master.c

Go to the documentation of this file.

/*
 *   This program is free software; you can redistribute it and/or modify
 *   it under the terms of the GNU General Public License as published by
 *   the Free Software Foundation; either version 2 of the License, or
 *   (at your option) any later version.
 *
 *   This program is distributed in the hope that it will be useful,
 *   but WITHOUT ANY WARRANTY; without even the implied warranty of
 *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *   GNU General Public License for more details.
 *
 *   You should have received a copy of the GNU General Public License
 *   along with this program; if not, write to the Free Software
 *   Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
 */
 
/**
 * $Id: d885d6052b16dfdcc4b21267f339c695a0fb01c1 $
 * @file io/master.c
 * @brief Master IO handler
 *
 * @copyright 2018 Alan DeKok (aland@freeradius.org)
 */
#include <freeradius-devel/io/listen.h>
#include <freeradius-devel/io/master.h>
 
#include <freeradius-devel/util/syserror.h>
 
FR_DLIST_TYPES(fr_io_nak_list)
FR_DLIST_TYPEDEFS(fr_io_nak_list, fr_io_nak_list_head_t, fr_io_nak_list_entry_t)
 
/** A negative cache entry.
 *
 *  This is smaller and simpler than using a #fr_io_client_t.  Plus, it doesn't affect the lifetime,
 *  ownership, or parenting of a #fr_io_client_t.
 */
typedef struct {
        fr_ipaddr_t                     src_ipaddr;                     //!< the address which was NAK'd
        fr_time_t                       expires;                        //!< when this entry stops being used
        fr_io_nak_list_entry_t          entry;
} fr_io_nak_entry_t;
 
FR_DLIST_FUNCS(fr_io_nak_list, fr_io_nak_entry_t, entry)
 
typedef struct {
        fr_event_list_t                 *el;                            //!< event list, for the master socket.
        fr_network_t                    *nr;                            //!< network for the master socket
 
        fr_trie_t                       *trie;                          //!< trie of clients
        fr_trie_t                       *nak_trie;                      //!< trie of NAK clients
        fr_io_nak_list_head_t           nak_list;                       //!< time ordered list of NAK entries.
 
        fr_heap_t                       *pending_clients;               //!< heap of pending clients
        fr_heap_t                       *alive_clients;                 //!< heap of active dynamic clients
 
        fr_listen_t                     *listen;                        //!< The master IO path
        fr_listen_t                     *child;                         //!< The child (app_io) IO path
        fr_schedule_t                   *sc;                            //!< the scheduler
 
        uint32_t                        num_connections;                //!< number of dynamic connections
        uint32_t                        num_pending_packets;            //!< number of pending packets
        uint64_t                        client_id;                      //!< Unique client identifier.
 
        struct {
                fr_rate_limit_t                 accept_failed;
                fr_rate_limit_t                 alloc_failed;
                fr_rate_limit_t                 bad_type;
                fr_rate_limit_t                 conn_alloc_failed;
                fr_rate_limit_t                 max_connections;
                fr_rate_limit_t                 queue_full;
                fr_rate_limit_t                 repeat_nak;
                fr_rate_limit_t                 too_many_pending;
                fr_rate_limit_t                 tracking_failed;
                fr_rate_limit_t                 unknown_client;
        } rate_limit;
} fr_io_thread_t;
 
/** A saved packet
 *
 */
typedef struct {
        fr_heap_index_t         heap_id;
        uint32_t                priority;
        fr_time_t               recv_time;
        fr_io_track_t           *track;
        uint8_t                 *buffer;
        size_t                  buffer_len;
} fr_io_pending_packet_t;
 
 
/** Client states
 *
 */
typedef enum : uint8_t {
        PR_CLIENT_INVALID = 0,
        PR_CLIENT_STATIC,                               //!< static / global clients
        PR_CLIENT_DYNAMIC,                              //!< dynamically defined client
        PR_CLIENT_CONNECTED,                            //!< dynamically defined client in a connected socket
        PR_CLIENT_PENDING,                              //!< dynamic client pending definition
} fr_io_client_state_t;
 
/*
 *      Dynamic clients are run through the normal src/process/foo state machine.
 *
 *      request->async->packet_ctx is an fr_io_track_t
 *
 *      track->dynamic is set to a non-zero value.
 *
 *      The dynamic client code returns a buffer of 1 byte for a NAK.
 *
 *      If the client creation is successful, then it does talloc(NULL, fr_client_t),
 *      fills out the structure, and sends the pointer in the buffer (8 bytes).
 *
 *      This code will take over ownership of the structure, and
 *      create the dynamic client.
 */
 
typedef struct fr_io_connection_s fr_io_connection_t;
 
/** Client definitions for master IO
 *
 */
struct fr_io_client_s {
        fr_io_connection_t              *connection;    //!< parent connection
        fr_io_client_state_t            state;          //!< state of this client
        fr_ipaddr_t                     src_ipaddr;     //!< packets come from this address
        fr_ipaddr_t                     network;        //!< network for dynamic clients
        fr_client_t                     *radclient;     //!< old-style definition of this client
 
        int                             packets;        //!< number of packets using this client
        fr_heap_index_t                 pending_id;     //!< for pending clients
        fr_heap_index_t                 alive_id;       //!< for all clients
 
        bool                            use_connected;  //!< does this client allow connected sub-sockets?
        bool                            ready_to_delete; //!< are we ready to delete this client?
        bool                            in_trie;        //!< is the client in the trie?
 
        fr_io_instance_t const          *inst;          //!< parent instance for master IO handler
        fr_io_thread_t                  *thread;
        fr_timer_t                      *ev;            //!< when we clean up the client
        fr_rb_tree_t                    *table;         //!< tracking table for packets
 
        fr_heap_t                       *pending;       //!< pending packets for this client
        fr_hash_table_t                 *addresses;     //!< list of src/dst addresses used by this client
 
        pthread_mutex_t                 mutex;          //!< for parent / child signaling
        fr_hash_table_t                 *ht;            //!< for tracking connected sockets
};
 
/** Track a connection
 *
 *  This structure contains information about the connection,
 *  a pointer to the library instance so that we can clean up on exit,
 *  and the listener.
 *
 *  It also points to a client structure which is for this connection,
 *  and only this connection.
 *
 *  Finally, a pointer to the parent client, so that the child can
 *  tell the parent it's alive, and the parent can push packets to the
 *  child.
 */
struct fr_io_connection_s {
        char const                      *name;          //!< taken from proto_FOO_TRANSPORT
        int                             packets;        //!< number of packets using this connection
        fr_io_address_t                 *address;       //!< full information about the connection.
        fr_listen_t                     *listen;        //!< master listener for this socket
        fr_listen_t                     *child;         //!< child listener (app_io) for this socket
        fr_io_client_t                  *client;        //!< our local client (pending or connected).
        fr_io_client_t                  *parent;        //!< points to the parent client.
        module_instance_t               *mi;            //!< for submodule
 
        bool                            dead;           //!< roundabout way to get the network side to close a socket
        bool                            paused;         //!< event filter doesn't like resuming something that isn't paused
        bool                            in_parent_hash; //!< for tracking thread issues
        fr_event_list_t                 *el;            //!< event list for this connection
        fr_network_t                    *nr;            //!< network for this connection
};
 
static fr_event_update_t pause_read[] = {
        FR_EVENT_SUSPEND(fr_event_io_func_t, read),
        { 0 }
};
 
static fr_event_update_t resume_read[] = {
        FR_EVENT_RESUME(fr_event_io_func_t, read),
        { 0 }
};
 
static int track_free(fr_io_track_t *track)
{
        FR_TIMER_DELETE_RETURN(&track->ev);
        talloc_free_children(track);
 
        fr_assert(track->client->packets > 0);
        track->client->packets--;
 
        return 0;
}
 
static int track_dedup_free(fr_io_track_t *track)
{
        void *found;
 
        fr_assert(track->client->table != NULL);
 
        /*
         *      If the tree is being freed, then we don't try to remove ourselves from it.  Doing so would
         *      free this node, and therefore corrupt the tree.
         *
         *      The talloc code will take care of cleaning up the children and events when this chunk is
         *      freed.
         */
        if (track->client->table->being_freed) return 0;
 
        fr_rb_find(&found, track->client->table, track);
        fr_assert(found != NULL);
 
        if (fr_rb_delete(track->client->table, track) != 0) {
                fr_assert(0);
        }
 
        return track_free(track);
}
 
/*
 *  Return negative numbers to put 'one' at the top of the heap.
 *  Return positive numbers to put 'two' at the top of the heap.
 */
static fr_cmp_ret_t pending_packet_cmp(void const *one, void const *two)
{
        fr_io_pending_packet_t const *a = talloc_get_type_abort_const(one, fr_io_pending_packet_t);
        fr_io_pending_packet_t const *b = talloc_get_type_abort_const(two, fr_io_pending_packet_t);
        int ret;
 
        /*
         *      Higher priority elements are larger than lower
         *      priority elements.  So if "a" is larger than "b", we
         *      wish to prefer "a".
         */
        ret = CMP_PREFER_LARGER(a->priority, b->priority);
        if (ret != 0) return ret;
 
        /*
         *      Smaller numbers mean packets were received earlier.
         *      We want to process packets in time order.  So if "a"
         *      is smaller than "b", we wish to prefer "a".
         *
         *      After that, it doesn't really matter what order the
         *      packets go in.  Since we'll never have two identical
         *      "recv_time" values, the code should never get here.
         */
        return CMP_PREFER_SMALLER(fr_time_unwrap(a->recv_time), fr_time_unwrap(b->recv_time));
}
 
/*
 *      Order clients in the pending_clients heap, based on the
 *      packets that they contain.
 */
static fr_cmp_ret_t pending_client_cmp(void const *one, void const *two)
{
        fr_io_pending_packet_t const *a;
        fr_io_pending_packet_t const *b;
 
        fr_io_client_t const *c1 = talloc_get_type_abort_const(one, fr_io_client_t);
        fr_io_client_t const *c2 = talloc_get_type_abort_const(two, fr_io_client_t);
 
        a = fr_heap_peek(c1->pending);
        b = fr_heap_peek(c2->pending);
 
        fr_assert(a != NULL);
        fr_assert(b != NULL);
 
        return pending_packet_cmp(a, b);
}
 
 
static fr_cmp_ret_t address_cmp(void const *one, void const *two)
{
        fr_io_address_t const *a = talloc_get_type_abort_const(one, fr_io_address_t);
        fr_io_address_t const *b = talloc_get_type_abort_const(two, fr_io_address_t);
        fr_cmp_ret_t ret;
 
        CMP_RETURN(a, b, socket.inet.src_port);
        CMP_RETURN(a, b, socket.inet.dst_port);
        CMP_RETURN(a, b, socket.inet.ifindex);
 
        ret = fr_ipaddr_cmp(&a->socket.inet.src_ipaddr, &b->socket.inet.src_ipaddr);
        if (ret != 0) return ret;
 
        return fr_ipaddr_cmp(&a->socket.inet.dst_ipaddr, &b->socket.inet.dst_ipaddr);
}
 
static uint32_t connection_hash(void const *ctx)
{
        uint32_t hash;
        fr_io_connection_t const *c = talloc_get_type_abort_const(ctx, fr_io_connection_t);
 
        hash = fr_hash(&c->address->socket.inet.src_ipaddr, sizeof(c->address->socket.inet.src_ipaddr));
        hash = fr_hash_update(&c->address->socket.inet.src_port, sizeof(c->address->socket.inet.src_port), hash);
 
        hash = fr_hash_update(&c->address->socket.inet.ifindex, sizeof(c->address->socket.inet.ifindex), hash);
 
        hash = fr_hash_update(&c->address->socket.inet.dst_ipaddr, sizeof(c->address->socket.inet.dst_ipaddr), hash);
        return fr_hash_update(&c->address->socket.inet.dst_port, sizeof(c->address->socket.inet.dst_port), hash);
}
 
static fr_cmp_ret_t connection_cmp(void const *one, void const *two)
{
        fr_io_connection_t const *a = talloc_get_type_abort_const(one, fr_io_connection_t);
        fr_io_connection_t const *b = talloc_get_type_abort_const(two, fr_io_connection_t);
 
        return address_cmp(a->address, b->address);
}
 
 
static fr_cmp_ret_t track_cmp(void const *one, void const *two)
{
        fr_io_track_t const *a = talloc_get_type_abort_const(one, fr_io_track_t);
        fr_io_track_t const *b = talloc_get_type_abort_const(two, fr_io_track_t);
        int ret;
 
        fr_assert(a->client != NULL);
        fr_assert(b->client != NULL);
        fr_assert(a->client == b->client); /* tables are per-client */
 
        fr_assert(!a->client->connection);
        fr_assert(!b->client->connection);
 
        /*
         *      Unconnected sockets must check src/dst ip/port.
         */
        ret = address_cmp(a->address, b->address);
        if (ret != 0) return ret;
 
        /*
         *      Call the per-protocol comparison function.
         */
        ret = a->client->inst->app_io->track_compare(a->client->inst->app_io_instance,
                                                     a->client->thread->child->thread_instance,
                                                     a->client->radclient,
                                                     a->packet, b->packet);
        return CMP(ret, 0);
}
 
 
static fr_cmp_ret_t track_connected_cmp(void const *one, void const *two)
{
        fr_io_track_t const *a = talloc_get_type_abort_const(one, fr_io_track_t);
        fr_io_track_t const *b = talloc_get_type_abort_const(two, fr_io_track_t);
        int ret;
 
        fr_assert(a->client != NULL);
        fr_assert(b->client != NULL);
 
        fr_assert(a->client->connection);
        fr_assert(b->client->connection);
        fr_assert(a->client == b->client);
        fr_assert(a->client->connection == b->client->connection);
 
        /*
         *      Note that we pass the connection "client", as
         *      we may do negotiation specific to this connection.
         */
        ret = a->client->inst->app_io->track_compare(a->client->inst->app_io_instance,
                                                     a->client->connection->child->thread_instance,
                                                     a->client->connection->client->radclient,
                                                     a->packet, b->packet);
        return CMP(ret, 0);
}
 
 
static fr_io_pending_packet_t *pending_packet_pop(fr_io_thread_t *thread)
{
        fr_io_client_t *client;
        fr_io_pending_packet_t *pending;
 
        fr_heap_pop((void **)&client, &thread->pending_clients);
        if (!client) {
                fr_assert(thread->num_pending_packets == 0);
 
                /*
                 *      99% of the time we don't have pending clients.
                 *      So we might as well free this, so that the
                 *      caller doesn't keep checking us for every packet.
                 */
                talloc_free(thread->pending_clients);
                thread->pending_clients = NULL;
                return NULL;
        }
 
        fr_heap_pop((void **)&pending, &client->pending);
        fr_assert(pending != NULL);
 
        /*
         *      If the client has more packets pending, add it back to
         *      the heap.
         */
        if (fr_heap_num_elements(client->pending) > 0) {
                if (fr_heap_insert(&thread->pending_clients, client) < 0) {
                        fr_assert(0 == 1);
                }
        }
 
        fr_assert(thread->num_pending_packets > 0);
        thread->num_pending_packets--;
 
        return pending;
}
 
static fr_client_t *radclient_clone(TALLOC_CTX *ctx, fr_client_t const *parent)
{
        fr_client_t *c;
 
        if (!parent) return NULL;
 
        c = talloc_zero(ctx, fr_client_t);
        if (!c) return NULL;
 
        /*
         *      Do NOT set ipaddr or src_ipaddr.  The caller MUST do this!
         */
 
#define DUP_FIELD(_x) do { if (parent->_x) {c->_x = talloc_strdup(c, parent->_x); if (!c->_x) {goto error;}}} while (0)
#define COPY_FIELD(_x) c->_x = parent->_x
 
        DUP_FIELD(longname);
        DUP_FIELD(shortname);
        DUP_FIELD(secret);
        DUP_FIELD(nas_type);
        DUP_FIELD(server);
 
        COPY_FIELD(require_message_authenticator);
        COPY_FIELD(require_message_authenticator_is_set);
#ifdef NAS_VIOLATES_RFC
        COPY_FIELD(allow_vulnerable_clients);
#endif
        COPY_FIELD(limit_proxy_state);
        COPY_FIELD(limit_proxy_state_is_set);
        /* dynamic MUST be false */
        COPY_FIELD(server_cs);
        COPY_FIELD(cs);
        COPY_FIELD(proto);
        COPY_FIELD(active);
 
        COPY_FIELD(use_connected);
 
#ifdef WITH_TLS
        COPY_FIELD(tls_required);
#endif
 
        c->ipaddr = parent->ipaddr;
        c->src_ipaddr = parent->src_ipaddr;
 
        return c;
 
        /*
         *      @todo - fill in other fields, too!
         */
 
error:
        talloc_free(c);
        return NULL;
}
#undef COPY_FIELD
#undef DUP_FIELD
 
/** Expire entries in the cache.
 *
 *  Remove old entries from the head of the queue if they have expired, OR if the NAK list is large.
 *
 *  @todo - make the size of the NAK list configurable.  For now, a hard-coded limit of 1000 isn't terrible.
 *  If there are more NAK IPs than that, the admin should set up firewall rules to block traffic.
 */
static void fr_io_nak_expire(fr_io_thread_t *thread, fr_time_t now)
{
        fr_io_nak_entry_t *nak;
 
        while ((nak = fr_io_nak_list_head(&thread->nak_list)) != NULL) {
                if (fr_time_gt(nak->expires, now) &&
                    (fr_io_nak_list_num_elements(&thread->nak_list) <= 1000)) {
                        break;
                }
 
                /*
                 *      @todo - if the cache is full because we're expiring entries too quickly, then that's
                 *      likely a firewall issue for the admin to resolve.
                 */
 
                (void) fr_trie_remove_by_key(thread->nak_trie, &nak->src_ipaddr.addr, nak->src_ipaddr.prefix);
                fr_io_nak_list_remove(&thread->nak_list, nak);
                talloc_free(nak);
        }
 
}
 
/** Add a negative cache entry.
 *
 *  The new entry is either added, or it has its expiry updated.  Deleting the existing entry would cause
 *  problems for the lifetime of the NAK entry.
 *
 * @param[in] thread            thread instance
 * @param[in] src_ipaddr        to look add
 * @param[in] lifetime          how long the entry lives for.
 */
static void fr_io_nak_insert(fr_io_thread_t *thread, fr_ipaddr_t const *src_ipaddr, fr_time_delta_t lifetime)
{
        fr_time_t now = fr_time();
        fr_io_nak_entry_t *nak;
 
        fr_io_nak_expire(thread, now);
 
        nak = fr_trie_match_by_key(thread->nak_trie, &src_ipaddr->addr, src_ipaddr->prefix);
        if (nak) {
                fr_io_nak_list_remove(&thread->nak_list, nak);
                goto do_insert;
        }
 
        /*
         *      @todo - rate limited warnings if the NAK cache is full.
         */
        if (fr_io_nak_list_num_elements(&thread->nak_list) > 1000) {
                return;
        }
 
        /*
         *      No match, allocate and insert a new entry.
         */
        MEM(nak = talloc_zero(thread->nak_trie, fr_io_nak_entry_t));
        nak->src_ipaddr = *src_ipaddr;
        fr_io_nak_list_entry_init(nak);
 
        if (fr_trie_insert_by_key(thread->nak_trie, &src_ipaddr->addr, src_ipaddr->prefix, nak)) {
                talloc_free(nak);
                return;
        }
 
do_insert:
        /*
         *      This assumes that the lifetime is constant.  If it varies per IP, we will need to move the
         *      dlist to an RB tree.
         */
        nak->expires = fr_time_add(now, lifetime);
        fr_io_nak_list_insert_tail(&thread->nak_list, nak);
}
 
/** Look up a negative cache entry.
 *
 *  Entries are expired first, and then looked up.
 *
 * @param[in] thread            thread instance
 * @param[in] src_ipaddr        to look up.
 * @return
 *      - true if the address is NAK'd, and the caller must refuse it.
 *      - false if it is not NAK'd, or if the entry has expired.
 */
static bool fr_io_nak_find(fr_io_thread_t *thread, fr_ipaddr_t const *src_ipaddr)
{
        fr_time_t now = fr_time();
        fr_io_nak_entry_t *nak;
 
        fr_io_nak_expire(thread, now);
 
        nak = fr_trie_lookup_by_key(thread->nak_trie, &src_ipaddr->addr, src_ipaddr->prefix);
        return (nak != NULL);
}
 
 
/** Count the number of connections used by active clients.
 */
static int count_connections(UNUSED uint8_t const *key, UNUSED size_t keylen, void *data, void *ctx)
{
        fr_io_client_t *client = talloc_get_type_abort(data, fr_io_client_t);
        int connections;
 
        pthread_mutex_lock(&client->mutex);
 
        if (!client->ht) {
                pthread_mutex_unlock(&client->mutex);
                return 0;
        }
 
        connections = fr_hash_table_num_elements(client->ht);
        pthread_mutex_unlock(&client->mutex);
 
        /*
         *      Don't check "use_connected".  Pending dynamic clients get an entry in the connection tracking
         *      table before the client is defined, and therefore before "use_connected" is set.  As a result,
         *      we can't check the value of "use_connected" until much later.
         */
 
        *((uint32_t *) ctx) += connections;
 
        return 0;
}
 
 
static int _client_free(fr_io_client_t *client)
{
        /*
         *      The mutex is initialized whenever the connection tracking table is created, which can happen
         *      for pending clients which do not (yet) have "use_connected" set.  Since the mutex creation is
         *      conditional on the existence of the connection tracking table, we make the mutex deletion
         *      conditional on the existence of the tracking table.
         */
        if (client->ht) (void) pthread_mutex_destroy(&client->mutex);
 
        TALLOC_FREE(client->pending);
 
        return 0;
}
 
static void client_pending_free(fr_io_client_t *client)
{
        size_t num;
 
        fr_assert(!client->connection);
 
        if (!client->pending) return;
 
        num = fr_heap_num_elements(client->pending);
 
        fr_assert(client->thread->num_pending_packets >= num);
        client->thread->num_pending_packets -= num;
 
        TALLOC_FREE(client->pending);
}
 
 
static int connection_free(fr_io_connection_t *connection)
{
        /*
         *      This is it's own talloc context, as there are
         *      thousands of packets associated with it.
         */
        TALLOC_FREE(connection->client);
 
        return 0;
}
 
/** Create a new connection.
 *
 *  Called ONLY from the master socket.
 */
static fr_io_connection_t *fr_io_connection_alloc(fr_io_instance_t const *inst,
                                                  fr_io_thread_t *thread,
                                                  fr_io_client_t *client, int fd,
                                                  fr_io_address_t *address)
{
        int ret;
        fr_io_connection_t *connection;
        module_instance_t *mi = NULL;
        fr_listen_t *li;
        fr_client_t *radclient;
 
 
        /*
         *      Reload the app_io module as a "new" library.  This
         *      causes the link count for the library to be correct.
         *      It also allocates a new instance data for it, too.
         *      Passing CONF_SECTION of NULL ensures that there's no
         *      config for it, as we'll just clone it's contents from
         *      the original.  It also means that detach should be
         *      called when the instance data is freed.
         */
        {
                CONF_SECTION *cs;
                char *inst_name;
 
                if (inst->max_connections || client->radclient->limit.max_connections) {
                        uint32_t max_connections = inst->max_connections ? inst->max_connections : client->radclient->limit.max_connections;
 
                        /*
                         *      We've hit the connection limit.  Walk
                         *      over all clients with connections, and
                         *      count the number of connections used.
                         */
                        if (thread->num_connections >= max_connections) {
                                thread->num_connections = 0;
 
                                (void) fr_trie_walk(thread->trie, &thread->num_connections, count_connections);
 
                                if ((thread->num_connections + 1) >= max_connections) {
                                        RATE_LIMIT_LOCAL(&thread->rate_limit.max_connections, INFO,
                                                         "proto_%s - Ignoring connection from client %s - 'max_connections' limit reached.",
                                                         inst->app->common.name, client->radclient->shortname);
                                        if (fd >= 0) close(fd);
                                        return NULL;
                                }
                        }
                }
 
                /*
                 *      Add a client module into a sublist
                 */
                inst_name = talloc_asprintf(NULL, "%"PRIu64, thread->client_id++);
                mi = module_instance_copy(inst->clients, inst->submodule, inst_name);
                talloc_free(inst_name);
 
                cs = cf_section_dup(mi, NULL, inst->submodule->conf,
                                    cf_section_name1(inst->submodule->conf),
                                    cf_section_name2(inst->submodule->conf), false);
 
                if (!cs) {
                        cf_log_err(inst->server_cs, "Failed duplicating module config");
                        goto cleanup;
                }
 
                /*
                 *      Clear the "dynamic_clients" flag, so that the child instantiate routines don't check
                 *      the network allow / deny list when instantiating child connections.
                 *
                 *      This is a short-term and minimal hack to get the problem fixed.  A longer term
                 *      solution would be to update fr_master_io_network() so that it sets caches the trie
                 *      _and_ the dynamic client flag in connection data structure.  Which then means that the
                 *      mod_network_get() API could also go away.
                 *
                 *      But doing that involves more rearchitecture and code changes, which we're avoiding at
                 *      this time.
                 */
                cf_pair_replace_or_add(cs, "dynamic_clients", "no");
 
                if (module_instance_conf_parse(mi, cs) < 0) {
                        cf_log_err(inst->server_cs, "Failed parsing module config");
                        goto cleanup;
                }
 
                /* Thread local module lists never run bootstrap */
                if (module_instantiate(mi) < 0) {
                        cf_log_err(inst->server_cs, "Failed instantiating module");
                        goto cleanup;
                }
 
                if (module_thread_instantiate(mi, mi, thread->el) < 0) {
                        cf_log_err(inst->server_cs, "Failed instantiating module");
                        goto cleanup;
                }
 
                /*
                 *      FIXME - Instantiate the new module?!
                 */
                fr_assert(mi != NULL);
        }
 
        MEM(connection = talloc_zero(mi, fr_io_connection_t));
        MEM(connection->address = talloc_memdup(connection, address, sizeof(*address)));
        (void) talloc_set_name_const(connection->address, "fr_io_address_t");
 
        connection->parent = client;
        connection->mi = mi;
 
        MEM(connection->client = talloc_named(NULL, sizeof(fr_io_client_t), "fr_io_client_t"));
        memset(connection->client, 0, sizeof(*connection->client));
 
        MEM(connection->client->radclient = radclient = radclient_clone(connection->client, client->radclient));
 
        talloc_set_destructor(connection->client, _client_free);
        talloc_set_destructor(connection, connection_free);
 
        connection->client->pending_id = FR_HEAP_INDEX_INVALID;
        connection->client->alive_id = FR_HEAP_INDEX_INVALID;
        connection->client->connection = connection;
 
        /*
         *      Create the packet tracking table for this client.
         *
         *      #todo - unify the code with static clients?
         */
        if (inst->app_io->track_duplicates) {
                MEM(connection->client->table = fr_rb_inline_talloc_alloc(client, fr_io_track_t, node,
                                                                          track_connected_cmp, NULL));
        }
 
        /*
         *      Set this radclient to be dynamic, and active.
         */
        radclient->dynamic = true;
        radclient->active = true;
 
        /*
         *      address->socket.inet.client points to a "static" client.  We want
         *      to clean up everything associated with the connection
         *      when it closes.  So we need to point to our own copy
         *      of the client here.
         */
        connection->address->radclient = connection->client->radclient;
        connection->client->inst = inst;
        connection->client->thread = thread;
 
        /*
         *      Create a heap for packets which are pending for this
         *      client.
         */
        MEM(connection->client->pending = fr_heap_alloc(connection->client, pending_packet_cmp,
                                                         fr_io_pending_packet_t, heap_id, 0));
 
        /*
         *      Clients for connected sockets are always a /32 or /128.
         */
        connection->client->src_ipaddr = address->socket.inet.src_ipaddr;
        connection->client->network = address->socket.inet.src_ipaddr;
 
        /*
         *      Don't initialize mutex or hash table.
         *      Connections cannot spawn other connections.
         */
 
        /*
         *      If this client state is pending, then the connection
         *      state is pending, too.  That allows NAT gateways to be
         *      defined dynamically, AND for them to have multiple
         *      connections, each with a different client.  This
         *      allows for different shared secrets to be used for
         *      different connections.  Once the client gets defined
         *      for this connection, it will be either "connected" or
         *      not.  If connected, then the parent client remains
         *      PENDING.  Otherwise, the parent client is moved to
         *      DYNAMIC
         *
         *      If this client state is static or dynamic,
         *      then we're just using connected sockets behind
         *      that client.  The connections here all use the
         *      same shared secret, but they use different
         *      sockets, so they allow for sharing of IO
         *      across CPUs / threads.
         */
        switch (client->state) {
        case PR_CLIENT_PENDING:
                connection->client->state = PR_CLIENT_PENDING;
 
                /*
                 *      Needed for rlm_radius, which refuses to proxy packets
                 *      that define a dynamic client.
                 */
                radclient->active = false;
                break;
 
        case PR_CLIENT_STATIC:
        case PR_CLIENT_DYNAMIC:
                connection->client->state = PR_CLIENT_CONNECTED;
                break;
 
        case PR_CLIENT_INVALID:
        case PR_CLIENT_CONNECTED:
                fr_assert(0 == 1);
                goto cleanup;
        }
 
        {
                /*
                 *      Get the child listener.
                 */
                MEM(li = connection->child = talloc(connection, fr_listen_t));
                memcpy(li, thread->listen, sizeof(*li));
 
                /*
                 *      Glue in the actual app_io
                 */
                li->connected = true;
                li->app_io = thread->child->app_io;
                li->cs = inst->app_io_conf;
                li->thread_instance = connection;
                li->app_io_instance = mi->data;
                li->track_duplicates = thread->child->app_io->track_duplicates;
 
                /*
                 *      Create writable thread instance data.
                 */
                connection->child->thread_instance = talloc_zero_array(NULL, uint8_t,
                                                                       inst->app_io->common.thread_inst_size);
                talloc_set_destructor(connection->child, fr_io_listen_free);
                talloc_set_name(connection->child->thread_instance, "proto_%s_thread_t",
                                inst->app_io->common.name);
 
                /*
                 *      This is "const", and the user can't
                 *      touch it.  So we just reuse the same
                 *      configuration everywhere.
                 */
                connection->child->app_io_instance = inst->app_io_instance;
 
                /*
                 *      Create the listener, based on our listener.
                 */
                MEM(li = connection->listen = talloc(connection, fr_listen_t));
 
                /*
                 *      Note that our instance is effectively 'const'.
                 *
                 *      i.e. we can't add things to it.  Instead, we have to
                 *      put all variable data into the connection.
                 */
                memcpy(li, thread->listen, sizeof(*li));
 
                /*
                 *      Glue in the connection to the listener.
                 */
                fr_assert(li->app_io == &fr_master_app_io);
 
                li->connected = true;
                li->thread_instance = connection;
                li->cs = inst->app_io_conf;
                li->app_io_instance = li->thread_instance;
                li->track_duplicates = thread->child->app_io->track_duplicates;
 
                /*
                 *      Instantiate the child, and open the socket.
                 */
                fr_assert(inst->app_io->connection_set != NULL);
 
                if (inst->app_io->connection_set(connection->child, connection->address) < 0) {
                        DEBUG("proto_%s - Failed setting connection for socket.", inst->app->common.name);
                        goto cleanup;
                }
 
                /*
                 *      UDP sockets: open a new socket, and then
                 *      connect it to the client.  This emulates the
                 *      behavior of accept().
                 *
                 *      Note that there is a small window between the
                 *      bind() and connect() where UDP packets for the
                 *      wildcard socket can get received by this
                 *      socket.  We hope that this time frame is as
                 *      small as possible.
                 *
                 *      i.e. we ignore the problem, and don't
                 *      currently check dst ip/port for UDP packets
                 *      received on connected sockets.
                 */
                if (fd < 0) {
                        socklen_t salen;
                        struct sockaddr_storage src;
 
                        if (fr_ipaddr_to_sockaddr(&src, &salen,
                                                  &connection->address->socket.inet.src_ipaddr,
                                                  connection->address->socket.inet.src_port) < 0) {
                                DEBUG("proto_%s - Failed getting IP address", inst->app->common.name);
                                talloc_free(mi);
                                return NULL;
                        }
 
                        if (inst->app_io->open(connection->child) < 0) {
                                DEBUG("proto_%s - Failed opening connected socket.", inst->app->common.name);
                                talloc_free(mi);
                                return NULL;
                        }
 
                        fd = connection->child->fd;
 
                        if (connect(fd, (struct sockaddr *) &src, salen) < 0) {
                                ERROR("proto_%s - Failed in connect: %s", inst->app->common.name, fr_syserror(errno));
                                goto cleanup;
                        }
                } else {
                        connection->child->fd = fd;
                }
 
                /*
                 *      Set the new FD, and get the module to set it's connection name.
                 */
                if (inst->app_io->fd_set(connection->child, fd) < 0) {
                        DEBUG3("Failed setting FD to %s", inst->app_io->common.name);
                        goto cleanup;
                }
 
                li->fd = fd;
 
                if (!inst->app_io->get_name) {
                        connection->name = fr_asprintf(connection, "proto_%s from client %pV port "
                                                       "%u to server %pV port %u",
                                                       inst->app->common.name,
                                                       fr_box_ipaddr(connection->address->socket.inet.src_ipaddr),
                                                       connection->address->socket.inet.src_port,
                                                       fr_box_ipaddr(connection->address->socket.inet.dst_ipaddr),
                                                       connection->address->socket.inet.dst_port);
                } else {
                        connection->name = inst->app_io->get_name(connection->child);
                }
 
                /*
                 *      Set the names for the listeners.
                 */
                connection->listen->name = connection->name;
                connection->child->name = connection->name;
        }
 
        /*
         *      If the parent client is still PENDING, lazily create the hash table that tracks its child
         *      connections.  We need this so that fr_io_connection_allow() can later find any sibling
         *      connections, and add them to the scheduler once the dynamic client is defined.
         */
        if ((client->state == PR_CLIENT_PENDING) && !client->ht) {
                (void) pthread_mutex_init(&client->mutex, NULL);
                MEM(client->ht = fr_hash_table_alloc(client, connection_hash, connection_cmp, NULL));
        }
 
        /*
         *      Add the connection to the set of connections for this client.  Capture the pre-insert size so
         *      we can tell whether this is the first connection (which runs the dynamic-client verification)
         *      or a later connection (which is deferred until the first connection is verified).
         */
        pthread_mutex_lock(&client->mutex);
        if (client->ht) {
                size_t pre_size = fr_hash_table_num_elements(client->ht);
 
                ret = fr_hash_table_insert(client->ht, connection);
                client->ready_to_delete = false;
                connection->in_parent_hash = true;
 
                if (ret != 0) {
                        pthread_mutex_unlock(&client->mutex);
                        ERROR("proto_%s - Failed inserting connection into tracking table.  "
                              "Closing it, and discarding all packets for connection %s.",
                              inst->app_io->common.name, connection->name);
                        goto cleanup;
                }
 
                /*
                 *      The first connection for a PENDING parent runs the dynamic client definition.  All
                 *      later connections will be scheduled by fr_io_connection_allow(), once the parent is
                 *      defined (or not).
                 */
                if ((client->state == PR_CLIENT_PENDING) && (pre_size > 0)) {
                        pthread_mutex_unlock(&client->mutex);
                        DEBUG("proto_%s - deferring scheduling of connection %s until parent client %pV is defined",
                              inst->app_io->common.name, connection->name, fr_box_ipaddr(client->src_ipaddr));
                        thread->num_connections++;
                        return connection;
                }
        }
        pthread_mutex_unlock(&client->mutex);
 
        DEBUG("proto_%s - starting connection %s", inst->app_io->common.name, connection->name);
        connection->nr = fr_schedule_listen_add(thread->sc, connection->listen);
        if (!connection->nr) {
                ERROR("proto_%s - Failed inserting connection into scheduler.  "
                      "Closing it, and diuscarding all packets for connection %s.",
                      inst->app_io->common.name, connection->name);
                pthread_mutex_lock(&client->mutex);
                if (client->ht) (void) fr_hash_table_delete(client->ht, connection);
                pthread_mutex_unlock(&client->mutex);
 
        cleanup:
                if (fd >= 0) close(fd);
                talloc_free(mi);
                return NULL;
        }
 
        /*
         *      We have one more connection.  Note that we do
         *      NOT decrement this counter when a connection
         *      closes, as the close is done in a child
         *      thread.  Instead, we just let counter hit the
         *      limit, and then walk over the clients to reset
         *      the count.
         */
        thread->num_connections++;
 
        return connection;
}
 
 
/*
 *      And here we go into the rabbit hole...
 *
 *      @todo future - have a similar structure
 *      fr_io_connection_io, which will duplicate some code,
 *      but may make things simpler?
 */
static void get_inst(fr_listen_t *li, fr_io_instance_t const **inst, fr_io_thread_t **thread,
                     fr_io_connection_t **connection, fr_listen_t **child)
{
        if (!li->connected) {
                *inst = li->app_io_instance;
                if (thread) *thread = li->thread_instance;
                *connection = NULL;
                if (child) *child = ((fr_io_thread_t *)li->thread_instance)->child;
 
        } else {
                fr_assert(connection != NULL);
 
                *connection = li->thread_instance;
                *inst = (*connection)->client->inst;
                if (thread) *thread = NULL;
                if (child) *child = (*connection)->child;
        }
}
 
 
static fr_client_t *radclient_alloc(TALLOC_CTX *ctx, int ipproto, fr_io_address_t *address)
{
        fr_client_t     *radclient;
        char            *shortname;
 
        MEM(radclient = talloc_zero(ctx, fr_client_t));
 
        fr_value_box_aprint(radclient, &shortname, fr_box_ipaddr(address->socket.inet.src_ipaddr), NULL);
        radclient->longname = radclient->shortname = shortname;
 
        radclient->secret = radclient->nas_type = talloc_strdup(radclient, "");
 
        radclient->ipaddr = address->socket.inet.src_ipaddr;
 
        radclient->src_ipaddr = address->socket.inet.dst_ipaddr;
 
        radclient->proto = ipproto;
        radclient->dynamic = true;
 
        return radclient;
}
 
/*
 *      Remove a client from the list of "live" clients.
 *
 *      This function is only used for the "main" socket.  Clients
 *      from connections do not use it.
 */
static int _client_live_free(fr_io_client_t *client)
{
        talloc_get_type_abort(client, fr_io_client_t);
 
        fr_assert(client->in_trie);
        fr_assert(!client->connection);
        fr_assert(client->thread);
 
        if (client->pending) client_pending_free(client);
 
        (void) fr_trie_remove_by_key(client->thread->trie, &client->src_ipaddr.addr, client->src_ipaddr.prefix);
 
        /*
         *      A client with pending packets is also in the thread's heap of pending clients.  Nothing
         *      else takes it out of that heap, so do it here, or the heap is left holding a pointer to
         *      freed memory.
         */
        if (client->thread->pending_clients && fr_heap_entry_inserted(client->pending_id)) {
                (void) fr_heap_extract(&client->thread->pending_clients, client);
        }
 
        if (client->thread->alive_clients) {
                fr_assert(fr_heap_num_elements(client->thread->alive_clients) > 0);
                (void) fr_heap_extract(&client->thread->alive_clients, client);
        }
 
        /*
         *      The mutex/ht pair is initialized either for use_connected
         *      clients (in client_alloc) or for PENDING clients with
         *      deferred sibling connections (in fr_io_connection_alloc).
         *      The ht pointer is the canonical signal that the mutex
         *      was initialized.
         */
        if (client->ht) (void) pthread_mutex_destroy(&client->mutex);
 
        return 0;
}
 
/** Allocate a dynamic client.
 *
 */
static fr_io_client_t *client_alloc(TALLOC_CTX *ctx, fr_io_client_state_t state,
                                    fr_io_instance_t const *inst, fr_io_thread_t *thread, fr_client_t *radclient,
                                    fr_ipaddr_t const *network)
{
        fr_io_client_t *client;
 
        /*
         *      Create our own local client.  This client
         *      holds our state which really shouldn't go into
         *      fr_client_t.
         *
         *      Note that we create a new top-level talloc
         *      context for this client, as there may be tens
         *      of thousands of packets associated with this
         *      client.  And we want to avoid problems with
         *      O(N) issues in talloc.
         */
        MEM(client = talloc_named(ctx, sizeof(fr_io_client_t), "fr_io_client_t"));
        memset(client, 0, sizeof(*client));
 
        client->state = state;
        client->src_ipaddr = radclient->ipaddr;
        client->radclient = radclient;
        client->inst = inst;
        client->thread = thread;
 
        if (network) {
                client->network = *network;
        } else {
                client->network = client->src_ipaddr;
        }
 
        /*
         *      At this point, this variable can only be true
         *      for STATIC clients.  PENDING clients may set
         *      it to true later, after they've been defined.
         */
        client->use_connected = radclient->use_connected;
 
        /*
         *      Create the pending heap for pending clients.
         */
        if (state == PR_CLIENT_PENDING) {
                MEM(client->pending = fr_heap_alloc(client, pending_packet_cmp,
                                                    fr_io_pending_packet_t, heap_id, 0));
        }
 
        /*
         *      Create the packet tracking table for this client.
         */
        if (inst->app_io->track_duplicates) {
                fr_assert(inst->app_io->track_compare != NULL);
                MEM(client->table = fr_rb_inline_talloc_alloc(client, fr_io_track_t, node, track_cmp, NULL));
        }
 
        /*
         *      Allow connected sockets to be set on a
         *      per-client basis.
         */
        if (client->use_connected) {
                fr_assert(client->state == PR_CLIENT_STATIC);
 
                (void) pthread_mutex_init(&client->mutex, NULL);
                MEM(client->ht = fr_hash_table_alloc(client, connection_hash, connection_cmp, NULL));
        }
 
        /*
         *      Add the newly defined client to the trie of
         *      allowed clients.
         */
        if (fr_trie_insert_by_key(thread->trie, &client->src_ipaddr.addr, client->src_ipaddr.prefix, client)) {
                ERROR("proto_%s - Failed inserting client %s into tracking table.  Discarding client, and all packets for it.",
                      inst->app_io->common.name, client->radclient->shortname);
                if (client->ht) (void) pthread_mutex_destroy(&client->mutex);
                talloc_free(client);
                return NULL;
        }
 
        client->in_trie = true;
 
        /*
         *      It's a static client.  Don't insert it into the list of alive clients, as those are only for
         *      dynamic clients.
         */
        if (state == PR_CLIENT_STATIC) return client;
 
        fr_assert(thread->alive_clients != NULL);
 
        /*
         *      Track the live clients so that we can clean
         *      them up.
         */
        (void) fr_heap_insert(&thread->alive_clients, client);
        client->pending_id = FR_HEAP_INDEX_INVALID;
 
        /*
         *      Now that we've inserted it into the heap and
         *      incremented the numbers, set the destructor
         *      function.
         */
        talloc_set_destructor(client, _client_live_free);
 
        return client;
}
 
 
static fr_io_track_t *fr_io_track_add(fr_listen_t const *li, fr_io_client_t *client,
                                      fr_io_address_t *address,
                                      uint8_t const *packet, size_t packet_len,
                                      fr_time_t recv_time, bool *is_dup)
{
        size_t len;
        fr_io_track_t *track, *old;
        TALLOC_CTX *track_ctx = client->inst->app_io->track_duplicates ? (TALLOC_CTX *)client->table : (TALLOC_CTX *)client;
 
        *is_dup = false;
 
        /*
         *      Allocate a new tracking structure.  Most of the time
         *      there are no duplicates, so this is fine.
         */
        if (client->connection) {
                MEM(track = talloc_zero_pooled_object(track_ctx, fr_io_track_t, 1, sizeof(*track) + 64));
                track->address = client->connection->address;
        } else {
                fr_io_address_t *my_address;
 
                MEM(track = talloc_zero_pooled_object(track_ctx, fr_io_track_t, 1, sizeof(*track) + sizeof(*track->address) + 64));
                MEM(track->address = my_address = talloc(track, fr_io_address_t));
 
                *my_address = *address;
                my_address->radclient = client->radclient;
        }
 
        track->li = li;
        track->client = client;
 
        track->timestamp = recv_time;
        track->packets = 1;
 
        /*
         *      We're not tracking duplicates, so just return the
         *      tracking entry.  This tracks src/dst IP/port, client,
         *      receive time, etc.
         */
        if (!client->inst->app_io->track_duplicates) {
                client->packets++;
                talloc_set_destructor(track, track_free);
                return track;
        }
 
        /*
         *      We are checking for duplicates, see if there is a dup
         *      already in the tree.
         */
        track->packet = client->inst->app_io->track_create(client->inst->app_io_instance,
                                                           client->thread->child->thread_instance,
                                                           client->radclient,
                                                           track, packet, packet_len);
        if (!track->packet) {
                talloc_free(track);
                return NULL;
        }
 
        /*
         *      No existing duplicate.  Return the new tracking entry.
         */
        fr_rb_find((void **)&old, client->table, track);
        if (!old) goto do_insert;
 
        fr_assert(old->client == client);
 
        /*
         *      It cannot be both in the free list and in the tracking table.
         *
         *      2020-08-17, this assertion fails randomly in travis.
         *      Which means that "track" was in the free list, *and*
         *      in the rbtree.
         */
        fr_assert(old != track);
 
        /*
         *      The new packet has the same dedup fields as the old
         *      one, BUT it may be a conflicting packet.  Check for
         *      that via a simple memcmp().
         *
         *      It's an exact duplicate.  Drop the new one and
         *      use the old one.
         *
         *      If there's a cached reply, the caller will take care
         *      of sending it to the network layer.
         */
        len = talloc_array_length(old->packet);
        if ((len == talloc_array_length(track->packet)) &&
            (memcmp(old->packet, track->packet, len) == 0)) {
                fr_assert(old != track);
 
                /*
                 *      Ignore duplicates while the client is
                 *      still pending.
                 */
                if (client->state == PR_CLIENT_PENDING) {
                        DEBUG("Ignoring duplicate packet while client %s is still pending dynamic definition",
                              client->radclient->shortname);
                        talloc_free(track);
                        return NULL;
                }
 
                *is_dup = true;
                old->packets++;
                talloc_free(track);
 
                /*
                 *      Retransmits can sit in the outbound queue for
                 *      a while.  We don't want to time out this
                 *      struct while the packet is in the outbound
                 *      queue.
                 */
                FR_TIMER_DISARM(old->ev);
                return old;
        }
 
        /*
         *      Else it's a conflicting packet.  Which is OK if we
         *      already have a reply.  We just delete the old entry,
         *      and insert the new one.
         *
         *      If there's no reply, then the old request is still
         *      "live".  Delete the old one from the tracking tree,
         *      and return the new one.
         */
        if (old->reply_len || old->do_not_respond) {
                talloc_free(old);
 
        } else {
                fr_assert(client == old->client);
 
                if (fr_rb_delete(client->table, old) != 0) {
                        fr_assert(0);
                }
                FR_TIMER_DELETE(&old->ev);
 
                talloc_set_destructor(old, track_free);
 
                old->discard = true; /* don't send any reply, there's nowhere for it to go */
        }
 
do_insert:
        if (fr_rb_insert(client->table, track) != 0) {
                fr_assert(0);
        }
 
        client->packets++;
        talloc_set_destructor(track, track_dedup_free);
        return track;
}
 
 
static int pending_free(fr_io_pending_packet_t *pending)
{
        fr_io_track_t *track = pending->track;
 
        /*
         *      Note that we don't check timestamps, replies, etc.  If
         *      a packet is pending, then any conflicting packet gets
         *      the "pending" entry marked as such, and a new entry
         *      added.  Any duplicate packet gets suppressed.  And
         *      because the packets are pending, track->reply MUST be
         *      NULL.
         */
        fr_assert(track->packets > 0);
        track->packets--;
 
        /*
         *      No more packets using this tracking entry,
         *      delete it.
         */
        if (track->packets == 0) talloc_free(track);
 
        return 0;
}
 
static fr_io_pending_packet_t *fr_io_pending_alloc(fr_io_connection_t *connection, fr_io_client_t *client,
                                                   uint8_t const *buffer, size_t packet_len,
                                                   fr_io_track_t *track,
                                                   int priority)
{
        fr_io_pending_packet_t *pending;
 
        MEM(pending = talloc_zero(client->pending, fr_io_pending_packet_t));
 
        MEM(pending->buffer = talloc_memdup(pending, buffer, packet_len));
        pending->buffer_len = packet_len;
        pending->priority = priority;
        pending->track = track;
        pending->recv_time = track->timestamp; /* there can only be one */
 
        talloc_set_destructor(pending, pending_free);
 
        /*
         *      Insert the pending packet for this client.  If it
         *      fails, silently discard the packet.
         */
        if (fr_heap_insert(&client->pending, pending) < 0) {
                talloc_free(pending);
                return NULL;
        }
 
        /*
         *      We only track pending packets for the
         *      main socket.  For connected sockets,
         *      we pause the FD, so the number of
         *      pending packets will always be small.
         */
        if (!connection) client->thread->num_pending_packets++;
 
        return pending;
}
 
 
/*
 *      Order clients in the alive_clients heap, based on their IP
 *      address.
 *
 *      This function is only used for the "main" socket.  Clients
 *      from connections do not use it.
 */
static fr_cmp_ret_t alive_client_cmp(void const *one, void const *two)
{
        fr_io_client_t const *a = talloc_get_type_abort_const(one, fr_io_client_t);
        fr_io_client_t const *b = talloc_get_type_abort_const(two, fr_io_client_t);
 
        return fr_ipaddr_cmp(&a->src_ipaddr, &b->src_ipaddr);
}
 
/**  Implement 99% of the read routines.
 *
 *  The app_io->read does the transport-specific data read.
 */
static ssize_t mod_read(fr_listen_t *li, void **packet_ctx, fr_time_t *recv_time_p,
                        uint8_t *buffer, size_t buffer_len, size_t *leftover)
{
        fr_io_instance_t const  *inst;
        fr_io_thread_t          *thread;
        ssize_t                 packet_len = -1;
        fr_time_t               recv_time = fr_time_wrap(0);
        fr_io_client_t          *client;
        fr_io_address_t         address;
        fr_io_connection_t      my_connection, *connection;
        fr_io_pending_packet_t  *pending = NULL;
        fr_io_track_t           *track;
        fr_listen_t             *child;
        int                     value, accept_fd = -1;
        uint32_t                priority = PRIORITY_NORMAL;
 
/** Log that we ignore clients in debug mode, or when it's enabled for a listener
 */
#define LOG_IGNORED_CLIENTS(_inst) ((_inst)->log_ignored_clients || fr_debug_lvl >= 1)
 
        get_inst(li, &inst, &thread, &connection, &child);
 
        track = NULL;
 
        /*
         *      There was data left over from the previous read, go
         *      get the rest of it now.  We MUST do this instead of
         *      popping a pending packet, because the leftover bytes
         *      are already in the output buffer.
         */
        if (*leftover) goto do_read;
 
redo:
        /*
         *      Read one pending packet.  The packet may be pending
         *      because of dynamic client definitions, or because it's
         *      for a connected UDP socket, and was sent over by the
         *      "master" UDP socket.
         */
        if (connection) {
                /*
                 *      The connection is dead.  Tell the network side
                 *      to close it.
                 */
                if (connection->dead) {
                        DEBUG("Dead connection %s", connection->name);
                        return -1;
                }
 
                fr_heap_pop((void **)&pending, &connection->client->pending);
 
        } else if (thread->pending_clients) {
                pending = pending_packet_pop(thread);
 
        } else {
                pending = NULL;
        }
 
        if (pending) {
                fr_assert(buffer_len >= pending->buffer_len);
                track = pending->track;
 
                /*
                 *      Clear the destructor as we now own the
                 *      tracking entry.
                 */
                talloc_set_destructor(pending, NULL);
 
                /*
                 *      We received a conflicting packet while this
                 *      packet was pending.  Discard this entry and
                 *      try to get another one.
                 *
                 *      Note that the pending heap is *simple*.  We
                 *      just track priority and recv_time.  This means
                 *      it's fast, but also that it's hard to look up
                 *      random packets in the pending heap.
                 */
                if (fr_time_neq(pending->recv_time, track->timestamp)) {
                        DEBUG3("Discarding old packet");
                        TALLOC_FREE(pending);
                        goto redo;
                }
 
                /*
                 *      We have a valid packet.  Copy it over to the
                 *      caller, and return.
                 */
                *packet_ctx = track;
                *leftover = 0;
                recv_time = *recv_time_p = pending->recv_time;
                client = track->client;
 
                memcpy(buffer, pending->buffer, pending->buffer_len);
                packet_len = pending->buffer_len;
 
                /*
                 *      Shouldn't be necessary, but what the heck...
                 */
                memcpy(&address, track->address, sizeof(address));
                TALLOC_FREE(pending);
 
                /*
                 *      Skip over all kinds of logic to find /
                 *      allocate the client, when we don't need to do
                 *      it any more.
                 */
                goto have_client;
 
        } else if (!connection && (inst->ipproto == IPPROTO_TCP)) {
                struct sockaddr_storage saremote;
                socklen_t salen;
 
                salen = sizeof(saremote);
 
                /*
                 *      We're a TCP socket but are NOT connected.  We
                 *      must be the master socket.  Accept the new
                 *      connection, and figure out src/dst IP/port.
                 */
                accept_fd = accept(child->fd,
                                   (struct sockaddr *) &saremote, &salen);
 
                /*
                 *      Couldn't open a NEW socket, but THIS ONE is
                 *      OK.  So don't return -1.
                 */
                if (accept_fd < 0) {
                        RATE_LIMIT_LOCAL(&thread->rate_limit.accept_failed,
                                         INFO, "proto_%s - failed to accept new socket: %s",
                                         inst->app->common.name, fr_syserror(errno));
                        return 0;
                }
 
#ifdef STATIC_ANALYZER
                saremote.ss_family = AF_INET; /* static analyzer doesn't know that accept() initializes this */
#endif
 
                /*
                 *      Get IP addresses only if we have IP addresses.
                 */
                if ((saremote.ss_family == AF_INET) || (saremote.ss_family == AF_INET6)) {
                        memset(&address.socket, 0, sizeof(address.socket));
                        (void) fr_ipaddr_from_sockaddr(&address.socket.inet.src_ipaddr, &address.socket.inet.src_port,
                                                       &saremote, salen);
 
                        /*
                         *      If the IP is in the NAK cache, then ignore the new socket.
                         *
                         *      An entry which has passed its expiry time is removed here, and the
                         *      connection is accepted as normal.
                         */
                        if (fr_io_nak_find(thread, &address.socket.inet.src_ipaddr)) {
                                RATE_LIMIT_LOCAL(&thread->rate_limit.accept_failed,
                                                 INFO, "proto_%s - ignoring connection request from address %pV due to NAK cache",
                                                 inst->app->common.name, fr_box_ipaddr(address.socket.inet.src_ipaddr));
                                close(accept_fd);
                                return 0;
                        }
 
                        /*
                         *      @todo - only if the local listen address is "*".
                         */
                        salen = sizeof(saremote);
                        (void) getsockname(accept_fd, (struct sockaddr *) &saremote, &salen);
                        (void) fr_ipaddr_from_sockaddr(&address.socket.inet.dst_ipaddr, &address.socket.inet.dst_port,
                                                       &saremote, salen);
                        address.socket.type = (inst->ipproto == IPPROTO_TCP) ? SOCK_STREAM : SOCK_DGRAM;
                        address.socket.fd = accept_fd;
                }
 
                /*
                 *      Set the new descriptor to be non-blocking.
                 */
                (void) fr_nonblock(accept_fd);
 
 
        } else {
                fr_io_address_t *local_address;
 
                /*
                 *      We're either not a TCP socket, or we are a
                 *      connected TCP socket.  Just read it.
                 */
do_read:
                local_address = &address;
 
                /*
                 *      @todo - For connected TCP sockets which are
                 *      dynamically defined, the app_io read()
                 *      function should stop reading the socket if the
                 *      server is busy.  That change puts TCP
                 *      backpressure on the client.
                 *
                 *      @todo TLS - for TLS and dynamic sockets, do
                 *      the SSL setup here, but have a structure which
                 *      describes the TLS data and run THAT through
                 *      the dynamic client definition, instead of
                 *      using normal packets.  Or, rely on the app_io
                 *      read() function to do all TLS work?  Given
                 *      that some protocols have "starttls" beginning
                 *      after a clear-text exchange, it's likely best
                 *      to have yet another layer of trampoline
                 *      functions which do all of the TLS work.
                 */
                packet_len = inst->app_io->read(child, (void **) &local_address, &recv_time,
                                          buffer, buffer_len, leftover);
                if (packet_len <= 0) {
                        return packet_len;
                }
 
                /*
                 *      Not allowed?  Discard it.  The priority()
                 *      function has done any complaining, if
                 *      necessary.
                 */
                if (inst->app->priority) {
                        value = inst->app->priority(inst->app_instance, buffer, packet_len);
                        if (value <= 0) {
                                static fr_rate_limit_t bad_type;
 
                                /*
                                 *      @todo - unix sockets.  We need to use
                                 *      the "name" of the socket, in the
                                 *      listener?
                                 */
                                if (LOG_IGNORED_CLIENTS(inst)) {
                                        RATE_LIMIT_LOCAL(thread ? &thread->rate_limit.bad_type : &bad_type, INFO,
                                                         "proto_%s - ignoring packet from IP %pV. It is not configured as 'type = ...'",
                                                         inst->app_io->common.name, fr_box_ipaddr(address.socket.inet.src_ipaddr));
                                }
                                return 0;
                        }
                        priority = value;
                }
 
                /*
                 *      If the connection is pending, pause reading of
                 *      more packets.  If mod_write() accepts the
                 *      connection, it will resume reading.
                 *      Otherwise, it will close the socket without
                 *      resuming it.
                 */
                if (connection &&
                    (connection->client->state == PR_CLIENT_PENDING)) {
                        fr_assert(!connection->paused);
 
                        connection->paused = true;
                        (void) fr_event_filter_update(connection->el,
                                                      child->fd,
                                                      FR_EVENT_FILTER_IO, pause_read);
                }
        }
 
        /*
         *      Look up the client, unless we already have one (for a
         *      connected socket).
         */
        if (!connection) {
                client = fr_trie_lookup_by_key(thread->trie,
                                        &address.socket.inet.src_ipaddr.addr, address.socket.inet.src_ipaddr.prefix);
                fr_assert(!client || !client->connection);
 
                /*
                 *      Verify the cached client is the most specific match.
                 *      A broader subnet may have been cached first, shadowing
                 *      a more specific client definition.
                 */
                if (client && (client->state == PR_CLIENT_STATIC)) {
                        fr_client_t *radclient;
 
                        radclient = inst->app_io->client_find(thread->child,
                                                              &address.socket.inet.src_ipaddr, inst->ipproto);
                        if (radclient && (radclient->ipaddr.prefix > client->src_ipaddr.prefix)) {
                                client = NULL;
                        }
                }
 
        } else {
                client = connection->client;
 
                /*
                 *      We don't care what the read function says
                 *      about address.  We have it already.
                 */
                address = *connection->address;
        }
 
        /*
         *      No client was found, and we don't have a connection.  Check the NAK cache before defining a
         *      dynamic client.
         *
         *      The NAK cache has already been checked on the TCP path above, after the accept().
         */
        if (!client && !connection &&
            fr_io_nak_find(thread, &address.socket.inet.src_ipaddr)) {
                RATE_LIMIT_LOCAL(&thread->rate_limit.repeat_nak, ERROR,
                                 "proto_%s - Discarding packet from NAK'd dynamic client %pV",
                                 inst->app_io->common.name, fr_box_ipaddr(address.socket.inet.src_ipaddr));
                if (accept_fd >= 0) close(accept_fd);
                return 0;
        }
 
        /*
         *      If there's no client, try to pull one from the global
         *      / static client list.  Or if dynamic clients are
         *      allowed, try to define a dynamic client.
         */
        if (!client) {
                fr_client_t *radclient = NULL;
                fr_io_client_state_t state;
                fr_ipaddr_t const *network = NULL;
                char const *error;
 
                /*
                 *      We MUST be the master socket.
                 */
                fr_assert(!connection);
 
                radclient = inst->app_io->client_find(thread->child, &address.socket.inet.src_ipaddr, inst->ipproto);
                if (radclient) {
                        state = PR_CLIENT_STATIC;
 
                        /*
                         *      Make our own copy that we can modify it.
                         */
                        MEM(radclient = radclient_clone(thread, radclient));
                        radclient->active = true;
 
                } else if (inst->dynamic_clients) {
                        if (inst->max_clients && (fr_heap_num_elements(thread->alive_clients) >= inst->max_clients)) {
                                error = "Too many dynamic clients have been defined";
                                goto ignore;
                        }
 
                        /*
                         *      Look up the allowed networks.
                         */
                        network = fr_trie_lookup_by_key(inst->networks, &address.socket.inet.src_ipaddr.addr,
                                                 address.socket.inet.src_ipaddr.prefix);
                        if (!network) {
                                error = "Address is outside of the the 'allow' network range";
                                goto ignore;
                        }
 
                        /*
                         *      It exists, but it's a "deny" rule, ignore it.
                         */
                        if (network->af == AF_UNSPEC) {
                                error = "Address is forbidden by the 'deny' network range";
                                goto ignore;
                        }
 
                        /*
                         *      Allocate our local radclient as a
                         *      placeholder for the dynamic client.
                         */
                        radclient = radclient_alloc(thread, inst->ipproto, &address);
                        state = PR_CLIENT_PENDING;
 
                } else {
                        char const *msg;
 
                        error = "No matching 'client' definition was found";
 
                ignore:
                        if (accept_fd < 0) {
                                msg = "packet";
                        } else {
                                msg = "connection attempt";
                                close(accept_fd);
                        }
 
                        if (LOG_IGNORED_CLIENTS(inst)) {
                                static fr_rate_limit_t unknown_client;
                                RATE_LIMIT_LOCAL(thread ? &thread->rate_limit.unknown_client : &unknown_client,
                                                 ERROR, "proto_%s - Ignoring %s from IP address %pV - %s",
                                                 inst->app_io->common.name, msg, fr_box_ipaddr(address.socket.inet.src_ipaddr),
                                                 error);
                        }
 
                        return 0;
                }
 
                MEM(client = client_alloc(thread, state, inst, thread, radclient, network));
 
                /*
                 *      Parent the dynamic client radclient off the client - it
                 *      is the client which gets freed by the dynamic client timers.
                 */
                if (state == PR_CLIENT_PENDING) talloc_steal(client, radclient);
        }
 
have_client:
        fr_assert(client->state != PR_CLIENT_INVALID);
 
        /*
         *      We've accepted a new connection.  Go allocate it, and
         *      let it read from the socket.
         */
        if (accept_fd >= 0) {
                connection = fr_io_connection_alloc(inst, thread, client, accept_fd, &address);
                if (!connection) {
                        static fr_rate_limit_t alloc_failed;
 
                        RATE_LIMIT_LOCAL(thread ? &thread->rate_limit.conn_alloc_failed : &alloc_failed,
                                         ERROR, "Failed to allocate connection from client %s", client->radclient->shortname);
                        return -1;
                }
 
                /*
                 *      The parent is in use - ensure the cleanup timer is disarmed.
                 */
                if (fr_timer_armed(connection->parent->ev)) {
                        FR_TIMER_DISARM_RETURN(connection->parent->ev);
                        connection->parent->ready_to_delete = false;
                }
 
                return 0;
        }
 
        /*
         *      No connected sockets, OR we are the connected socket.
         *
         *      Track this packet and return it if necessary.
         */
        if (connection || !client->use_connected) {
                fr_io_track_t *to_free = NULL;
 
                /*
                 *      Add the packet to the tracking table, if it's
                 *      not already there.  Pending packets will be in
                 *      the tracking table, but won't be counted as
                 *      "live" packets.
                 */
                if (!track) {
                        static  fr_rate_limit_t tracking_failed;
                        bool    is_dup = false;
 
                        track = fr_io_track_add(li, client, &address, buffer, packet_len, recv_time, &is_dup);
                        if (!track) {
                                RATE_LIMIT_LOCAL(thread ? &thread->rate_limit.tracking_failed : &tracking_failed,
                                                 ERROR, "Failed tracking packet from client %s - discarding it",
                                                 client->radclient->shortname);
                                return 0;
                        }
 
                        /*
                         *      If there's a cached reply, just send that and don't do anything else.
                         */
                        if (is_dup) {
                                fr_network_t *nr;
 
                                if (track->do_not_respond) {
                                        DEBUG("Ignoring retransmit from client %s - we are not responding to this request", client->radclient->shortname);
                                        return 0;
                                }
 
                                if (track->discard) {
                                        DEBUG("Ignoring transmit from client %s - we previously received a newer / conflicting packet", client->radclient->shortname);
                                        return 0;
                                }
 
                                if (!track->reply) {
                                        fr_assert(!track->finished);
                                        DEBUG("Ignoring retransmit from client %s - we are still processing the request", client->radclient->shortname);
                                        return 0;
                                }
 
                                if (connection) {
                                        nr = connection->nr;
                                } else {
                                        nr = thread->nr;
                                }
 
                                /*
                                 *      @todo - mark things up so that we know to keep 'track' around
                                 *      until the packet is actually written to the network.  OR, add
                                 *      a network API so that the talloc_free() function can remove
                                 *      the packet from the queue of packets to be retransmitted.
                                 *
                                 *      Perhaps via having fr_network_listen_write() return a pointer
                                 *      to the localized message, and then caching that in the tracking
                                 *      structure.
                                 */
                                DEBUG("Sending duplicate reply to client %s", client->radclient->shortname);
                                fr_network_listen_write(nr, li, track->reply, track->reply_len,
                                                        track, track->timestamp);
                                return 0;
                        }
 
                        /*
                         *      Got to free this if we don't process the packet.
                         */
                        to_free = track;
                }
 
                /*
                 *      This is a pending dynamic client.  See if we
                 *      have to either run the dynamic client code to
                 *      define the client, OR to push the packet onto
                 *      the pending queue for this client.
                 */
                if (client->state == PR_CLIENT_PENDING) {
                        /*
                         *      Track pending packets for the master
                         *      socket.  Connected sockets are paused
                         *      as soon as they are defined, so we
                         *      won't be reading any more packets from
                         *      them.
                         *
                         *      Since we don't have pending packets
                         *      for connected sockets, we don't need
                         *      to track pending packets.
                         */
                        if (!connection && inst->max_pending_packets && (thread->num_pending_packets >= inst->max_pending_packets)) {
                                RATE_LIMIT_LOCAL(&thread->rate_limit.too_many_pending,
                                                 ERROR, "Too many pending dynamic client packets for listener - discarding packet from %pV",
                                                 fr_box_ipaddr(client->src_ipaddr));
 
                        discard:
                                talloc_free(to_free);
                                return 0;
                        }
 
                        /*
                         *      Allocate the pending packet structure.
                         */
                        pending = fr_io_pending_alloc(connection, client, buffer, packet_len,
                                                      track, priority);
                        if (!pending) {
                                static fr_rate_limit_t alloc_failed;
                                RATE_LIMIT_LOCAL(thread ? &thread->rate_limit.alloc_failed : &alloc_failed,
                                                 ERROR, "proto_%s - Failed allocating space for dynamic client %pV - discarding packet",
                                                 inst->app_io->common.name, fr_box_ipaddr(client->src_ipaddr));
                                goto discard;
                        }
 
                        if (fr_heap_num_elements(client->pending) > 1) {
                                DEBUG("Verification is still pending for dynamic client %pV - queuing additional packet(s)",
                                      fr_box_ipaddr(client->src_ipaddr));
                                return 0;
                        }
 
                        /*
                         *      Tell this packet that it's defining a
                         *      dynamic client.
                         */
                        track->dynamic = recv_time;
 
                        INFO("proto_%s - Verification started for packet from dynamic client %pV - queuing new packets",
                             inst->app_io->common.name, fr_box_ipaddr(client->src_ipaddr));
                }
 
                /*
                 *      Remove all cleanup timers for the client /
                 *      connection.  It's still in use, so we don't
                 *      want to clean it up.
                 */
                if (fr_timer_armed(client->ev)) {
                        FR_TIMER_DISARM_RETURN(client->ev);
                        client->ready_to_delete = false;
                }
 
                /*
                 *      Remove cleanup timers for the connection parent.
                 */
                if (connection && fr_timer_armed(connection->parent->ev)) {
                        FR_TIMER_DISARM_RETURN(connection->parent->ev);
                        connection->parent->ready_to_delete = false;
                }
 
                /*
                 *      Return the packet.
                 */
                *recv_time_p = track->timestamp;
                *packet_ctx = track;
                return packet_len;
        }
 
        /*
         *
         */
        fr_assert(!pending);
 
        /*
         *      This must be the main UDP socket which creates
         *      connections.
         */
        fr_assert(inst->ipproto == IPPROTO_UDP);
 
        /*
         *      We're using connected sockets, but this socket isn't
         *      connected.  It must be the master socket.  The master
         *      can either be STATIC, DYNAMIC, or PENDING.  Whatever
         *      the state, the child socket will take care of handling
         *      the packet.  e.g. dynamic clients, etc.
         */
        my_connection.address = &address;
 
        pthread_mutex_lock(&client->mutex);
        fr_hash_table_find((void **)&connection, client->ht, &my_connection);
        pthread_mutex_unlock(&client->mutex);
 
        /*
         *      No existing connection, create one.
         */
        if (!connection) {
                connection = fr_io_connection_alloc(inst, thread, client, -1, &address);
                if (!connection) {
                        RATE_LIMIT_LOCAL(&thread->rate_limit.conn_alloc_failed,
                                         ERROR, "Failed to allocate connection from client %s.  Discarding packet.", client->radclient->shortname);
                        return 0;
                }
        }
 
        DEBUG("Sending packet to connection %s", connection->name);
 
        /*
         *      Inject the packet into the connected socket.  It will
         *      process the packet as if it came in from the network.
         *
         *      @todo future - after creating the connection, put the
         *      current packet into connection->pending, instead of
         *      inject?, and then call fr_network_listen_read() from
         *      the child's instantiation routine???
         *
         *      @todo TCP - for ACCEPT sockets, we don't have a
         *      packet, so don't do this.  Instead, the connection
         *      will take care of figuring out what to do.
         *
         *      We don't need "to_free" after this, as it will be
         *      tracked in the connected socket.
         */
        if (fr_network_listen_inject(connection->nr, connection->listen,
                                     buffer, packet_len, recv_time) < 0) {
                RATE_LIMIT_LOCAL(&thread->rate_limit.queue_full, PERROR,
                                 "proto_%s - Discarding packet from dynamic client %pV - cannot push packet to connected socket",
                                 inst->app_io->common.name, fr_box_ipaddr(address.socket.inet.src_ipaddr));
                /*
                 *      Don't return an error, because that will cause the listener to close its socket.
                 */
        }
 
        return 0;
}
 
/** Inject a packet to a connection.
 *
 *  Always called in the context of the network.
 */
static int mod_inject(fr_listen_t *li, uint8_t const *buffer, size_t buffer_len, fr_time_t recv_time)
{
        fr_io_instance_t const *inst;
        int             priority;
        bool            is_dup = false;
        fr_io_connection_t *connection;
        fr_io_pending_packet_t *pending;
        fr_io_track_t *track;
 
        get_inst(li, &inst, NULL, &connection, NULL);
 
        if (!connection) {
                DEBUG2("Received injected packet for an unconnected socket.");
                return -1;
        }
 
        if (inst->app->priority) {
                priority = inst->app->priority(inst->app_instance, buffer, buffer_len);
                if (priority <= 0) {
                        return -1;
                }
        } else {
                priority = PRIORITY_NORMAL;
        }
 
        /*
         *      Track this packet, because that's what mod_read expects.
         */
        track = fr_io_track_add(li, connection->client, connection->address,
                                buffer, buffer_len, recv_time, &is_dup);
        if (!track) {
                DEBUG2("Failed injecting packet to tracking table");
                return -1;
        }
 
        talloc_get_type_abort(track, fr_io_track_t);
 
        /*
         *      @todo future - what to do with duplicates?
         */
        fr_assert(!is_dup);
 
        /*
         *      Remember to restore this packet later.
         */
        pending = fr_io_pending_alloc(connection, connection->client, buffer, buffer_len,
                                      track, priority);
        if (!pending) {
                DEBUG2("Failed injecting packet due to allocation error");
                return -1;
        }
 
        return 0;
}
 
/** Open a new listener
 *
 */
static int mod_open(fr_listen_t *li)
{
        fr_io_thread_t *thread;
        fr_io_instance_t const *inst;
 
        thread = li->thread_instance;
        inst = li->app_io_instance;
 
        if (inst->app_io->open(thread->child) < 0) return -1;
 
        li->fd = thread->child->fd;     /* copy this back up */
 
        /*
         *      Set the name of the socket.
         */
        if (!li->app_io->get_name) {
                li->name = li->app_io->common.name;
        } else {
                li->name = li->app_io->get_name(li);
        }
 
        /*
         *      Note that we're opening a child socket, so we don't
         *      put it into the list of global listeners.
         */
 
        return 0;
}
 
 
/** Set the event list for a new socket
 *
 * @param[in] li the listener
 * @param[in] el the event list
 * @param[in] nr context from the network side
 */
static void mod_event_list_set(fr_listen_t *li, fr_event_list_t *el, void *nr)
{
        fr_io_instance_t const *inst;
        fr_io_connection_t *connection;
        fr_io_thread_t *thread;
        fr_listen_t *child;
 
        get_inst(li, &inst, &thread, &connection, &child);
 
        /*
         *      We're not doing IO, so there are no timers for
         *      cleaning up packets, dynamic clients, or connections.
         */
        if (!inst->submodule) return;
 
        if (inst->app_io->event_list_set) {
                inst->app_io->event_list_set(child, el, nr);
        }
 
        /*
         *      Set event list and network side for this socket.
         */
        if (!connection) {
                thread->el = el;
                thread->nr = nr;
 
        } else {
                connection->el = el;
                connection->nr = nr;
        }
}
 
 
/** Delete a client and any connection it belongs to.
 *
 *  A connected socket owns a client, so we have to remove the connection from the parent "accept" connection.
 *  We also mark the connection as dead, and tell the network side to read from the connection, which then
 *  closes it.
 */
static void fr_io_client_delete(fr_io_client_t *client)
{
        fr_io_connection_t *connection = client->connection;
 
        fr_assert(client->packets == 0);
 
        if (connection) {
                pthread_mutex_lock(&connection->parent->mutex);
                if (connection->in_parent_hash) {
                        connection->in_parent_hash = false;
                        (void) fr_hash_table_delete(connection->parent->ht, connection);
                }
                pthread_mutex_unlock(&connection->parent->mutex);
 
                connection->dead = true;
                fr_network_listen_read(connection->nr, connection->listen);
                return;
        }
 
        talloc_free(client);
}
 
static void client_expiry_timer(fr_timer_list_t *tl, fr_time_t now, void *uctx)
{
        fr_io_client_t          *client = talloc_get_type_abort(uctx, fr_io_client_t);
        fr_io_instance_t const  *inst;
        fr_io_connection_t      *connection;
        fr_time_delta_t         delay;
        int                     connections;
 
        /*
         *      No event list?  We don't need to expire the client.
         */
        if (!tl) return;
 
        // @todo - print out what we plan on doing next
        connection = client->connection;
        inst = client->inst;
 
        fr_assert(client->state != PR_CLIENT_STATIC);
 
        /*
         *      Called from the read or write functions with
         *      now==0, to signal that we have to *set* the timer.
         */
        if (fr_time_eq(now, fr_time_wrap(0))) {
                /*
                 *      The timer is already set, don't do anything.
                 */
                if (fr_timer_armed(client->ev)) return;
 
                switch (client->state) {
                case PR_CLIENT_CONNECTED:
                        fr_assert(connection != NULL);
                        delay = inst->idle_timeout;
                        if (fr_time_delta_ispos(client->radclient->limit.idle_timeout) &&
                            (fr_time_delta_lt(client->radclient->limit.idle_timeout, inst->idle_timeout))) {
                                delay = client->radclient->limit.idle_timeout;
                        }
                        break;
 
                case PR_CLIENT_DYNAMIC:
                        delay = inst->dynamic_timeout;
                        break;
 
                default:
                        fr_assert(0 == 1);
                        return;
                }
 
                DEBUG("TIMER - setting idle timeout to %pVs for connection from client %s", fr_box_time_delta(delay), client->radclient->shortname);
 
                goto reset_timer;
        }
 
        DEBUG2("TIMER - checking status of dynamic client %s %pV", client->radclient->shortname, fr_box_ipaddr(client->src_ipaddr));
 
        /*
         *      It's a dynamically defined client.  If no one is using
         *      it, clean it up after an idle timeout.
         */
        if ((client->state == PR_CLIENT_DYNAMIC) ||
            (client->state == PR_CLIENT_CONNECTED)) {
                if (client->packets > 0) {
                        client->ready_to_delete = false;
                        return;
                }
 
                /*
                 *      No packets, check / set idle timeout.
                 */
                goto idle_timeout;
        }
 
        /*
         *      The client is pending definition.  It's either a
         *      dynamic client which has timed out, OR it's a
         *      "place-holder" client for connected sockets.
         */
        fr_assert(client->state == PR_CLIENT_PENDING);
 
        /*
         *      This is a dynamic client pending definition.
         *      But it's taken too long to define, so we just
         *      delete the client, and all packets for it.  A
         *      new packet will cause the dynamic definition
         *      to be run again.
         */
        if (!client->use_connected) {
                if (!client->packets) {
                        DEBUG("proto_%s - No packets are using unconnected socket", inst->app_io->common.name);
                        fr_io_client_delete(client);
                        return;
                }
 
                /*
                 *      Tell the writer to NOT dynamically define the
                 *      client.  We've run into a problem.  Then,
                 *      return.  The writer will take care of calling
                 *      us again when it notices that a PENDING client
                 *      is ready to delete.
                 *
                 *      TBH... that shouldn't happen?  We should rely
                 *      on the write to do this all of the time...
                 */
                client->ready_to_delete = true;
                return;
        }
 
        fr_assert(!connection);
 
        /*
         *      Find out how many connections are using this
         *      client.
         */
        pthread_mutex_lock(&client->mutex);
        fr_assert(client->ht != NULL);
        connections = fr_hash_table_num_elements(client->ht);
        pthread_mutex_unlock(&client->mutex);
 
        /*
         *      No connections are using this client.  If
         *      we've passed the idle timeout, then just
         *      delete it.  Otherwise, set an idle timeout (as
         *      above);
         */
        if (!connections) {
idle_timeout:
                /*
                 *      We didn't receive any packets during the
                 *      idle_timeout, just delete it.
                 */
                if (client->ready_to_delete) {
                        if (connection) {
                                DEBUG("proto_%s - idle timeout for connection %s", inst->app_io->common.name, connection->name);
                        } else {
                                DEBUG("proto_%s - idle timeout for client %s", inst->app_io->common.name, client->radclient->shortname);
                        }
                        fr_io_client_delete(client);
                        return;
                }
 
                /*
                 *      No packets and no idle timeout set, go set
                 *      idle timeut.
                 */
                client->ready_to_delete = true;
                delay = client->state == PR_CLIENT_DYNAMIC ? inst->dynamic_timeout : inst->idle_timeout;
                goto reset_timer;
        }
 
        /*
         *      There are live sub-connections.  Poll again after a
         *      long period of time.  Once all of the connections are
         *      closed, we can then delete this client.
         *
         *      @todo - maybe just leave it?  we want to be able to
         *      clean up this client after a while tho... especially
         *      if the total number of clients is limited.
         */
        client->ready_to_delete = false;
        delay = inst->check_interval;
 
reset_timer:
        if (fr_timer_in(client, tl, &client->ev,
                        delay, false, client_expiry_timer, client) < 0) {
                ERROR("proto_%s - Failed adding timeout for dynamic client %s.  It will be permanent!",
                      inst->app_io->common.name, client->radclient->shortname);
                return;
        }
 
        return;
}
 
 
/*
 *      Expire cached packets after cleanup_delay time
 */
static void packet_expiry_timer(fr_timer_list_t *tl, fr_time_t now, void *uctx)
{
        fr_io_track_t *track = talloc_get_type_abort(uctx, fr_io_track_t);
        fr_io_client_t *client = track->client;
        fr_io_instance_t const *inst = client->inst;
 
        /*
         *      Insert the timer if requested.
         *
         *      On duplicates this also extends the expiry timer.
         */
        if (fr_time_eq(now, fr_time_wrap(0)) && !track->discard && inst->app_io->track_duplicates) {
                fr_assert(fr_time_delta_ispos(inst->cleanup_delay));
                fr_assert(track->do_not_respond || track->reply_len);
 
                track->expires = fr_time_add(fr_time(), inst->cleanup_delay);
 
                /*
                 *      if the timer succeeds, then "track"
                 *      will be cleaned up when the timer
                 *      fires.
                 */
                if (fr_timer_at(track, tl, &track->ev,
                                track->expires,
                                false, packet_expiry_timer, track) == 0) {
                        DEBUG("proto_%s - cleaning up request in %.6fs", inst->app_io->common.name,
                              fr_time_delta_unwrap(inst->cleanup_delay) / (double)NSEC);
                        return;
                }
 
                DEBUG("proto_%s - Failed adding cleanup_delay for packet.  Discarding packet immediately",
                      inst->app_io->common.name);
        }
 
        /*
         *      So that all cleanup paths can come here, not just the
         *      timeout ones.
         */
        if (fr_time_neq(now, fr_time_wrap(0))) {
                DEBUG2("TIMER - proto_%s - cleanup delay", inst->app_io->common.name);
        } else {
                DEBUG2("proto_%s - cleaning up", inst->app_io->common.name);
        }
 
        /*
         *      Delete the tracking entry.
         */
        talloc_free(track);
 
        /*
         *      The client isn't dynamic, stop here.
         */
        if (client->state == PR_CLIENT_STATIC) return;
 
        fr_assert(client->state != PR_CLIENT_PENDING);
 
        /*
         *      If necessary, call the client expiry timer to clean up
         *      the client.
         */
        if (client->packets == 0) {
                client_expiry_timer(tl, now, client);
        }
}
 
static void update_client(fr_io_client_t *client, fr_client_t *radclient)
{
 
        /*
         *      The new client is mostly OK.  Copy the various fields
         *      over.
         */
#define COPY_FIELD(_dest, _x) _dest->radclient->_x = radclient->_x
#define DUP_FIELD(_dest, _x) _dest->radclient->_x = talloc_strdup(_dest->radclient, radclient->_x)
 
        /*
         *      Only these two fields are set.  Other strings in
         *      radclient are copies of these ones.
         */
        talloc_const_free(client->radclient->shortname);
        talloc_const_free(client->radclient->secret);
 
        DUP_FIELD(client, longname);
        DUP_FIELD(client, shortname);
        DUP_FIELD(client, secret);
        DUP_FIELD(client, nas_type);
 
        COPY_FIELD(client, ipaddr);
        COPY_FIELD(client, src_ipaddr);
        COPY_FIELD(client, require_message_authenticator);
        COPY_FIELD(client, require_message_authenticator_is_set);
#ifdef NAS_VIOLATES_RFC
        COPY_FIELD(client, allow_vulnerable_clients);
#endif
        COPY_FIELD(client, limit_proxy_state);
        COPY_FIELD(client, limit_proxy_state_is_set);
        COPY_FIELD(client, use_connected);
        COPY_FIELD(client, cs);
}
 
/** Tear down any deferred sibling connections under a pending parent.
 *
 *  This function is Cclled from mod_write() when the first connection dynamic-client verification fails (NAK
 *  or hard reject).  Any other connections that were inserted into parent->ht are then freed.
 *
 *  Walk parent->ht, find every deferred sibling, remove it from the hash table, close the socket, and frees
 *  its module instance (which cascades to the connection itself).  The first connection (the one whose
 *  verification just failed) is identified by conn->nr != NULL and is left alone, as the caller already owns
 *  its cleanup.
 *
 *  Uses a find-one-then-restart loop rather than iterate-while-deleting, since the deferred sibling count is
 *  small and the cost is irrelevant on the failure path.
 */
static void fr_io_connection_deny(fr_io_client_t *parent)
{
        if (!parent->ht) return;
 
        while (true) {
                fr_hash_iter_t          iter;
                fr_io_connection_t      *target = NULL;
                fr_io_connection_t      *conn;
 
                pthread_mutex_lock(&parent->mutex);
                for (conn = fr_hash_table_iter_init(parent->ht, &iter);
                     conn != NULL;
                     conn = fr_hash_table_iter_next(parent->ht, &iter)) {
                        if (conn->nr != NULL) continue;                         /* first child, owned by caller */
                        if (conn->client->state != PR_CLIENT_PENDING) continue; /* already NAK or promoted */
                        target = conn;
                        break;
                }
                if (target && target->in_parent_hash) {
                        target->in_parent_hash = false;
                        (void) fr_hash_table_delete(parent->ht, target);
                }
                pthread_mutex_unlock(&parent->mutex);
 
                if (!target) break;
 
                DEBUG("proto_%s - cleaning up deferred connection %s after verification failure",
                      target->client->inst->app_io->common.name, target->name);
 
                if (target->child) {
                        if (target->client->inst->app_io->close) {
                                (void) target->client->inst->app_io->close(target->child);
                        } else if (target->child->fd >= 0) {
                                close(target->child->fd);
                        }
                }
 
                talloc_free(target->mi);
        }
}
 
/** Promote a pending dynamic-client parent and all of its child connections.
 *
 *  When multiple TCP connections from the same source IP arrive while the parent client is still
 *  PR_CLIENT_PENDING, only the first connection is added to the scheduler.  Later connection are inserted
 *  into the parent's hash table, but their fr_schedule_listen_add() call is deferred to this function.  When
 *  the first connection verification completes successfully, every other connection is finished and
 *  scheduled.
 *
 *  This function:
 *
 *  * Promotes the parent itself to PR_CLIENT_DYNAMIC and re-parents the cs.
 *
 *  * Walks the parent's hash table of connections.  For the first connection (already in the scheduler with
 *    connection->nr set), it resumes reads if the connection was paused.  For every later connection,
 *    (connection->nr == NULL), it copies the verified radclient definition onto the connection and then calls
 *    fr_schedule_listen_add() to add the connection to the scheduler.
 *
 *  @param parent     the parent client whose state and child connections to promote
 *  @param radclient  the verified radclient (typically the calling child's radclient)
 */
static void fr_io_connection_allow(fr_io_client_t *parent, fr_client_t *radclient)
{
        fr_hash_iter_t          iter;
        fr_io_connection_t      *conn;
        fr_schedule_t           *sc = parent->thread->sc;
 
        /*
         *      Promote the parent itself.  Only the first connection to reach here does the work; later
         *      connections will already see the parent as PR_CLIENT_DYNAMIC.
         */
        if (parent->state == PR_CLIENT_PENDING) {
                parent->radclient->active = true;
                parent->state = PR_CLIENT_DYNAMIC;
 
                update_client(parent, radclient);
 
                /*
                 *      Re-parent the conf section used to build this
                 *      client so its lifetime is linked to the parent
                 *      client.
                 */
                talloc_steal(parent->radclient, parent->radclient->cs);
        } else {
                fr_assert(parent->state == PR_CLIENT_DYNAMIC);
        }
 
        /*
         *      Walk every child connection of this parent and promote the pending ones.  The calling
         *      connection is itself in parent->ht, so its per-child work is also done here.
         */
        pthread_mutex_lock(&parent->mutex);
        if (parent->ht) {
                for (conn = fr_hash_table_iter_init(parent->ht, &iter);
                     conn != NULL;
                     conn = fr_hash_table_iter_next(parent->ht, &iter)) {
                        fr_io_client_t *child = conn->client;
 
                        if (child->state != PR_CLIENT_PENDING) continue;
 
                        /*
                         *      Connections can't spawn new connections.
                         */
                        child->use_connected = child->radclient->use_connected = false;
 
                        if (conn->nr == NULL) {
                                /*
                                 *      Deferred connection: its radclient was cloned from the parent's
                                 *      placeholder before verification, so copy the now-verified fields onto
                                 *      it before adding it to the scheduler.
                                 */
                                update_client(child, radclient);
 
                                child->state = PR_CLIENT_DYNAMIC;
                                child->radclient->active = true;
 
                                DEBUG("proto_%s - scheduling deferred connection %s",
                                      child->inst->app_io->common.name, conn->name);
 
                                conn->nr = fr_schedule_listen_add(sc, conn->listen);
                                if (!conn->nr) {
                                        ERROR("proto_%s - Failed scheduling deferred connection %s",
                                              child->inst->app_io->common.name, conn->name);
                                        /*
                                         *      Leave the entry in the hash table; the usual connection
                                         *      cleanup path will eventually remove it.
                                         */
                                }
                        } else {
                                /*
                                 *      The first connection is already scheduled.  Resume reads if it was
                                 *      paused, while waiting on verification.
                                 */
                                if (conn->paused) {
                                        conn->paused = false;
                                        (void) fr_event_filter_update(conn->el, conn->child->fd,
                                                                      FR_EVENT_FILTER_IO, resume_read);
                                }
 
                                child->state = PR_CLIENT_DYNAMIC;
                                child->radclient->active = true;
                        }
                }
        }
        pthread_mutex_unlock(&parent->mutex);
}
 
static ssize_t mod_write(fr_listen_t *li, void *packet_ctx, fr_time_t request_time,
                         uint8_t *buffer, size_t buffer_len, size_t written)
{
        fr_io_instance_t const *inst;
        fr_io_thread_t *thread;
        fr_io_connection_t *connection;
        fr_io_track_t *track = talloc_get_type_abort(packet_ctx, fr_io_track_t);
        fr_io_client_t *client;
        fr_client_t *radclient;
        fr_listen_t *child;
        fr_event_list_t *el;
        char const *name;
 
        get_inst(li, &inst, &thread, &connection, &child);
 
        client = track->client;
        if (connection) {
                el = connection->el;
                name = connection->name;
        } else {
                el = thread->el;
                name = li->name;
        }
 
        DEBUG3("Processing reply for %s", name);
 
        /*
         *      A fully defined client means that we just send the reply.
         */
        if (client->state != PR_CLIENT_PENDING) {
                ssize_t packet_len;
 
                track->finished = true;
 
                /*
                 *      The request received a conflicting packet, so we
                 *      discard this one.
                 */
                if (fr_time_neq(track->timestamp, request_time) || track->discard) {
                        fr_assert(track->packets > 0);
                        track->packets--;
                        DEBUG3("Suppressing reply as we have a newer / conflicing packet from the same source");
                        track->discard = true;
                        goto setup_timer;
                }
 
                /*
                 *      We have a NAK packet, or the request has timed
                 *      out, or it was discarded due to a conflicting
                 *      packet.  We don't respond, but we do cache the
                 *      "do not respond" reply for a period of time.
                 */
                if ((buffer_len == 1) || track->do_not_respond) {
                        DEBUG3("Not sending response to request - it is marked as 'do not respond'");
                        track->do_not_respond = true;
                        goto setup_timer;
                }
 
                /*
                 *      We have a real packet, write it to the network
                 *      via the underlying transport write.
                 */
                packet_len = inst->app_io->write(child, track, request_time,
                                                 buffer, buffer_len, written);
                if (packet_len <= 0) {
                        ERROR("Failed writing the reply - not sending any response on %s", name);
                        track->discard = true;
                        packet_expiry_timer(el->tl, fr_time_wrap(0), track);
                        return packet_len;
                }
 
                /*
                 *      Only a partial write.  The network code will
                 *      take care of calling us again, and we will set
                 *      the expiry timer at that point.
                 */
                if ((size_t) packet_len < buffer_len) {
                        DEBUG3("Partial write (%zd < %zu)", packet_len, buffer_len);
                        return packet_len;
                }
 
                /*
                 *      We're not tracking duplicates, so just expire
                 *      the packet now.
                 */
                if (!inst->app_io->track_duplicates) {
                        DEBUG3("Not tracking duplicates - expiring the request");
                        goto setup_timer;
                }
 
                /*
                 *      Cache the reply packet if we're doing dedup.
                 *
                 *      On resend duplicate reply, the reply is
                 *      already filled out.  So we don't do that twice.
                 */
                if (!track->reply) {
                        DEBUG3("Caching reply");
                        MEM(track->reply = talloc_memdup(track, buffer, buffer_len));
                        track->reply_len = buffer_len;
                }
 
                /*
                 *      Set the timer to expire the packet.
                 *
                 *      On dedup this also extends the timer.
                 */
        setup_timer:
                packet_expiry_timer(el->tl, fr_time_wrap(0), track);
                return buffer_len;
        }
 
        /*
         *      The client is pending, so we MUST have dynamic clients.
         *
         *      If there's a connected socket and no dynamic clients, then the
         *      client state is set to CONNECTED when the client is created.
         */
        fr_assert(inst->dynamic_clients);
        fr_assert(client->pending != NULL);
 
        /*
         *      We have a response that indicates the dynamic client creation failed.  Delete the client, and
         *      all pending packets.  Then, add the connection to the NAK cache.
         */
        if (buffer_len == 1) {
                if (*buffer == true) {
                        DEBUG("Request failed trying to define a new client.  Discarding client and pending packets.");
                } else {
                        INFO("proto_%s - Verification failed for packet from dynamic client %pV - adding IP address to the NAK cache",
                             inst->app_io->common.name, fr_box_ipaddr(client->src_ipaddr));
 
                        fr_io_nak_insert(client->thread, &client->src_ipaddr, inst->nak_lifetime);
                }
 
                if (!connection) {
                        talloc_free(client);
                        return buffer_len;
                }
 
                /*
                 *      Free pending packets and tracking table.
                 *      The table is parented by connection->parent, so won't
                 *      be auto-freed when connection->client is freed.
                 */
                TALLOC_FREE(client->pending);
                if (client->table) TALLOC_FREE(client->table);
 
                /*
                 *      Remove from parent's hash table so new packets won't
                 *      be routed to this connection.
                 */
                pthread_mutex_lock(&connection->parent->mutex);
                if (connection->in_parent_hash) {
                        connection->in_parent_hash = false;
                        (void) fr_hash_table_delete(connection->parent->ht, connection);
                }
                pthread_mutex_unlock(&connection->parent->mutex);
 
                /*
                 *      Tear down any sibling connections that were
                 *      deferred waiting on this verification.
                 */
                fr_io_connection_deny(connection->parent);
 
                /*
                 *      Mark the connection as dead, then trigger the
                 *      standard cleanup path via fr_network_listen_read().
                 *      This calls mod_read(), which sees connection->dead,
                 *      returns -1, and the network layer closes the
                 *      connection through its normal error handling.
                 */
                connection->dead = true;
                fr_network_listen_read(connection->nr, connection->listen);
 
                return buffer_len;
        }
 
        fr_assert(buffer_len == sizeof(radclient));
 
        memcpy(&radclient, buffer, sizeof(radclient));
 
        if (!connection) {
                fr_ipaddr_t ipaddr;
 
                /*
                 *      Check the encapsulating network against the
                 *      address that the user wants to use, but only
                 *      for unconnected sockets.
                 */
                if (client->network.af != radclient->ipaddr.af) {
                        DEBUG("Client IP address %pV IP family does not match the source network %pV of the packet.",
                              fr_box_ipaddr(radclient->ipaddr), fr_box_ipaddr(client->network));
                        goto error;
                }
 
                /*
                 *      Network prefix is more restrictive than the one given
                 *      by the client... that's bad.
                 */
                if (client->network.prefix > radclient->ipaddr.prefix) {
                        DEBUG("Client IP address %pV is not within the prefix with the defined network %pV",
                              fr_box_ipaddr(radclient->ipaddr), fr_box_ipaddr(client->network));
                        goto error;
                }
 
                ipaddr = radclient->ipaddr;
                fr_ipaddr_mask(&ipaddr, client->network.prefix);
                if (fr_ipaddr_cmp(&ipaddr, &client->network) != 0) {
                        DEBUG("Client IP address %pV is not within the defined network %pV.",
                              fr_box_ipaddr(radclient->ipaddr), fr_box_ipaddr(client->network));
                        goto error;
                }
 
                /*
                 *      We can't define dynamic clients as networks (for now).
                 *
                 *      @todo - If we did allow it, we would have to remove
                 *      this client from the trie, update it's IP address, and
                 *      re-add it.  We can PROBABLY do this if this client
                 *      isn't already connected, AND radclient->use_connected
                 *      is true.  But that's for later...
                 */
                if (((radclient->ipaddr.af == AF_INET) &&
                     (radclient->ipaddr.prefix != 32)) ||
                    ((radclient->ipaddr.af == AF_INET6) &&
                     (radclient->ipaddr.prefix != 128))) {
                        ERROR("Cannot define a dynamic client as a network");
 
                error:
                        talloc_free(radclient);
 
                        /*
                         *      Remove the pending client from the trie.
                         */
                        fr_assert(!connection);
                        talloc_free(client);
                        return buffer_len;
                }
        }
 
        update_client(client, radclient);
 
        // @todo - fill in other fields?
 
        talloc_free(radclient);
 
        radclient = client->radclient; /* laziness */
        radclient->server_cs = inst->server_cs;
        radclient->server = cf_section_name2(inst->server_cs);
 
        /*
         *      This is a connected socket, and it's just been
         *      allowed.  Go poke the network side to read from the
         *      socket.
         */
        if (connection) {
                fr_assert(connection != NULL);
                fr_assert(connection->client == client);
                fr_assert(client->connection != NULL);
 
                /*
                 *      Promote the parent and every sibling connection.
                 *      This also promotes the current child.
                 */
                fr_io_connection_allow(connection->parent, radclient);
 
                INFO("proto_%s - Verification succeeded for packet from dynamic client %pV - processing queued packets",
                     inst->app_io->common.name, fr_box_ipaddr(client->src_ipaddr));
                goto finish;
        } else {
                /*
                 *      Re-parent the conf section used to build this client
                 *      so its lifetime is linked to the client
                 */
                talloc_steal(radclient, radclient->cs);
        }
 
        fr_assert(connection == NULL);
        fr_assert(client->use_connected == false); /* we weren't sure until now */
 
        /*
         *      Disallow unsupported configurations.
         */
        if (radclient->use_connected && !inst->app_io->connection_set) {
                DEBUG("proto_%s - cannot use connected sockets as underlying 'transport = %s' does not support it.",
                      inst->app_io->common.name, inst->submodule->module->exported->name);
                goto error;
        }
 
 
        /*
         *      Dynamic clients can spawn new connections.
         */
        client->use_connected = radclient->use_connected;
 
        /*
         *      The admin has defined a client which uses connected
         *      sockets.  Go spawn it
         */
        if (client->use_connected) {
                fr_assert(connection == NULL);
 
 
                /*
                 *      Leave the state as PENDING.  Each connection
                 *      will then cause a dynamic client to be
                 *      defined.
                 */
                (void) pthread_mutex_init(&client->mutex, NULL);
                MEM(client->ht = fr_hash_table_alloc(client, connection_hash, connection_cmp, NULL));
 
        } else {
                /*
                 *      The client has been allowed.
                 */
                client->state = PR_CLIENT_DYNAMIC;
                client->radclient->active = true;
 
                INFO("proto_%s - Verification succeeded for packet from dynamic client %pV - processing %u queued packets",
                     inst->app_io->common.name, fr_box_ipaddr(client->src_ipaddr),
                     fr_heap_num_elements(client->pending));
        }
 
        /*
         *      Add this client to the master socket, so that
         *      mod_read() will see the pending client, pop the
         *      pending packet, and process it.
         *
         */
        if (!thread->pending_clients) {
                MEM(thread->pending_clients = fr_heap_alloc(thread, pending_client_cmp,
                                                           fr_io_client_t, pending_id, 0));
        }
 
        fr_assert(!fr_heap_entry_inserted((client->pending_id)));
        (void) fr_heap_insert(&thread->pending_clients, client);
 
finish:
        /*
         *      Maybe we defined the client, but the original packet
         *      timed out, so there's nothing more to do.  In that case, set up the expiry timers.
         */
        if (client->packets == 0) {
                client_expiry_timer(el->tl, fr_time_wrap(0), client);
        }
 
        /*
         *      If there are pending packets (and there should be at
         *      least one), tell the network socket to call our read()
         *      function again.
         */
        if (fr_heap_num_elements(client->pending) > 0) {
                if (connection) {
                        fr_network_listen_read(connection->nr, connection->listen);
                } else {
                        fr_network_listen_read(thread->nr, thread->listen);
                }
        }
 
        return buffer_len;
}
 
/** Close the socket.
 *
 */
static int mod_close(fr_listen_t *li)
{
        fr_io_instance_t const *inst;
        fr_io_connection_t *connection;
        fr_listen_t *child;
 
        get_inst(li, &inst, NULL, &connection, &child);
 
        if (inst->app_io->close) {
                int ret;
 
                ret = inst->app_io->close(child);
                if (ret < 0) return ret;
        } else {
                close(child->fd);
//              child->fd = -1;
        }
 
        if (!connection) return 0;
 
        /*
         *      We allocated this, so we're responsible for closing
         *      it.
         */
        DEBUG("Closing connection %s", connection->name);
        if (connection->client->pending) {
                TALLOC_FREE(connection->client->pending); /* for any pending packets */
        }
 
        /*
         *      Remove connection from parent hash table
         */
        pthread_mutex_lock(&connection->parent->mutex);
        if (connection->in_parent_hash) {
                connection->in_parent_hash = false;
                (void) fr_hash_table_delete(connection->parent->ht, connection);
        }
 
        /*
         *      If this is a dynamic client, and the parent has no more connections
         *      set up the timer to expire the dynamic client.
         */
        if ((connection->parent->state == PR_CLIENT_DYNAMIC) &&
            ((!connection->parent->ht) || (fr_hash_table_num_elements(connection->parent->ht) == 0))) {
                client_expiry_timer(connection->el->tl, fr_time_wrap(0), connection->parent);
        }
        pthread_mutex_unlock(&connection->parent->mutex);
 
        talloc_free(connection->mi);
 
        return 0;
}
 
static int mod_instantiate(module_inst_ctx_t const *mctx)
{
        fr_io_instance_t *inst = mctx->mi->data;
        CONF_SECTION *conf = mctx->mi->conf;
 
        inst->mi = mctx->mi;
        inst->app_io = (fr_app_io_t const *) inst->submodule->exported;
        inst->app_io_conf = inst->submodule->conf;
        inst->app_io_instance = inst->submodule->data;
 
        /*
         *      If we're not tracking duplicates then we don't need a
         *      cleanup delay.
         *
         *      If we are tracking duplicates, then we must have a non-zero cleanup delay.
         */
        if (!inst->app_io->track_duplicates) {
                inst->cleanup_delay = fr_time_delta_wrap(0);
 
        } else {
                FR_TIME_DELTA_BOUND_CHECK("cleanup_delay", inst->cleanup_delay, >=, fr_time_delta_from_sec(1));
 
                if (!inst->app_io->track_create) {
                        cf_log_err(inst->app_io_conf, "Internal error: 'track_duplicates' is set, but there is no 'track create' function");
                        return -1;
                }
        }
 
        /*
         *      Get various information after bootstrapping the
         *      application IO module.
         */
        if (inst->app_io->network_get) {
                inst->app_io->network_get(&inst->ipproto, &inst->dynamic_clients, &inst->networks, inst->app_io_instance);
        }
 
        if ((inst->ipproto == IPPROTO_TCP) && !inst->app_io->connection_set) {
                cf_log_err(inst->app_io_conf, "Missing 'connection set' API for proto_%s", inst->app_io->common.name);
                return -1;
        }
 
        /*
         *      Ensure that the dynamic client sections exist
         */
        if (inst->dynamic_clients) {
                CONF_SECTION *server = cf_item_to_section(cf_parent(conf));
 
                if (!cf_section_find(server, "new", "client")) {
                        cf_log_err(conf, "Cannot use 'dynamic_clients = yes' as the virtual server has no 'new client { ... }' section defined.");
                        return -1;
                }
 
                if (!cf_section_find(server, "add", "client")) {
                        cf_log_warn(conf, "No 'add client { ... }' section was defined.");
                }
 
                if (!cf_section_find(server, "deny", "client")) {
                        cf_log_warn(conf, "No 'deny client { ... }' section was defined.");
                }
        }
 
        /*
         *      Create a list of client modules.
         *
         *      FIXME - Probably only want to do this for connected sockets?
         *
         *      FIXME - We probably want write protect enabled?
         */
        inst->clients = module_list_alloc(inst, &module_list_type_thread_local, "clients", false);
        module_list_mask_set(inst->clients, MODULE_INSTANCE_BOOTSTRAPPED);
 
        return 0;
}
 
 
static char const *mod_name(fr_listen_t *li)
{
        fr_io_thread_t *thread;
        fr_io_connection_t *connection;
        fr_listen_t *child;
        fr_io_instance_t const *inst;
 
        get_inst(li, &inst, &thread, &connection, &child);
 
        fr_assert(child != NULL);
        return child->app_io->get_name(child);
}
 
/** Create a trie from arrays of allow / deny IP addresses
 *
 * @param ctx   the talloc ctx
 * @param af    the address family to allow
 * @param allow the array of IPs / networks to allow.  MUST be talloc'd
 * @param deny  the array of IPs / networks to deny.  MAY be NULL, MUST be talloc'd
 * @return
 *      - fr_trie_t on success
 *      - NULL on error
 */
fr_trie_t *fr_master_io_network(TALLOC_CTX *ctx, int af, fr_ipaddr_t *allow, fr_ipaddr_t *deny)
{
        fr_trie_t *trie;
        size_t i, num;
 
        MEM(trie = fr_trie_alloc(ctx, NULL, NULL));
 
        num = talloc_array_length(allow);
        fr_assert(num > 0);
 
        for (i = 0; i < num; i++) {
                fr_ipaddr_t *network;
 
                /*
                 *      Can't add v4 networks to a v6 socket, or vice versa.
                 */
                if (allow[i].af != af) {
                        fr_strerror_printf("Address family in entry %zu - 'allow = %pV' "
                                           "does not match 'ipaddr'", i + 1, fr_box_ipaddr(allow[i]));
                        talloc_free(trie);
                        return NULL;
                }
 
                /*
                 *      Duplicates are bad.
                 */
                network = fr_trie_match_by_key(trie,
                                        &allow[i].addr, allow[i].prefix);
                if (network) {
                        fr_strerror_printf("Cannot add duplicate entry 'allow = %pV'",
                                           fr_box_ipaddr(allow[i]));
                        talloc_free(trie);
                        return NULL;
                }
 
                /*
                 *      Look for overlapping entries.
                 *      i.e. the networks MUST be disjoint.
                 *
                 *      Note that this catches 192.168.1/24
                 *      followed by 192.168/16, but NOT the
                 *      other way around.  The best fix is
                 *      likely to add a flag to
                 *      fr_trie_alloc() saying "we can only
                 *      have terminal fr_trie_user_t nodes"
                 */
                network = fr_trie_lookup_by_key(trie,
                                         &allow[i].addr, allow[i].prefix);
                if (network && (network->prefix <= allow[i].prefix)) {
                        fr_strerror_printf("Cannot add overlapping entry 'allow = %pV'", fr_box_ipaddr(allow[i]));
                        fr_strerror_const("Entry is completely enclosed inside of a previously defined network.");
                        talloc_free(trie);
                        return NULL;
                }
 
                /*
                 *      Insert the network into the trie.
                 *      Lookups will return the fr_ipaddr_t of
                 *      the network.
                 */
                if (fr_trie_insert_by_key(trie,
                                   &allow[i].addr, allow[i].prefix,
                                   &allow[i]) < 0) {
                        fr_strerror_printf("Failed adding 'allow = %pV' to tracking table", fr_box_ipaddr(allow[i]));
                        talloc_free(trie);
                        return NULL;
                }
        }
 
        /*
         *      And now check denied networks.
         */
        num = talloc_array_length(deny);
        if (!num) return trie;
 
        /*
         *      Since the default is to deny, you can only add
         *      a "deny" inside of a previous "allow".
         */
        for (i = 0; i < num; i++) {
                fr_ipaddr_t *network;
 
                /*
                 *      Can't add v4 networks to a v6 socket, or vice versa.
                 */
                if (deny[i].af != af) {
                        fr_strerror_printf("Address family in entry %zu - 'deny = %pV' "
                                           "does not match 'ipaddr'", i + 1, fr_box_ipaddr(deny[i]));
                        talloc_free(trie);
                        return NULL;
                }
 
                /*
                 *      Duplicates are bad.
                 */
                network = fr_trie_match_by_key(trie,
                                        &deny[i].addr, deny[i].prefix);
                if (network) {
                        fr_strerror_printf("Cannot add duplicate entry 'deny = %pV'", fr_box_ipaddr(deny[i]));
                        talloc_free(trie);
                        return NULL;
                }
 
                /*
                 *      A "deny" can only be within a previous "allow".
                 */
                network = fr_trie_lookup_by_key(trie,
                                         &deny[i].addr, deny[i].prefix);
                if (!network) {
                        fr_strerror_printf("The network in entry %zu - 'deny = %pV' is not "
                                           "contained within a previous 'allow'", i + 1, fr_box_ipaddr(deny[i]));
                        talloc_free(trie);
                        return NULL;
                }
 
                /*
                 *      We hack the AF in "deny" rules.  If
                 *      the lookup gets AF_UNSPEC, then we're
                 *      adding a "deny" inside of a "deny".
                 */
                if (network->af != af) {
                        fr_strerror_printf("The network in entry %zu - 'deny = %pV' is overlaps "
                                           "with another 'deny' rule", i + 1, fr_box_ipaddr(deny[i]));
                        talloc_free(trie);
                        return NULL;
                }
 
                /*
                 *      Insert the network into the trie.
                 *      Lookups will return the fr_ipaddr_t of
                 *      the network.
                 */
                if (fr_trie_insert_by_key(trie,
                                   &deny[i].addr, deny[i].prefix,
                                   &deny[i]) < 0) {
                        fr_strerror_printf("Failed adding 'deny = %pV' to tracking table", fr_box_ipaddr(deny[i]));
                        talloc_free(trie);
                        return NULL;
                }
 
                /*
                 *      Hack it to make it a deny rule.
                 */
                deny[i].af = AF_UNSPEC;
        }
 
        return trie;
}
 
 
int fr_io_listen_free(fr_listen_t *li)
{
        if (!li->thread_instance) return 0;
 
        talloc_free(li->thread_instance);
        return 0;
}
 
int fr_master_io_listen(fr_io_instance_t *inst, fr_schedule_t *sc,
                        size_t default_message_size, size_t num_messages)
{
        fr_listen_t     *li, *child;
        fr_io_thread_t  *thread;
 
        /*
         *      No IO paths, so we don't initialize them.
         */
        if (!inst->app_io) {
                fr_assert(!inst->dynamic_clients);
                return 0;
        }
 
        if (!inst->app_io->common.thread_inst_size) {
                fr_strerror_const("IO modules MUST set 'thread_inst_size' when using the master IO handler.");
                return -1;
        }
 
        /*
         *      Build the #fr_listen_t.  This describes the complete
         *      path data takes from the socket to the decoder and
         *      back again.
         */
        MEM(li = talloc_zero(NULL, fr_listen_t));
        talloc_set_destructor(li, fr_io_listen_free);
 
        /*
         *      The first listener is the one for the application
         *      (e.g. RADIUS).  However, we mangle the IO path to
         *      point to the master IO handler.  That allows all of
         *      the high-level work (dynamic client checking,
         *      connected sockets, etc.) to be handled by the master
         *      IO handler.
         *
         *      This listener is then passed to the network code,
         *      which calls our trampoline functions to do the actual
         *      work.
         */
        li->app = inst->app;
        li->app_instance = inst->app_instance;
        li->server_cs = inst->server_cs;
 
        /*
         *      Set configurable parameters for message ring buffer.
         */
        li->default_message_size = default_message_size;
        li->num_messages = num_messages;
 
        /*
         *      Per-socket data lives here.
         */
        thread = talloc_zero(NULL, fr_io_thread_t);
        thread->listen = li;
        thread->sc = sc;
 
        /*
         *      Create the trie of clients and NAK clients.
         */
        MEM(thread->trie = fr_trie_alloc(thread, NULL, NULL));
        MEM(thread->nak_trie = fr_trie_alloc(thread, NULL, NULL));
        fr_io_nak_list_init(&thread->nak_list);
 
        if (inst->dynamic_clients) {
                MEM(thread->alive_clients = fr_heap_alloc(thread, alive_client_cmp,
                                                          fr_io_client_t, alive_id, 0));
        }
 
        /*
         *      Set the listener to call our master trampoline function.
         */
        li->cs = inst->app_io_conf;
        li->app_io = &fr_master_app_io;
        li->thread_instance = thread;
        li->app_io_instance = inst;
        li->track_duplicates = inst->app_io->track_duplicates;
        if (inst->app_io->hexdump_set) inst->app_io->hexdump_set(li, inst->app_io_instance);
 
        /*
         *      The child listener points to the *actual* IO path.
         *
         *      We need to create a complete listener here (e.g.
         *      RADIUS + RADIUS_UDP), because the underlying IO
         *      functions expect to get passed a full listener.
         *
         *      Once the network side calls us, we will call the child
         *      listener to do the actual IO.
         */
        child = thread->child = talloc_zero(li, fr_listen_t);
        memcpy(child, li, sizeof(*child));
 
        /*
         *      Reset these fields to point to the IO instance data.
         */
        child->app_io = inst->app_io;
        child->track_duplicates = inst->app_io->track_duplicates;
 
        if (child->app_io->common.thread_inst_size > 0) {
                child->thread_instance = talloc_zero_array(NULL, uint8_t,
                                                           inst->app_io->common.thread_inst_size);
                talloc_set_destructor(child, fr_io_listen_free);
 
                talloc_set_name(child->thread_instance, "proto_%s_thread_t",
                                inst->app_io->common.name);
 
                /*
                 *      This is "const", and the user can't
                 *      touch it.  So we just reuse the same
                 *      configuration everywhere.
                 */
                child->app_io_instance = inst->app_io_instance;
 
        } else {
                child->thread_instance = inst->app_io_instance;
                child->app_io_instance = child->thread_instance;
        }
 
        /*
         *      Don't call connection_set() for the main socket.  It's
         *      not connected.  Instead, tell the IO path to open the
         *      socket for us.
         */
        if (inst->app_io->open(child) < 0) {
                talloc_free(li);
                return -1;
        }
 
        li->fd = child->fd;     /* copy this back up */
 
        if (!child->app_io->get_name) {
                child->name = child->app_io->common.name;
        } else {
                child->name = child->app_io->get_name(child);
        }
        li->name = child->name;
 
        /*
         *      Record which socket we opened.
         */
        if (child->app_io_addr) {
                fr_listen_t *other;
 
                other = listen_find_any(thread->child);
                if (other) {
                        cf_log_err(other->cs, "Already opened socket %s", other->name);
                        cf_log_err(li->cs, "Failed opening duplicate socket - cannot use the same configuration for two different listen sections");
 
                        talloc_free(li);
                        return -1;
                }
 
                (void) listen_record(child);
        }
 
        /*
         *      Add the socket to the scheduler, where it might end up
         *      in a different thread.
         */
        if (!fr_schedule_listen_add(sc, li)) {
                talloc_free(li);
                return -1;
        }
 
        return 0;
}
 
/*
 *      Used to create a tracking structure for fr_network_sendto_worker()
 */
fr_io_track_t *fr_master_io_track_alloc(fr_listen_t *li, fr_client_t *radclient, fr_ipaddr_t const *src_ipaddr, int src_port,
                                        fr_ipaddr_t const *dst_ipaddr, int dst_port)
{
        fr_io_instance_t const  *inst;
        fr_io_thread_t          *thread;
        fr_io_connection_t      *connection;
        fr_listen_t             *child;
        fr_io_track_t           *track;
        fr_io_client_t          *client;
        fr_io_address_t         *address;
        fr_listen_t             *parent = talloc_parent(li);
 
        (void) talloc_get_type_abort(parent, fr_listen_t);
 
        get_inst(parent, &inst, &thread, &connection, &child);
 
        fr_assert(child == li);
 
        if (unlikely(!thread)) return NULL;
        fr_assert(thread->trie != NULL);
 
        client = fr_trie_lookup_by_key(thread->trie, &src_ipaddr->addr, src_ipaddr->prefix);
        if (!client) {
                MEM(client = client_alloc(thread, PR_CLIENT_STATIC, inst, thread, radclient, NULL));
        }
 
        MEM(track = talloc_zero_pooled_object(client->table, fr_io_track_t, 1, sizeof(*track) + sizeof(*track->address) + 64));
        MEM(track->address = address = talloc_zero(track, fr_io_address_t));
 
        track->li = li;
        track->client = client;
 
        address->socket.inet.src_port = src_port;
        address->socket.inet.dst_port = dst_port;
 
        address->socket.inet.src_ipaddr = *src_ipaddr;
        address->socket.inet.dst_ipaddr = *dst_ipaddr;
        address->radclient = radclient;
 
        return track;
}
 
 
fr_app_io_t fr_master_app_io = {
        .common = {
                .magic                  = MODULE_MAGIC_INIT,
                .name                   = "radius_master_io",
 
                .instantiate            = mod_instantiate,
        },
        .default_message_size   = 4096,
        .track_duplicates       = true,
 
        .read                   = mod_read,
        .write                  = mod_write,
        .inject                 = mod_inject,
 
        .open                   = mod_open,
        .close                  = mod_close,
        .event_list_set         = mod_event_list_set,
        .get_name               = mod_name,
};