush(ofproto->connmgr);

    if (!list_is_empty(&group->ofproto_node)) {
        ovs_assert(ofproto->n_pending > 0);
        ofproto->n_pending--;
        list_remove(&group->ofproto_node);
    }
    if (!list_is_empty(&group->ofconn_node)) {
        list_remove(&group->ofconn_node);
        if (error) {
            ofconn_send_error(group->ofconn, group->request, error);
        }
        connmgr_retry(ofproto->connmgr);
    }

    free(group->request);
    free(group);
}

static void
delete_flow__(struct rule *rule, struct ofopgroup *group)
{
    struct ofproto *ofproto = rule->ofproto;

    ofproto_rule_send_removed(rule, OFPRR_DELETE);

    ofoperation_create(group, rule, OFOPERATION_DELETE, OFPRR_DELETE);
    oftable_remove_rule(rule);
    ofproto->ofproto_class->rule_destruct(rule);
}

/* Deletes the rules listed in 'rules'.
 *
 * Returns 0 on success, otherwise an OpenFlow error code. */
static enum ofperr
delete_flows__(struct ofproto *ofproto, struct ofconn *ofconn,
               const struct ofp_header *request, struct list *rules)
{
    struct rule *rule, *next;
    struct ofopgroup *group;

    group = ofopgroup_create(ofproto, ofconn, request, UINT32_MAX);
    LIST_FOR_EACH_SAFE (rule, next, ofproto_node, rules) {
        delete_flow__(rule, group);
    }
    ofopgroup_submit(group);

    return 0;
}