/src/freeradius-server/src/lib/server/trunk.c

Source
/*
 *   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: fc85e0cd7a091e18670aa5b6916ffb5d84743c69 $
 *
 * @file src/lib/server/trunk.c
 * @brief A management API for bonding multiple connections together.
 *
 * @copyright 2019-2020 Arran Cudbard-Bell (a.cudbardb@freeradius.org)
 * @copyright 2019-2020 The FreeRADIUS server project
 */

#define LOG_PREFIX trunk->log_prefix

#ifdef NDEBUG
#  define TALLOC_GET_TYPE_ABORT_NOOP 1
#endif

typedef struct trunk_request_s trunk_request_t;
typedef struct trunk_connection_s trunk_connection_t;
typedef struct trunk_s trunk_t;
#define _TRUNK_PRIVATE 1
#include <freeradius-devel/server/trunk.h>

#include <freeradius-devel/server/trigger.h>
#include <freeradius-devel/util/debug.h>
#include <freeradius-devel/util/misc.h>
#include <freeradius-devel/util/syserror.h>
#include <freeradius-devel/util/minmax_heap.h>

#ifdef HAVE_STDATOMIC_H
#  include <stdatomic.h>
#  ifndef ATOMIC_VAR_INIT
#    define ATOMIC_VAR_INIT(_x) (_x)
#  endif
#else
#  include <freeradius-devel/util/stdatomic.h>
#endif

static atomic_uint_fast64_t request_counter = ATOMIC_VAR_INIT(1);

#ifdef TESTING_TRUNK
static fr_time_t test_time_base = fr_time_wrap(1);

static fr_time_t test_time(void)
{
  return test_time_base;
}

#define fr_time test_time
#endif

#ifndef NDEBUG
/** The maximum number of state logs to record per request
 *
 */
#define TRUNK_REQUEST_STATE_LOG_MAX 20

/** Trace state machine changes for a particular request
 *
 */
typedef struct {
  fr_dlist_head_t     *log_head;  //!< To allow the log entry to remove itself on free.
  fr_dlist_t      entry;    //!< Entry in the linked list.
  trunk_request_state_t   from;   //!< What state we transitioned from.
  trunk_request_state_t   to;   //!< What state we transitioned to.

  trunk_connection_t    *tconn;   //!< The request was associated with.
              ///< Pointer may now be invalid, do no de-reference.

  uint64_t      tconn_id; //!< If the treq was associated with a connection
              ///< the connection ID.
  trunk_connection_state_t  tconn_state;  //!< If the treq was associated with a connection
              ///< the connection state at the time of the
              ///< state transition.

  char const            *function;  //!< State change occurred in.
  int       line;   //!< Line change occurred on.
} trunk_request_state_log_t;
#endif

/** Wraps a normal request
 *
 */
struct trunk_request_s {
  struct trunk_request_pub_s  pub;    //!< Public fields in the trunk request.
              ///< This *MUST* be the first field in this
              ///< structure.

  uint64_t    id;     //!< Trunk request ID.

  fr_heap_index_t   heap_id;    //!< Used to track the request conn->pending heap.

  fr_dlist_t    entry;      //!< Used to track the trunk request in the conn->sent
              ///< or trunk->backlog request.

  trunk_cancel_reason_t cancel_reason;    //!< Why this request was cancelled.

  fr_time_t   last_freed;   //!< Last time this request was freed.

  bool      bound_to_conn;    //!< Fail the request if there's an attempt to
              ///< re-enqueue it.

  bool      sent;     //!< Trunk request has been sent at least once.
              ///< Used so that re-queueing doesn't increase trunk
              ///< `sent` count.

#ifndef NDEBUG
  fr_dlist_head_t   log;      //!< State change log.
#endif
};


/** Associates request queues with a connection
 *
 * @dotfile src/lib/server/trunk_conn.gv "Trunk connection state machine"
 * @dotfile src/lib/server/trunk_req.gv "Trunk request state machine"
 */
struct trunk_connection_s {
  struct trunk_connection_pub_s pub;    //!< Public fields in the trunk connection.
              ///< This *MUST* be the first field in this
              ///< structure.

  fr_heap_index_t   heap_id;    //!< Used to track the connection in the connected
              ///< heap.

  fr_dlist_t    entry;      //!< Used to track the connection in the connecting,
              ///< full and failed lists.

  /** @name State
   * @{
   */
  trunk_connection_event_t events;    //!< The current events we expect to be notified on.
  /** @} */

  /** @name Request lists
   * @{
   */
  fr_heap_t   *pending;   //!< Requests waiting to be sent.

  trunk_request_t   *partial;   //!< Partially written request.

  fr_dlist_head_t   sent;     //!< Sent request.

  fr_dlist_head_t   reapable;   //!< Idle request.

  fr_dlist_head_t   cancel;     //!< Requests in the cancel state.

  trunk_request_t   *cancel_partial;  //!< Partially written cancellation request.

  fr_dlist_head_t   cancel_sent;    //!< Sent cancellation request.
  /** @} */

  /** @name Statistics
   * @{
   */
  uint64_t    sent_count;   //!< The number of requests that have been sent using
              ///< this connection.
  /** @} */

  /** @name Timers
   * @{
   */
    fr_timer_t  *lifetime_ev;   //!< Maximum time this connection can be open.
    /** @} */
};

/** An entry in a trunk watch function list
 *
 */
typedef struct trunk_watch_entry_s {
  fr_dlist_t    entry;      //!< List entry.
  trunk_watch_t   func;     //!< Function to call when a trunk enters
              ///< the state this list belongs to
  bool      oneshot;    //!< Remove the function after it's called once.
  bool      enabled;    //!< Whether the watch entry is enabled.
  void      *uctx;      //!< User data to pass to the function.
} trunk_watch_entry_t;

/** Map connection states to trigger names
 *
 * Must stay in the same order as #trunk_connection_state_t
 */
static fr_table_num_indexed_bit_pos_t const trunk_conn_trigger_names[] = {
  { L("pool.connection_halted"),      TRUNK_CONN_HALTED     },  /* 0x0000 - bit 0 */
  { L("pool.connection_init"),      TRUNK_CONN_INIT       },  /* 0x0001 - bit 1 */
  { L("pool.connection_connecting"),    TRUNK_CONN_CONNECTING     },  /* 0x0002 - bit 2 */
  { L("pool.connection_active"),      TRUNK_CONN_ACTIVE     },  /* 0x0004 - bit 3 */
  { L("pool.connection_closed"),      TRUNK_CONN_CLOSED     },  /* 0x0008 - bit 4 */
  { L("pool.connection_full"),      TRUNK_CONN_FULL       },  /* 0x0010 - bit 5 */
  { L("pool.connection_inactive"),    TRUNK_CONN_INACTIVE     },  /* 0x0020 - bit 6 */
  { L("pool.connection_inactive_draining"), TRUNK_CONN_INACTIVE_DRAINING    },  /* 0x0040 - bit 7 */
  { L("pool.connection_draining"),    TRUNK_CONN_DRAINING     },  /* 0x0080 - bit 8 */
  { L("pool.connection_draining_to_free"),  TRUNK_CONN_DRAINING_TO_FREE   } /* 0x0100 - bit 9 */
};
static size_t trunk_conn_trigger_names_len = NUM_ELEMENTS(trunk_conn_trigger_names);

/** Main trunk management handle
 *
 */
struct trunk_s {
  struct trunk_pub_s  pub;      //!< Public fields in the trunk connection.
              ///< This *MUST* be the first field in this
              ///< structure.

  char const    *log_prefix;    //!< What to prepend to messages.

  fr_event_list_t   *el;      //!< Event list used by this trunk and the connection.

  trunk_conf_t    conf;     //!< Trunk common configuration.

  fr_dlist_head_t   free_requests;    //!< Requests in the unassigned state.  Waiting to be
              ///< enqueued.

  fr_heap_t   *backlog;   //!< The request backlog.  Requests we couldn't
              ///< immediately assign to a connection.

  /** @name Connection lists
   *
   * A connection must always be in exactly one of these lists
   * or trees.
   *
   * @{
   */
  fr_dlist_head_t   init;     //!< Connections which have not yet started
              ///< connecting.

  fr_dlist_head_t   connecting;   //!< Connections which are not yet in the open state.

  fr_minmax_heap_t  *active;    //!< Connections which can service requests.

  fr_dlist_head_t   full;     //!< Connections which have too many outstanding
              ///< requests.

  fr_dlist_head_t   inactive;   //!< Connections which have been signalled to be
              ///< inactive by the API client.

  fr_dlist_head_t   inactive_draining;  //!< Connections which have been signalled to be
              ///< inactive by the API client, which the trunk
              ///< manager is draining to close.

  fr_dlist_head_t   failed;     //!< Connections that'll be reconnected shortly.

  fr_dlist_head_t   closed;     //!< Connections that have closed. Either due to
              ///< shutdown, reconnection or failure.

  fr_dlist_head_t   draining;   //!< Connections that will be freed once all their
              ///< requests are complete, but can be reactivated.

  fr_dlist_head_t   draining_to_free; //!< Connections that will be freed once all their
              ///< requests are complete.

  fr_dlist_head_t   to_free;    //!< Connections we're done with and will free on
              //!< the next call to trunk_manage.
              //!< This prevents connections from being freed
              //!< whilst we're inside callbacks.
  /** @} */

  /** @name Callbacks
   * @{
   */
  trunk_io_funcs_t  funcs;      //!< I/O functions.

  void      *in_handler;    //!< Which handler we're inside.

  void      *uctx;      //!< Uctx data to pass to alloc.

  fr_dlist_head_t   watch[TRUNK_STATE_MAX]; //!< To be called when trunk changes state.

  trunk_watch_entry_t *next_watcher;    //!< Watcher about to be run. Used to prevent nested watchers.
  /** @} */

  /** @name Timers
   * @{
   */
  fr_timer_t    *manage_ev;   //!< Periodic connection management event.
  /** @} */

  /** @name Log rate limiting entries
   * @{
   */
  fr_rate_limit_t   limit_max_requests_alloc_log; //!< Rate limit on "Refusing to alloc requests - Limit of * requests reached"

  fr_rate_limit_t   limit_last_failure_log; //!< Rate limit on "Refusing to enqueue requests - No active conns"
  /** @} */

  /** @name State
   * @{
   */
  bool      freeing;    //!< Trunk is being freed, don't spawn new
              ///< connections or re-enqueue.

  bool      started;    //!< Has the trunk been started.

  bool      managing_connections; //!< Whether the trunk is allowed to manage
              ///< (open/close) connections.

  uint64_t    last_req_per_conn;  //!< The last request to connection ratio we calculated.
  /** @} */

  fr_pair_list_t    *trigger_args;    //!< Passed to trigger

  bool      trigger_undef[NUM_ELEMENTS(trunk_conn_trigger_names)];  //!< Record that a specific trigger is undefined.

  CONF_PAIR   *trigger_cp[NUM_ELEMENTS(trunk_conn_trigger_names)];  //!< Cached trigger CONF_PAIRs
};

int trunk_trigger_cf_parse(TALLOC_CTX *ctx, void *out, void *parent, CONF_ITEM *ci, conf_parser_t const *rule);

static conf_parser_t const trunk_config_request[] = {
  { FR_CONF_OFFSET("per_connection_max", trunk_conf_t, max_req_per_conn), .dflt = "2000" },
  { FR_CONF_OFFSET("per_connection_target", trunk_conf_t, target_req_per_conn), .dflt = "1000" },
  { FR_CONF_OFFSET("free_delay", trunk_conf_t, req_cleanup_delay), .dflt = "10.0" },
  { FR_CONF_OFFSET("triggers", trunk_conf_t, req_triggers), .func = trunk_trigger_cf_parse },

  CONF_PARSER_TERMINATOR
};

static conf_parser_t const trunk_config_connection[] = {
  { FR_CONF_OFFSET("connect_timeout", connection_conf_t, connection_timeout), .dflt = "3.0" },
  { FR_CONF_OFFSET("reconnect_delay", connection_conf_t, reconnection_delay), .dflt = "1" },

  CONF_PARSER_TERMINATOR
};

#ifndef TRUNK_TESTS
conf_parser_t const trunk_config[] = {
  { FR_CONF_OFFSET("start", trunk_conf_t, start), .dflt = "1" },
  { FR_CONF_OFFSET("min", trunk_conf_t, min), .dflt = "1" },
  { FR_CONF_OFFSET("max", trunk_conf_t, max), .dflt = "5" },
  { FR_CONF_OFFSET("connecting", trunk_conf_t, connecting), .dflt = "2" },
  { FR_CONF_OFFSET("uses", trunk_conf_t, max_uses), .dflt = "0" },
  { FR_CONF_OFFSET("lifetime", trunk_conf_t, lifetime), .dflt = "0" },
  { FR_CONF_OFFSET("idle_timeout", trunk_conf_t, idle_timeout), .dflt = "0" },

  { FR_CONF_OFFSET("open_delay", trunk_conf_t, open_delay), .dflt = "0.2" },
  { FR_CONF_OFFSET("close_delay", trunk_conf_t, close_delay), .dflt = "10.0" },

  { FR_CONF_OFFSET("manage_interval", trunk_conf_t, manage_interval), .dflt = "0.2" },

  { FR_CONF_OFFSET("max_backlog", trunk_conf_t, max_backlog), .dflt = "1000" },

  { FR_CONF_OFFSET("backlog_on_failed_conn", trunk_conf_t, backlog_on_failed_conn), },

  { FR_CONF_OFFSET("triggers", trunk_conf_t, conn_triggers), .func = trunk_trigger_cf_parse },

  { FR_CONF_OFFSET_SUBSECTION("connection", 0, trunk_conf_t, conn_conf, trunk_config_connection), .subcs_size = sizeof(trunk_config_connection) },
  { FR_CONF_POINTER("request", 0, CONF_FLAG_SUBSECTION, NULL), .subcs = (void const *) trunk_config_request },

  CONF_PARSER_TERMINATOR
};
#endif

#ifndef NDEBUG
/** Map request states to trigger names
 *
 * Must stay in the same order as #trunk_connection_state_t
 */
static fr_table_num_indexed_bit_pos_t const trunk_req_trigger_names[] = {
  { L("pool.request_init"),     TRUNK_REQUEST_STATE_INIT    },  /* 0x0000 - bit 0 */
  { L("pool.request_unassigned"),     TRUNK_REQUEST_STATE_UNASSIGNED    },  /* 0x0001 - bit 1 */
  { L("pool.request_backlog"),      TRUNK_REQUEST_STATE_BACKLOG   },  /* 0x0002 - bit 2 */
  { L("pool.request_pending"),      TRUNK_REQUEST_STATE_PENDING   },  /* 0x0004 - bit 3 */
  { L("pool.request_partial"),      TRUNK_REQUEST_STATE_PARTIAL   },  /* 0x0008 - bit 4 */
  { L("pool.request_sent"),     TRUNK_REQUEST_STATE_SENT    },  /* 0x0010 - bit 5 */
  { L("pool.request_state_reapable"),   TRUNK_REQUEST_STATE_REAPABLE    },  /* 0x0020 - bit 6 */
  { L("pool.request_complete"),     TRUNK_REQUEST_STATE_COMPLETE    },  /* 0x0040 - bit 7 */
  { L("pool.request_state_failed"),   TRUNK_REQUEST_STATE_FAILED    },  /* 0x0080 - bit 8 */
  { L("pool.request_state_cancel"),   TRUNK_REQUEST_STATE_CANCEL    },  /* 0x0100 - bit 9 */
  { L("pool.request_state_cancel_sent"),    TRUNK_REQUEST_STATE_CANCEL_SENT   },  /* 0x0200 - bit 10 */
  { L("pool.request_state_cancel_partial"), TRUNK_REQUEST_STATE_CANCEL_PARTIAL  },  /* 0x0400 - bit 11 */
  { L("pool.request_state_cancel_complete"),  TRUNK_REQUEST_STATE_CANCEL_COMPLETE },  /* 0x0800 - bit 12 */
};
static size_t trunk_req_trigger_names_len = NUM_ELEMENTS(trunk_req_trigger_names);
#endif

static fr_table_num_ordered_t const trunk_request_states[] = {
  { L("INIT"),          TRUNK_REQUEST_STATE_INIT    },
  { L("UNASSIGNED"),        TRUNK_REQUEST_STATE_UNASSIGNED    },
  { L("BACKLOG"),         TRUNK_REQUEST_STATE_BACKLOG   },
  { L("PENDING"),         TRUNK_REQUEST_STATE_PENDING   },
  { L("PARTIAL"),         TRUNK_REQUEST_STATE_PARTIAL   },
  { L("SENT"),          TRUNK_REQUEST_STATE_SENT    },
  { L("REAPABLE"),        TRUNK_REQUEST_STATE_REAPABLE    },
  { L("COMPLETE"),        TRUNK_REQUEST_STATE_COMPLETE    },
  { L("FAILED"),          TRUNK_REQUEST_STATE_FAILED    },
  { L("CANCEL"),          TRUNK_REQUEST_STATE_CANCEL    },
  { L("CANCEL-SENT"),       TRUNK_REQUEST_STATE_CANCEL_SENT   },
  { L("CANCEL-PARTIAL"),        TRUNK_REQUEST_STATE_CANCEL_PARTIAL  },
  { L("CANCEL-COMPLETE"),       TRUNK_REQUEST_STATE_CANCEL_COMPLETE }
};
static size_t trunk_request_states_len = NUM_ELEMENTS(trunk_request_states);

static fr_table_num_ordered_t const trunk_states[] = {
  { L("IDLE"),          TRUNK_STATE_IDLE      },
  { L("ACTIVE"),          TRUNK_STATE_ACTIVE      },
  { L("PENDING"),         TRUNK_STATE_PENDING     },
  { L("FULL"),          TRUNK_STATE_FULL      },
  { L("FAILED"),          TRUNK_STATE_FAILED      }
};
static size_t trunk_states_len = NUM_ELEMENTS(trunk_states);

static fr_table_num_ordered_t const trunk_connection_states[] = {
  { L("INIT"),          TRUNK_CONN_INIT       },
  { L("HALTED"),          TRUNK_CONN_HALTED     },
  { L("CONNECTING"),        TRUNK_CONN_CONNECTING     },
  { L("ACTIVE"),          TRUNK_CONN_ACTIVE     },
  { L("CLOSED"),          TRUNK_CONN_CLOSED     },
  { L("FULL"),          TRUNK_CONN_FULL       },
  { L("INACTIVE"),        TRUNK_CONN_INACTIVE     },
  { L("INACTIVE-DRAINING"),     TRUNK_CONN_INACTIVE_DRAINING    },
  { L("DRAINING"),        TRUNK_CONN_DRAINING     },
  { L("DRAINING-TO-FREE"),      TRUNK_CONN_DRAINING_TO_FREE   }
};
static size_t trunk_connection_states_len = NUM_ELEMENTS(trunk_connection_states);

static fr_table_num_ordered_t const trunk_cancellation_reasons[] = {
  { L("TRUNK_CANCEL_REASON_NONE"),    TRUNK_CANCEL_REASON_NONE    },
  { L("TRUNK_CANCEL_REASON_SIGNAL"),    TRUNK_CANCEL_REASON_SIGNAL    },
  { L("TRUNK_CANCEL_REASON_MOVE"),    TRUNK_CANCEL_REASON_MOVE    },
  { L("TRUNK_CANCEL_REASON_REQUEUE"),   TRUNK_CANCEL_REASON_REQUEUE   }
};
static size_t trunk_cancellation_reasons_len = NUM_ELEMENTS(trunk_cancellation_reasons);

static fr_table_num_ordered_t const trunk_connection_events[] = {
  { L("TRUNK_CONN_EVENT_NONE"),   TRUNK_CONN_EVENT_NONE     },
  { L("TRUNK_CONN_EVENT_READ"),   TRUNK_CONN_EVENT_READ   },
  { L("TRUNK_CONN_EVENT_WRITE"),    TRUNK_CONN_EVENT_WRITE    },
  { L("TRUNK_CONN_EVENT_BOTH"),   TRUNK_CONN_EVENT_BOTH   },
};
static size_t trunk_connection_events_len = NUM_ELEMENTS(trunk_connection_events);

#define CONN_TRIGGER(_state) do { \
  uint8_t idx = fr_high_bit_pos(_state); \
  if (trunk->conf.conn_triggers && !trunk->trigger_undef[idx]) { \
    if (trigger(unlang_interpret_get_thread_default(), trunk->conf.conn_trigger_cs, \
          &trunk->trigger_cp[idx], \
          fr_table_str_by_value(trunk_conn_trigger_names, _state, \
              "<INVALID>"), true, trunk->trigger_args) == -1) { \
      trunk->trigger_undef[idx] = true; \
    } \
  } \
} while (0)

#define CONN_STATE_TRANSITION(_new, _log) \
do { \
  _log("[%" PRIu64 "] Trunk connection changed state %s -> %s", \
       tconn->pub.conn->id, \
       fr_table_str_by_value(trunk_connection_states, tconn->pub.state, "<INVALID>"), \
       fr_table_str_by_value(trunk_connection_states, _new, "<INVALID>")); \
  tconn->pub.state = _new; \
  CONN_TRIGGER(_new); \
  trunk_requests_per_connection(NULL, NULL, trunk, fr_time(), false); \
} while (0)

#define CONN_BAD_STATE_TRANSITION(_new) \
do { \
  if (!fr_cond_assert_msg(0, "[%" PRIu64 "] Trunk connection invalid transition %s -> %s", \
        tconn->pub.conn->id, \
        fr_table_str_by_value(trunk_connection_states, tconn->pub.state, "<INVALID>"),  \
        fr_table_str_by_value(trunk_connection_states, _new, "<INVALID>"))) return; \
} while (0)

#ifndef NDEBUG
void trunk_request_state_log_entry_add(char const *function, int line,
               trunk_request_t *treq, trunk_request_state_t new) CC_HINT(nonnull);

#define REQUEST_TRIGGER(_state) do { \
  if (trunk->conf.req_triggers) { \
    trigger(unlang_interpret_get_thread_default(), \
      trunk->conf.req_trigger_cs, NULL, fr_table_str_by_value(trunk_req_trigger_names, _state, \
               "<INVALID>"), true, trunk->trigger_args); \
  } \
} while (0)

/** Record a request state transition and log appropriate output
 *
 */
#define REQUEST_STATE_TRANSITION(_new) \
do { \
  request_t *request = treq->pub.request; \
  ROPTIONAL(RDEBUG3, DEBUG3, "Trunk request %" PRIu64 " changed state %s -> %s", \
      treq->id, \
      fr_table_str_by_value(trunk_request_states, treq->pub.state, "<INVALID>"), \
      fr_table_str_by_value(trunk_request_states, _new, "<INVALID>")); \
  trunk_request_state_log_entry_add(__FUNCTION__, __LINE__, treq, _new); \
  treq->pub.state = _new; \
  REQUEST_TRIGGER(_new); \
} while (0)
#define REQUEST_BAD_STATE_TRANSITION(_new) \
do { \
  trunk_request_state_log(&default_log, L_ERR, __FILE__, __LINE__, treq); \
  if (!fr_cond_assert_msg(0, "Trunk request %" PRIu64 " invalid transition %s -> %s", \
        treq->id, \
        fr_table_str_by_value(trunk_request_states, treq->pub.state, "<INVALID>"), \
        fr_table_str_by_value(trunk_request_states, _new, "<INVALID>"))) return; \
} while (0)
#else
/** Record a request state transition
 *
 */
#define REQUEST_STATE_TRANSITION(_new) \
do { \
  request_t *request = treq->pub.request; \
  ROPTIONAL(RDEBUG3, DEBUG3, "Trunk request %" PRIu64 " changed state %s -> %s", \
      treq->id, \
      fr_table_str_by_value(trunk_request_states, treq->pub.state, "<INVALID>"), \
      fr_table_str_by_value(trunk_request_states, _new, "<INVALID>")); \
  treq->pub.state = _new; \
} while (0)
#define REQUEST_BAD_STATE_TRANSITION(_new) \
do { \
  if (!fr_cond_assert_msg(0, "Trunk request %" PRIu64 " invalid transition %s -> %s", \
        treq->id, \
        fr_table_str_by_value(trunk_request_states, treq->pub.state, "<INVALID>"), \
        fr_table_str_by_value(trunk_request_states, _new, "<INVALID>"))) return; \
} while (0)
#endif


/** Call the cancel callback if set
 *
 */
#define DO_REQUEST_CANCEL(_treq, _reason) \
do { \
  if ((_treq)->pub.trunk->funcs.request_cancel) { \
    request_t *request = (_treq)->pub.request; \
    void *_prev = (_treq)->pub.trunk->in_handler; \
    (_treq)->pub.trunk->in_handler = (void *)(_treq)->pub.trunk->funcs.request_cancel; \
    ROPTIONAL(RDEBUG3, DEBUG3, "Calling request_cancel(conn=%p, preq=%p, reason=%s, uctx=%p)", \
        (_treq)->pub.tconn->pub.conn, \
        (_treq)->pub.preq, \
        fr_table_str_by_value(trunk_cancellation_reasons, \
        (_reason), \
        "<INVALID>"), \
        (_treq)->pub.trunk->uctx); \
    (_treq)->pub.trunk->funcs.request_cancel((_treq)->pub.tconn->pub.conn, (_treq)->pub.preq, (_reason), (_treq)->pub.trunk->uctx); \
    (_treq)->pub.trunk->in_handler = _prev; \
  } \
} while(0)

/** Call the "conn_release" callback (if set)
 *
 */
#define DO_REQUEST_CONN_RELEASE(_treq) \
do { \
  if ((_treq)->pub.trunk->funcs.request_conn_release) { \
    request_t *request = (_treq)->pub.request; \
    void *_prev = (_treq)->pub.trunk->in_handler; \
    (_treq)->pub.trunk->in_handler = (void *)(_treq)->pub.trunk->funcs.request_conn_release; \
    ROPTIONAL(RDEBUG3, DEBUG3, "Calling request_conn_release(conn=%p, preq=%p, uctx=%p)", \
        (_treq)->pub.tconn->pub.conn, \
        (_treq)->pub.preq, \
        (_treq)->pub.trunk->uctx); \
    (_treq)->pub.trunk->funcs.request_conn_release((_treq)->pub.tconn->pub.conn, (_treq)->pub.preq, (_treq)->pub.trunk->uctx); \
    (_treq)->pub.trunk->in_handler = _prev; \
  } \
} while(0)

/** Call the complete callback (if set)
 *
 */
#define DO_REQUEST_COMPLETE(_treq) \
do { \
  if ((_treq)->pub.trunk->funcs.request_complete) { \
    request_t *request = (_treq)->pub.request; \
    void *_prev = (_treq)->pub.trunk->in_handler; \
    ROPTIONAL(RDEBUG3, DEBUG3, "Calling request_complete(request=%p, preq=%p, rctx=%p, uctx=%p)", \
        (_treq)->pub.request, \
        (_treq)->pub.preq, \
        (_treq)->pub.rctx, \
        (_treq)->pub.trunk->uctx); \
    (_treq)->pub.trunk->in_handler = (void *)(_treq)->pub.trunk->funcs.request_complete; \
    (_treq)->pub.trunk->funcs.request_complete((_treq)->pub.request, (_treq)->pub.preq, (_treq)->pub.rctx, (_treq)->pub.trunk->uctx); \
    (_treq)->pub.trunk->in_handler = _prev; \
  } \
} while(0)

/** Call the fail callback (if set)
 *
 */
#define DO_REQUEST_FAIL(_treq, _prev_state) \
do { \
  if ((_treq)->pub.trunk->funcs.request_fail) { \
    request_t *request = (_treq)->pub.request; \
    void *_prev = (_treq)->pub.trunk->in_handler; \
    ROPTIONAL(RDEBUG3, DEBUG3, "Calling request_fail(request=%p, preq=%p, rctx=%p, state=%s uctx=%p)", \
        (_treq)->pub.request, \
        (_treq)->pub.preq, \
        (_treq)->pub.rctx, \
        fr_table_str_by_value(trunk_request_states, (_prev_state), "<INVALID>"), \
        (_treq)->pub.trunk->uctx); \
    (_treq)->pub.trunk->in_handler = (void *)(_treq)->pub.trunk->funcs.request_fail; \
    (_treq)->pub.trunk->funcs.request_fail((_treq)->pub.request, (_treq)->pub.preq, (_treq)->pub.rctx, _prev_state, (_treq)->pub.trunk->uctx); \
    (_treq)->pub.trunk->in_handler = _prev; \
  } \
} while(0)

/** Call the free callback (if set)
 *
 */
#define DO_REQUEST_FREE(_treq) \
do { \
  if ((_treq)->pub.trunk->funcs.request_free) { \
    request_t *request = (_treq)->pub.request; \
    void *_prev = (_treq)->pub.trunk->in_handler; \
    ROPTIONAL(RDEBUG3, DEBUG3, "Calling request_free(request=%p, preq=%p, uctx=%p)", \
        (_treq)->pub.request, \
        (_treq)->pub.preq, \
        (_treq)->pub.trunk->uctx); \
    (_treq)->pub.trunk->in_handler = (void *)(_treq)->pub.trunk->funcs.request_free; \
    (_treq)->pub.trunk->funcs.request_free((_treq)->pub.request, (_treq)->pub.preq, (_treq)->pub.trunk->uctx); \
    (_treq)->pub.trunk->in_handler = _prev; \
  } \
} while(0)

/** Write one or more requests to a connection
 *
 */
#define DO_REQUEST_MUX(_tconn) \
do { \
  void *_prev = (_tconn)->pub.trunk->in_handler; \
  DEBUG3("[%" PRIu64 "] Calling request_mux(el=%p, tconn=%p, conn=%p, uctx=%p)", \
         (_tconn)->pub.conn->id, \
         (_tconn)->pub.trunk->el, \
         (_tconn), \
         (_tconn)->pub.conn, \
         (_tconn)->pub.trunk->uctx); \
  (_tconn)->pub.trunk->in_handler = (void *)(_tconn)->pub.trunk->funcs.request_mux; \
  (_tconn)->pub.trunk->funcs.request_mux((_tconn)->pub.trunk->el, (_tconn), (_tconn)->pub.conn, (_tconn)->pub.trunk->uctx); \
  (_tconn)->pub.trunk->in_handler = _prev; \
} while(0)

/** Read one or more requests from a connection
 *
 */
#define DO_REQUEST_DEMUX(_tconn) \
do { \
  void *_prev = (_tconn)->pub.trunk->in_handler; \
  DEBUG3("[%" PRIu64 "] Calling request_demux(tconn=%p, conn=%p, uctx=%p)", \
         (_tconn)->pub.conn->id, \
         (_tconn), \
         (_tconn)->pub.conn, \
         (_tconn)->pub.trunk->uctx); \
  (_tconn)->pub.trunk->in_handler = (void *)(_tconn)->pub.trunk->funcs.request_demux; \
  (_tconn)->pub.trunk->funcs.request_demux((_tconn)->pub.trunk->el, (_tconn), (_tconn)->pub.conn, (_tconn)->pub.trunk->uctx); \
  (_tconn)->pub.trunk->in_handler = _prev; \
} while(0)

/** Write one or more cancellation requests to a connection
 *
 */
#define DO_REQUEST_CANCEL_MUX(_tconn) \
do { \
  if ((_tconn)->pub.trunk->funcs.request_cancel_mux) { \
    void *_prev = (_tconn)->pub.trunk->in_handler; \
    DEBUG3("[%" PRIu64 "] Calling request_cancel_mux(tconn=%p, conn=%p, uctx=%p)", \
           (_tconn)->pub.conn->id, \
           (_tconn), \
           (_tconn)->pub.conn, \
           (_tconn)->pub.trunk->uctx); \
    (_tconn)->pub.trunk->in_handler = (void *)(_tconn)->pub.trunk->funcs.request_cancel_mux; \
    (_tconn)->pub.trunk->funcs.request_cancel_mux((_tconn)->pub.trunk->el, (_tconn), (_tconn)->pub.conn, (_tconn)->pub.trunk->uctx); \
    (_tconn)->pub.trunk->in_handler = _prev; \
  } \
} while(0)

/** Allocate a new connection
 *
 */
#define DO_CONNECTION_ALLOC(_tconn) \
do { \
  void *_prev = trunk->in_handler; \
  DEBUG3("Calling connection_alloc(tconn=%p, el=%p, conf=%p, log_prefix=\"%s\", uctx=%p)", \
         (_tconn), \
         (_tconn)->pub.trunk->el, \
         (_tconn)->pub.trunk->conf.conn_conf, \
         trunk->log_prefix, \
         (_tconn)->pub.trunk->uctx); \
  (_tconn)->pub.trunk->in_handler = (void *) (_tconn)->pub.trunk->funcs.connection_alloc; \
  (_tconn)->pub.conn = trunk->funcs.connection_alloc((_tconn), (_tconn)->pub.trunk->el, (_tconn)->pub.trunk->conf.conn_conf, (_tconn)->pub.trunk->log_prefix, trunk->uctx); \
  (_tconn)->pub.trunk->in_handler = _prev; \
  if (!(_tconn)->pub.conn) { \
    ERROR("Failed creating new connection"); \
    talloc_free(tconn); \
    return -1; \
  } \
} while(0)

/** Change what events the connection should be notified about
 *
 */
#define DO_CONNECTION_NOTIFY(_tconn, _events) \
do { \
  if ((_tconn)->pub.trunk->funcs.connection_notify) { \
    void *_prev = (_tconn)->pub.trunk->in_handler; \
    DEBUG3("[%" PRIu64 "] Calling connection_notify(tconn=%p, conn=%p, el=%p, events=%s, uctx=%p)", \
           (_tconn)->pub.conn->id, \
           (_tconn), \
           (_tconn)->pub.conn, \
           (_tconn)->pub.trunk->el, \
           fr_table_str_by_value(trunk_connection_events, (_events), "<INVALID>"), \
           (_tconn)->pub.trunk->uctx); \
    (_tconn)->pub.trunk->in_handler = (void *)(_tconn)->pub.trunk->funcs.connection_notify; \
    (_tconn)->pub.trunk->funcs.connection_notify((_tconn), (_tconn)->pub.conn, (_tconn)->pub.trunk->el, (_events), (_tconn)->pub.trunk->uctx); \
    (_tconn)->pub.trunk->in_handler = _prev; \
  } \
} while(0)

#define IN_HANDLER(_trunk)    (((_trunk)->in_handler) != NULL)
#define IN_REQUEST_MUX(_trunk)    (((_trunk)->funcs.request_mux) && ((_trunk)->in_handler == (void *)(_trunk)->funcs.request_mux))
#define IN_REQUEST_DEMUX(_trunk)  (((_trunk)->funcs.request_demux) && ((_trunk)->in_handler == (void *)(_trunk)->funcs.request_demux))
#define IN_REQUEST_CANCEL_MUX(_trunk) (((_trunk)->funcs.request_cancel_mux) && ((_trunk)->in_handler == (void *)(_trunk)->funcs.request_cancel_mux))

#define IS_SERVICEABLE(_tconn)    ((_tconn)->pub.state & TRUNK_CONN_SERVICEABLE)
#define IS_PROCESSING(_tconn)   ((_tconn)->pub.state & TRUNK_CONN_PROCESSING)

/** Remove the current request from the backlog
 *
 */
#define REQUEST_EXTRACT_BACKLOG(_treq) \
do { \
  int _ret; \
  _ret = fr_heap_extract(&(_treq)->pub.trunk->backlog, _treq); \
  if (!fr_cond_assert_msg(_ret == 0, "Failed extracting conn from backlog heap: %s", fr_strerror())) break; \
} while (0)

/** Remove the current request from the pending list
 *
 */
#define REQUEST_EXTRACT_PENDING(_treq) \
do { \
  int _ret; \
  _ret = fr_heap_extract(&(_treq)->pub.tconn->pending, _treq); \
  if (!fr_cond_assert_msg(_ret == 0, "Failed extracting conn from pending heap: %s", fr_strerror())) break; \
} while (0)

/** Remove the current request from the partial slot
 *
 */
#define REQUEST_EXTRACT_PARTIAL(_treq) \
do { \
  fr_assert((_treq)->pub.tconn->partial == treq); \
  tconn->partial = NULL; \
} while (0)

/** Remove the current request from the sent list
 *
 */
#define REQUEST_EXTRACT_SENT(_treq) fr_dlist_remove(&tconn->sent, treq)

/** Remove the current request from the reapable list
 *
 */
#define REQUEST_EXTRACT_REAPABLE(_treq) fr_dlist_remove(&tconn->reapable, treq)

/** Remove the current request from the cancel list
 *
 */
#define REQUEST_EXTRACT_CANCEL(_treq) fr_dlist_remove(&tconn->cancel, treq)

/** Remove the current request from the cancel_partial slot
 *
 */
#define REQUEST_EXTRACT_CANCEL_PARTIAL(_treq) \
do { \
  fr_assert((_treq)->pub.tconn->cancel_partial == treq); \
  tconn->cancel_partial = NULL; \
} while (0)

/** Remove the current request from the cancel sent list
 *
 */
#define REQUEST_EXTRACT_CANCEL_SENT(_treq) fr_dlist_remove(&tconn->cancel_sent, treq)

/** Reorder the connections in the active heap
 *
 * fr_heap_extract will also error out if heap_id is bad - no need for assert
 */
#define CONN_REORDER(_tconn) \
do { \
  int _ret; \
  if ((fr_minmax_heap_num_elements((_tconn)->pub.trunk->active) == 1)) break; \
  if (!fr_cond_assert((_tconn)->pub.state == TRUNK_CONN_ACTIVE)) break; \
  _ret = fr_minmax_heap_extract((_tconn)->pub.trunk->active, (_tconn)); \
  if (!fr_cond_assert_msg(_ret == 0, "Failed extracting conn from active heap: %s", fr_strerror())) break; \
  fr_minmax_heap_insert((_tconn)->pub.trunk->active, (_tconn)); \
} while (0)

DIAG_OFF(unused-function)

#define FR_TRUNK_LIST_FUNC(_list,_type) \
static inline CC_HINT(nonnull, always_inline) void trunk_list_ ## _list ## _add(trunk_t *trunk, _type *arg) \
{ \
  fr_dlist_insert_head(&trunk->_list, arg); \
} \
static inline CC_HINT(nonnull, always_inline) _type *trunk_list_ ## _list ##_peek(trunk_t *trunk) \
{ \
  return fr_dlist_tail(&trunk->_list); \
} \
static inline CC_HINT(nonnull, always_inline) _type *trunk_list_ ## _list ##_pop(trunk_t *trunk) \
{ \
  return fr_dlist_pop_head(&trunk->_list); \
} \
static inline CC_HINT(nonnull, always_inline) void trunk_list_ ## _list ##_remove(trunk_t *trunk, _type *arg) \
{ \
  fr_dlist_remove(&trunk->_list, arg); \
}

FR_TRUNK_LIST_FUNC(free_requests, trunk_request_t)
FR_TRUNK_LIST_FUNC(full, trunk_connection_t)
FR_TRUNK_LIST_FUNC(inactive, trunk_connection_t)
FR_TRUNK_LIST_FUNC(inactive_draining, trunk_connection_t)
FR_TRUNK_LIST_FUNC(draining, trunk_connection_t)

DIAG_ON(unused-function)

/** Call a list of watch functions associated with a state
 *
 */
static inline void trunk_watch_call(trunk_t *trunk, fr_dlist_head_t *list, trunk_state_t state)
{
  /*
   *  Nested watcher calls are not allowed
   *  and shouldn't be possible because of
   *  deferred signal processing.
   */
  fr_assert(trunk->next_watcher == NULL);

  while ((trunk->next_watcher = fr_dlist_next(list, trunk->next_watcher))) {
    trunk_watch_entry_t *entry = trunk->next_watcher;
    bool      oneshot = entry->oneshot; /* Watcher could be freed, so store now */

    if (!entry->enabled) continue;
    if (oneshot) trunk->next_watcher = fr_dlist_remove(list, entry);

    entry->func(trunk, trunk->pub.state, state, entry->uctx);

    if (oneshot) talloc_free(entry);
  }
  trunk->next_watcher = NULL;
}

/** Call the state change watch functions
 *
 */
#define CALL_WATCHERS(_trunk, _state) \
do { \
  if (fr_dlist_empty(&(_trunk)->watch[_state])) break; \
  trunk_watch_call((_trunk), &(_trunk)->watch[_state], _state); \
} while(0)

/** Remove a watch function from a trunk state list
 *
 * @param[in] trunk The trunk to remove the watcher from.
 * @param[in] state to remove the watch from.
 * @param[in] watch Function to remove.
 * @return
 *  - 0 if the function was removed successfully.
 *  - -1 if the function wasn't present in the watch list.
 *  - -2 if an invalid state was passed.
 */
int trunk_del_watch(trunk_t *trunk, trunk_state_t state, trunk_watch_t watch)
{
  trunk_watch_entry_t *entry = NULL;
  fr_dlist_head_t         *list;

  if (state >= TRUNK_STATE_MAX) return -2;

  list = &trunk->watch[state];
  while ((entry = fr_dlist_next(list, entry))) {
    if (entry->func == watch) {
      if (trunk->next_watcher == entry) {
        trunk->next_watcher = fr_dlist_remove(list, entry);
      } else {
        fr_dlist_remove(list, entry);
      }
      talloc_free(entry);
      return 0;
    }
  }

  return -1;
}

/** Add a watch entry to the trunk state list
 *
 * @param[in] trunk The trunk to add the watcher to.
 * @param[in] state to watch for.
 * @param[in] watch Function to add.
 * @param[in] oneshot Should this watcher only be run once.
 * @param[in] uctx  Context to pass to function.
 * @return
 *  - NULL if an invalid state is passed.
 *  - A new watch entry handle on success.
 */
trunk_watch_entry_t *trunk_add_watch(trunk_t *trunk, trunk_state_t state,
             trunk_watch_t watch, bool oneshot, void const *uctx)
{
  trunk_watch_entry_t *entry;
  fr_dlist_head_t   *list;

  if (state >= TRUNK_STATE_MAX) return NULL;

  list = &trunk->watch[state];
  MEM(entry = talloc_zero(trunk, trunk_watch_entry_t));

  entry->func = watch;
  entry->oneshot = oneshot;
  entry->enabled = true;
  memcpy(&entry->uctx, &uctx, sizeof(entry->uctx));
  fr_dlist_insert_tail(list, entry);

  return entry;
}

#define TRUNK_STATE_TRANSITION(_new) \
do { \
  DEBUG3("Trunk changed state %s -> %s", \
         fr_table_str_by_value(trunk_states, trunk->pub.state, "<INVALID>"), \
         fr_table_str_by_value(trunk_states, _new, "<INVALID>")); \
  CALL_WATCHERS(trunk, _new); \
  trunk->pub.state = _new; \
} while (0)

static void trunk_request_enter_backlog(trunk_request_t *treq, bool new);
static void trunk_request_enter_pending(trunk_request_t *treq, trunk_connection_t *tconn, bool new);
static void trunk_request_enter_partial(trunk_request_t *treq);
static void trunk_request_enter_sent(trunk_request_t *treq);
static void trunk_request_enter_reapable(trunk_request_t *treq);
static void trunk_request_enter_failed(trunk_request_t *treq);
static void trunk_request_enter_complete(trunk_request_t *treq);
static void trunk_request_enter_cancel(trunk_request_t *treq, trunk_cancel_reason_t reason);
static void trunk_request_enter_cancel_sent(trunk_request_t *treq);
static void trunk_request_enter_cancel_complete(trunk_request_t *treq);

static uint64_t trunk_requests_per_connection(uint16_t *conn_count_out, uint32_t *req_conn_out,
                trunk_t *trunk, fr_time_t now, NDEBUG_UNUSED bool verify);

static int trunk_connection_spawn(trunk_t *trunk, fr_time_t now);
static inline void trunk_connection_auto_full(trunk_connection_t *tconn);
static inline void trunk_connection_auto_unfull(trunk_connection_t *tconn);
static inline void trunk_connection_readable(trunk_connection_t *tconn);
static inline void trunk_connection_writable(trunk_connection_t *tconn);
static void trunk_connection_event_update(trunk_connection_t *tconn);
static void trunk_connection_enter_full(trunk_connection_t *tconn);
static void trunk_connection_enter_inactive(trunk_connection_t *tconn);
static void trunk_connection_enter_inactive_draining(trunk_connection_t *tconn);
static void trunk_connection_enter_draining(trunk_connection_t *tconn);
static void trunk_connection_enter_draining_to_free(trunk_connection_t *tconn);
static void trunk_connection_enter_active(trunk_connection_t *tconn);

static void trunk_rebalance(trunk_t *trunk);
static void trunk_manage(trunk_t *trunk, fr_time_t now);
static void _trunk_timer(fr_timer_list_t *tl, fr_time_t now, void *uctx);
static void trunk_backlog_drain(trunk_t *trunk);

/** Compare two protocol requests
 *
 * Allows protocol requests to be prioritised with a function
 * specified by the API client.  Defaults to by pointer address
 * if no function is specified.
 *
 * @param[in] a treq to compare to b.
 * @param[in] b treq to compare to a.
 * @return
 *  - +1 if a > b.
 *  - 0 if a == b.
 *  - -1 if a < b.
 */
static int8_t _trunk_request_prioritise(void const *a, void const *b)
{
  trunk_request_t const *treq_a = talloc_get_type_abort_const(a, trunk_request_t);
  trunk_request_t const *treq_b = talloc_get_type_abort_const(b, trunk_request_t);

  fr_assert(treq_a->pub.trunk == treq_b->pub.trunk);

  return treq_a->pub.trunk->funcs.request_prioritise(treq_a->pub.preq, treq_b->pub.preq);
}

/** Remove a request from all connection lists
 *
 * A common function used by init, fail, complete state functions to disassociate
 * a request from a connection in preparation for freeing or reassignment.
 *
 * Despite its unassuming name, this function is *the* place to put calls to
 * functions which need to be called when the number of requests associated with
 * a connection changes.
 *
 * Trunk requests will always be passed to this function before they're removed
 * from a connection, even if the requests are being freed.
 *
 * @param[in] treq  to trigger a state change for.
 */
static void trunk_request_remove_from_conn(trunk_request_t *treq)
{
  trunk_connection_t  *tconn = treq->pub.tconn;
  trunk_t   *trunk = treq->pub.trunk;

  if (!fr_cond_assert(!tconn || (tconn->pub.trunk == trunk))) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_UNASSIGNED:
    return; /* Not associated with connection */

  case TRUNK_REQUEST_STATE_PENDING:
    REQUEST_EXTRACT_PENDING(treq);
    break;

  case TRUNK_REQUEST_STATE_PARTIAL:
    REQUEST_EXTRACT_PARTIAL(treq);
    break;

  case TRUNK_REQUEST_STATE_SENT:
    REQUEST_EXTRACT_SENT(treq);
    break;

  case TRUNK_REQUEST_STATE_REAPABLE:
    REQUEST_EXTRACT_REAPABLE(treq);
    break;

  case TRUNK_REQUEST_STATE_CANCEL:
    REQUEST_EXTRACT_CANCEL(treq);
    break;

  case TRUNK_REQUEST_STATE_CANCEL_PARTIAL:
    REQUEST_EXTRACT_CANCEL_PARTIAL(treq);
    break;

  case TRUNK_REQUEST_STATE_CANCEL_SENT:
    REQUEST_EXTRACT_CANCEL_SENT(treq);
    break;

  default:
    fr_assert(0);
    break;
  }

  /*
   *  If the request wasn't associated with a
   *  connection, then there's nothing more
   *  to do.
   */
  if (!tconn) return;

  {
    request_t *request = treq->pub.request;

    ROPTIONAL(RDEBUG3, DEBUG3, "%s Trunk connection released request %" PRIu64,
        tconn->pub.conn->name, treq->id);
  }
  /*
   *  Release any connection specific resources the
   *  treq holds.
   */
  DO_REQUEST_CONN_RELEASE(treq);

  switch (tconn->pub.state){
  case TRUNK_CONN_FULL:
    trunk_connection_auto_unfull(tconn);    /* Check if we can switch back to active */
    if (tconn->pub.state == TRUNK_CONN_FULL) break; /* Only fallthrough if conn is now active */
    FALL_THROUGH;

  case TRUNK_CONN_ACTIVE:
    CONN_REORDER(tconn);
    break;

  default:
    break;
  }

  treq->pub.tconn = NULL;

  /*
   *  Request removed from the connection
   *  see if we need up deregister I/O events.
   */
  trunk_connection_event_update(tconn);
}

/** Transition a request to the unassigned state, in preparation for re-assignment
 *
 * @note treq->tconn may be inviable after calling
 *  if treq->conn and connection_signals_pause are not used.
 *  This is due to call to trunk_request_remove_from_conn.
 *
 * @param[in] treq  to trigger a state change for.
 */
static void trunk_request_enter_unassigned(trunk_request_t *treq)
{
  trunk_t   *trunk = treq->pub.trunk;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_UNASSIGNED:
    return;

  case TRUNK_REQUEST_STATE_BACKLOG:
    REQUEST_EXTRACT_BACKLOG(treq);
    break;

  case TRUNK_REQUEST_STATE_PENDING:
  case TRUNK_REQUEST_STATE_CANCEL:
  case TRUNK_REQUEST_STATE_CANCEL_PARTIAL:
  case TRUNK_REQUEST_STATE_CANCEL_SENT:
    trunk_request_remove_from_conn(treq);
    break;

  default:
    REQUEST_BAD_STATE_TRANSITION(TRUNK_REQUEST_STATE_UNASSIGNED);
  }

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_UNASSIGNED);
}

/** Transition a request to the backlog state, adding it to the backlog of the trunk
 *
 * @note treq->tconn and treq may be inviable after calling
 *  if treq->conn and connection_signals_pause are not used.
 *  This is due to call to trunk_manage.
 *
 * @param[in] treq  to trigger a state change for.
 * @param[in] new Whether this is a new request.
 */
static void trunk_request_enter_backlog(trunk_request_t *treq, bool new)
{
  trunk_connection_t  *tconn = treq->pub.tconn;
  trunk_t     *trunk = treq->pub.trunk;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_INIT:
  case TRUNK_REQUEST_STATE_UNASSIGNED:
    break;

  case TRUNK_REQUEST_STATE_PENDING:
    REQUEST_EXTRACT_PENDING(treq);
    break;

  case TRUNK_REQUEST_STATE_CANCEL:
    REQUEST_EXTRACT_CANCEL(treq);
    break;

  default:
    REQUEST_BAD_STATE_TRANSITION(TRUNK_REQUEST_STATE_BACKLOG);
  }

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_BACKLOG);
  fr_heap_insert(&trunk->backlog, treq);  /* Insert into the backlog heap */

  /*
   *  A new request has entered the trunk.
   *  Re-calculate request/connection ratios.
   */
  if (new) trunk_requests_per_connection(NULL, NULL, trunk, fr_time(), false);

  /*
   *  To reduce latency, if there's no connections
   *      in the connecting state, call the trunk manage
   *  function immediately.
   *
   *  Likewise, if there's draining connections
   *  which could be moved back to active call
   *  the trunk manage function.
   *
   *  Remember requests only enter the backlog if
   *  there's no connections which can service them.
   */
  if ((trunk_connection_count_by_state(treq->pub.trunk, TRUNK_CONN_CONNECTING) == 0) ||
      (trunk_connection_count_by_state(treq->pub.trunk, TRUNK_CONN_DRAINING) > 0)) {
    trunk_connection_manage_schedule(treq->pub.trunk);
  }
}

/** Transition a request to the pending state, adding it to the backlog of an active connection
 *
 * All trunk requests being added to a connection get passed to this function.
 * All trunk requests being removed from a connection get passed to #trunk_request_remove_from_conn.
 *
 * @note treq->tconn and treq may be inviable after calling
 *  if treq->conn and connection_signals_pause is not used.
 *  This is due to call to trunk_connection_event_update.
 *
 * @param[in] treq  to trigger a state change for.
 * @param[in] tconn to enqueue the request on.
 * @param[in] new Whether this is a new request.
 */
static void trunk_request_enter_pending(trunk_request_t *treq, trunk_connection_t *tconn, bool new)
{
  trunk_t   *trunk = treq->pub.trunk;

  fr_assert(tconn->pub.trunk == trunk);
  fr_assert(IS_PROCESSING(tconn));

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_INIT:
  case TRUNK_REQUEST_STATE_UNASSIGNED:
    fr_assert(!treq->pub.tconn);
    break;

  case TRUNK_REQUEST_STATE_BACKLOG:
    fr_assert(!treq->pub.tconn);
    REQUEST_EXTRACT_BACKLOG(treq);
    break;

  case TRUNK_REQUEST_STATE_CANCEL:  /* Moved from another connection */
    REQUEST_EXTRACT_CANCEL(treq);
    break;

  default:
    REQUEST_BAD_STATE_TRANSITION(TRUNK_REQUEST_STATE_PENDING);
  }

  /*
   *  Assign the new connection first this first so
   *      it appears in the state log.
   */
  treq->pub.tconn = tconn;

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_PENDING);

  {
    request_t *request = treq->pub.request;

    ROPTIONAL(RDEBUG, DEBUG3, "%s Trunk connection assigned request %"PRIu64,
        tconn->pub.conn->name, treq->id);
  }
  fr_heap_insert(&tconn->pending, treq);

  /*
   *  A new request has entered the trunk.
   *  Re-calculate request/connection ratios.
   */
  if (new) trunk_requests_per_connection(NULL, NULL, trunk, fr_time(), false);

  /*
   *  Check if we need to automatically transition the
   *  connection to full.
   */
  trunk_connection_auto_full(tconn);

  /*
   *  Reorder the connection in the heap now it has an
   *  additional request.
   */
  if (tconn->pub.state == TRUNK_CONN_ACTIVE) CONN_REORDER(tconn);

  /*
   *  We have a new request, see if we need to register
   *  for I/O events.
   */
  trunk_connection_event_update(tconn);
}

/** Transition a request to the partial state, indicating that is has been partially sent
 *
 * @param[in] treq  to trigger a state change for.
 */
static void trunk_request_enter_partial(trunk_request_t *treq)
{
  trunk_connection_t *tconn = treq->pub.tconn;
  trunk_t *trunk = treq->pub.trunk;

  if (!fr_cond_assert(!tconn || (tconn->pub.trunk == trunk))) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_PENDING: /* All requests go through pending, even requeued ones */
    REQUEST_EXTRACT_PENDING(treq);
    break;

  default:
    REQUEST_BAD_STATE_TRANSITION(TRUNK_REQUEST_STATE_PARTIAL);
  }

  fr_assert(!tconn->partial);
  tconn->partial = treq;

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_PARTIAL);
}

/** Transition a request to the sent state, indicating that it's been sent in its entirety
 *
 * @note treq->tconn and treq may be inviable after calling
 *  if treq->conn and connection_signals_pause is not used.
 *  This is due to call to trunk_connection_event_update.
 *
 * @param[in] treq  to trigger a state change for.
 */
static void trunk_request_enter_sent(trunk_request_t *treq)
{
  trunk_connection_t  *tconn = treq->pub.tconn;
  trunk_t     *trunk = treq->pub.trunk;

  if (!fr_cond_assert(!tconn || (tconn->pub.trunk == trunk))) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_PENDING:
    REQUEST_EXTRACT_PENDING(treq);
    break;

  case TRUNK_REQUEST_STATE_PARTIAL:
    REQUEST_EXTRACT_PARTIAL(treq);
    break;

  default:
    REQUEST_BAD_STATE_TRANSITION(TRUNK_REQUEST_STATE_SENT);
  }

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_SENT);
  fr_dlist_insert_tail(&tconn->sent, treq);

  /*
   *  Update the connection's sent stats if this is the
   *  first time this request is being sent.
   */
  if (!treq->sent) {
    trunk->pub.last_write_success = fr_time();

    tconn->pub.last_write_success = trunk->pub.last_write_success;
    tconn->sent_count++;
    treq->sent = true;

    /*
     *  Enforces max_uses
     */
    if ((trunk->conf.max_uses > 0) && (tconn->sent_count >= trunk->conf.max_uses)) {
      DEBUG3("Trunk hit max uses %" PRIu64 " at %d", trunk->conf.max_uses, __LINE__);
      trunk_connection_enter_draining_to_free(tconn);
    }
  }

  /*
   *  We just sent a request, we probably need
   *  to tell the event loop we want to be
   *  notified if there's data available.
   */
  trunk_connection_event_update(tconn);
}

/** Transition a request to the reapable state, indicating that it's been sent in its entirety, but no response is expected
 *
 * @note Largely a replica of trunk_request_enter_sent.
 *
 * @param[in] treq  to trigger a state change for.
 */
static void trunk_request_enter_reapable(trunk_request_t *treq)
{
  trunk_connection_t  *tconn = treq->pub.tconn;
  trunk_t     *trunk = treq->pub.trunk;

  if (!fr_cond_assert(!tconn || (tconn->pub.trunk == trunk))) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_PENDING:
    REQUEST_EXTRACT_PENDING(treq);
    break;

  case TRUNK_REQUEST_STATE_PARTIAL:
    REQUEST_EXTRACT_PARTIAL(treq);
    break;

  default:
    REQUEST_BAD_STATE_TRANSITION(TRUNK_REQUEST_STATE_REAPABLE);
  }

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_REAPABLE);
  fr_dlist_insert_tail(&tconn->reapable, treq);

  if (!treq->sent) {
    tconn->sent_count++;
    treq->sent = true;

    if ((trunk->conf.max_uses > 0) && (tconn->sent_count >= trunk->conf.max_uses)) {
      DEBUG3("Trunk hit max uses %" PRIu64 " at %d", trunk->conf.max_uses, __LINE__);
      trunk_connection_enter_draining_to_free(tconn);
    }
  }

  trunk_connection_event_update(tconn);
}

/** Transition a request to the cancel state, placing it in a connection's cancellation list
 *
 * If a request_cancel_send callback is provided, that callback will
 * be called periodically for requests which were cancelled due to
 * a signal.
 *
 * The request_cancel_send callback will dequeue cancelled requests
 * and inform a remote server that the result is no longer required.
 *
 * A request must enter this state before being added to the backlog
 * of another connection if it's been sent or partially sent.
 *
 * @note treq->tconn and treq may be inviable after calling
 *  if treq->conn and connection_signals_pause is not used.
 *  This is due to call to trunk_connection_event_update.
 *
 * @param[in] treq  to trigger a state change for.
 * @param[in] reason  Why the request was cancelled.
 *      Should be one of:
 *      - TRUNK_CANCEL_REASON_SIGNAL request cancelled
 *        because of a signal from the interpreter.
 *      - TRUNK_CANCEL_REASON_MOVE request cancelled
 *        because the connection failed and it needs
 *        to be assigned to a new connection.
 *      - TRUNK_CANCEL_REASON_REQUEUE request cancelled
 *        as it needs to be resent on the same connection.
 */
static void trunk_request_enter_cancel(trunk_request_t *treq, trunk_cancel_reason_t reason)
{
  trunk_connection_t  *tconn = treq->pub.tconn;
  trunk_t     *trunk = treq->pub.trunk;

  if (!fr_cond_assert(!tconn || (tconn->pub.trunk == trunk))) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_PARTIAL:
    REQUEST_EXTRACT_PARTIAL(treq);
    break;

  case TRUNK_REQUEST_STATE_SENT:
    REQUEST_EXTRACT_SENT(treq);
    break;

  case TRUNK_REQUEST_STATE_REAPABLE:
    REQUEST_EXTRACT_REAPABLE(treq);
    break;

  default:
    REQUEST_BAD_STATE_TRANSITION(TRUNK_REQUEST_STATE_CANCEL);
  }

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_CANCEL);
  fr_dlist_insert_tail(&tconn->cancel, treq);
  treq->cancel_reason = reason;

  DO_REQUEST_CANCEL(treq, reason);

  /*
   *  Our treq is no longer bound to an actual
   *      request_t *, as we can't guarantee the
   *  lifetime of the original request_t *.
   */
  if (treq->cancel_reason == TRUNK_CANCEL_REASON_SIGNAL) treq->pub.request = NULL;

  /*
   *  Register for I/O write events if we need to.
   */
  trunk_connection_event_update(treq->pub.tconn);
}

/** Transition a request to the cancel_partial state, placing it in a connection's cancel_partial slot
 *
 * The request_demux function is then responsible for signalling
 * that the cancel request is complete when the remote server
 * acknowledges the cancellation request.
 *
 * @param[in] treq  to trigger a state change for.
 */
static void trunk_request_enter_cancel_partial(trunk_request_t *treq)
{
  trunk_connection_t  *tconn = treq->pub.tconn;
  trunk_t     *trunk = treq->pub.trunk;

  if (!fr_cond_assert(!tconn || (tconn->pub.trunk == trunk))) return;
  fr_assert(trunk->funcs.request_cancel_mux);
  fr_assert(treq->cancel_reason == TRUNK_CANCEL_REASON_SIGNAL);

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_CANCEL:  /* The only valid state cancel_sent can be reached from */
    REQUEST_EXTRACT_CANCEL(treq);
    break;

  default:
    REQUEST_BAD_STATE_TRANSITION(TRUNK_REQUEST_STATE_CANCEL_PARTIAL);
  }

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_CANCEL_PARTIAL);
  fr_assert(!tconn->cancel_partial);
  tconn->cancel_partial = treq;
}

/** Transition a request to the cancel_sent state, placing it in a connection's cancel_sent list
 *
 * The request_demux function is then responsible for signalling
 * that the cancel request is complete when the remote server
 * acknowledges the cancellation request.
 *
 * @note treq->tconn and treq may be inviable after calling
 *  if treq->conn and connection_signals_pause is not used.
 *  This is due to call to trunk_connection_event_update.
 *
 * @param[in] treq  to trigger a state change for.
 */
static void trunk_request_enter_cancel_sent(trunk_request_t *treq)
{
  trunk_connection_t  *tconn = treq->pub.tconn;
  trunk_t     *trunk = treq->pub.trunk;

  if (!fr_cond_assert(!tconn || (tconn->pub.trunk == trunk))) return;
  fr_assert(trunk->funcs.request_cancel_mux);
  fr_assert(treq->cancel_reason == TRUNK_CANCEL_REASON_SIGNAL);

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_CANCEL_PARTIAL:
    REQUEST_EXTRACT_CANCEL_PARTIAL(treq);
    break;

  case TRUNK_REQUEST_STATE_CANCEL:
    REQUEST_EXTRACT_CANCEL(treq);
    break;

  default:
    REQUEST_BAD_STATE_TRANSITION(TRUNK_REQUEST_STATE_CANCEL_SENT);
  }

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_CANCEL_SENT);
  fr_dlist_insert_tail(&tconn->cancel_sent, treq);

  /*
   *  De-register for I/O write events
   *  and register the read events
   *  to drain the cancel ACKs.
   */
  trunk_connection_event_update(treq->pub.tconn);
}

/** Cancellation was acked, the request is complete, free it
 *
 * The API client will not be informed, as the original request_t *
 * will likely have been freed by this point.
 *
 * @note treq will be inviable after a call to this function.
 *      treq->tconn may be inviable after calling
 *  if treq->conn and connection_signals_pause is not used.
 *  This is due to call to trunk_request_remove_from_conn.
 *
 * @param[in] treq  to mark as complete.
 */
static void trunk_request_enter_cancel_complete(trunk_request_t *treq)
{
  trunk_connection_t  *tconn = treq->pub.tconn;
  trunk_t     *trunk = treq->pub.trunk;

  if (!fr_cond_assert(!tconn || (tconn->pub.trunk == trunk))) return;
  if (!fr_cond_assert(!treq->pub.request)) return; /* Only a valid state for request_t * which have been cancelled */

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_CANCEL_SENT:
  case TRUNK_REQUEST_STATE_CANCEL:
    break;

  default:
    REQUEST_BAD_STATE_TRANSITION(TRUNK_REQUEST_STATE_CANCEL_COMPLETE);
  }

  trunk_request_remove_from_conn(treq);

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_CANCEL_COMPLETE);
  trunk_request_free(&treq);  /* Free the request */
}

/** Request completed successfully, inform the API client and free the request
 *
 * @note treq will be inviable after a call to this function.
 *  treq->tconn may also be inviable due to call to
 *  trunk_request_remove_from_conn.
 *
 * @param[in] treq  to mark as complete.
 */
static void trunk_request_enter_complete(trunk_request_t *treq)
{
  trunk_connection_t  *tconn = treq->pub.tconn;
  trunk_t     *trunk = treq->pub.trunk;

  if (!fr_cond_assert(!tconn || (tconn->pub.trunk == trunk))) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_SENT:
  case TRUNK_REQUEST_STATE_PENDING:
  case TRUNK_REQUEST_STATE_REAPABLE:
    trunk_request_remove_from_conn(treq);
    break;

  default:
    REQUEST_BAD_STATE_TRANSITION(TRUNK_REQUEST_STATE_COMPLETE);
  }

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_COMPLETE);
  DO_REQUEST_COMPLETE(treq);
  trunk_request_free(&treq);  /* Free the request */
}

/** Request failed, inform the API client and free the request
 *
 * @note treq will be inviable after a call to this function.
 *  treq->tconn may also be inviable due to call to
 *  trunk_request_remove_from_conn.
 *
 * @param[in] treq  to mark as failed.
 */
static void trunk_request_enter_failed(trunk_request_t *treq)
{
  trunk_connection_t  *tconn = treq->pub.tconn;
  trunk_t     *trunk = treq->pub.trunk;
  trunk_request_state_t prev = treq->pub.state;

  if (!fr_cond_assert(!tconn || (tconn->pub.trunk == trunk))) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_BACKLOG:
    REQUEST_EXTRACT_BACKLOG(treq);
    break;

  default:
    trunk_request_remove_from_conn(treq);
    break;
  }

  REQUEST_STATE_TRANSITION(TRUNK_REQUEST_STATE_FAILED);
  DO_REQUEST_FAIL(treq, prev);
  trunk_request_free(&treq);  /* Free the request */
}

/** Check to see if a trunk request can be enqueued
 *
 * @param[out] tconn_out  Connection the request may be enqueued on.
 * @param[in] trunk   To enqueue requests on.
 * @param[in] request   associated with the treq (if any).
 * @return
 *  - TRUNK_ENQUEUE_OK      caller should enqueue request on provided tconn.
 *  - TRUNK_ENQUEUE_IN_BACKLOG    Request should be queued in the backlog.
 *  - TRUNK_ENQUEUE_NO_CAPACITY   Unable to enqueue request as we have no spare
 *            connections or backlog space.
 *  - TRUNK_ENQUEUE_DST_UNAVAILABLE Can't enqueue because the destination is
 *            unreachable.
 */
static trunk_enqueue_t trunk_request_check_enqueue(trunk_connection_t **tconn_out, trunk_t *trunk,
                  request_t *request)
{
  trunk_connection_t  *tconn;
  /*
   *  If we have an active connection then
   *  return that.
   */
  tconn = fr_minmax_heap_min_peek(trunk->active);
  if (tconn) {
    *tconn_out = tconn;
    return TRUNK_ENQUEUE_OK;
  }

  /*
   *  Unlike the connection pool, we don't need
   *  to drive any internal processes by feeding
   *  it requests.
   *
   *  If the last event to occur was a failure
   *  we refuse to enqueue new requests until
   *  one or more connections comes online.
   */
  if (!trunk->conf.backlog_on_failed_conn &&
      fr_time_gt(trunk->pub.last_failed, fr_time_wrap(0)) &&
      fr_time_lt(trunk->pub.last_connected, trunk->pub.last_failed)) {
        RATE_LIMIT_LOCAL_ROPTIONAL(&trunk->limit_last_failure_log,
             RWARN, WARN, "Refusing to enqueue requests - "
             "No active connections and last event was a connection failure");

    return TRUNK_ENQUEUE_DST_UNAVAILABLE;
  }


  /*
   *  Only enforce if we're limiting maximum
   *  number of connections, and maximum
   *  number of requests per connection.
   */
  if (trunk->conf.max_req_per_conn && trunk->conf.max) {
    uint64_t  limit;

    limit = trunk->conf.max * (uint64_t)trunk->conf.max_req_per_conn;
    if (limit > 0) {
      uint64_t  total_reqs;

      total_reqs = trunk_request_count_by_state(trunk, TRUNK_CONN_ALL,
                     TRUNK_REQUEST_STATE_ALL);
      if (total_reqs >= (limit + trunk->conf.max_backlog)) {
        RATE_LIMIT_LOCAL_ROPTIONAL(&trunk->limit_max_requests_alloc_log,
                 RWARN, WARN, "Refusing to alloc requests - "
                 "Limit of %"PRIu64" (max = %u * per_connection_max = %u) "
                 "plus %u backlog requests reached",
                 limit, trunk->conf.max, trunk->conf.max_req_per_conn,
                 trunk->conf.max_backlog);
        return TRUNK_ENQUEUE_NO_CAPACITY;
      }
    }
  }

  return TRUNK_ENQUEUE_IN_BACKLOG;
}

/** Enqueue a request which has never been assigned to a connection or was previously cancelled
 *
 * @param[in] treq  to re enqueue.  Must have been removed
 *      from its existing connection with
 *      #trunk_connection_requests_dequeue.
 * @return
 *  - TRUNK_ENQUEUE_OK      Request was re-enqueued.
 *  - TRUNK_ENQUEUE_NO_CAPACITY   Request enqueueing failed because we're at capacity.
 *  - TRUNK_ENQUEUE_DST_UNAVAILABLE Enqueuing failed for some reason.
 *                Usually because the connection to the resource is down.
 */
static trunk_enqueue_t trunk_request_enqueue_existing(trunk_request_t *treq)
{
  trunk_t     *trunk = treq->pub.trunk;
  trunk_connection_t  *tconn = NULL;
  trunk_enqueue_t   ret;

  /*
   *  Must *NOT* still be assigned to another connection
   */
  fr_assert(!treq->pub.tconn);

  ret = trunk_request_check_enqueue(&tconn, trunk, treq->pub.request);
  switch (ret) {
  case TRUNK_ENQUEUE_OK:
    if (trunk->conf.always_writable) {
      connection_signals_pause(tconn->pub.conn);
      trunk_request_enter_pending(treq, tconn, false);
      trunk_connection_writable(tconn);
      connection_signals_resume(tconn->pub.conn);
    } else {
      trunk_request_enter_pending(treq, tconn, false);
    }
    break;

  case TRUNK_ENQUEUE_IN_BACKLOG:
    /*
     *  No more connections and request
     *  is already in the backlog.
     *
     *  Signal our caller it should stop
     *  trying to drain the backlog.
     */
    if (treq->pub.state == TRUNK_REQUEST_STATE_BACKLOG) return TRUNK_ENQUEUE_NO_CAPACITY;
    trunk_request_enter_backlog(treq, false);
    break;

  default:
    break;
  }

  return ret;
}

/** Shift requests in the specified states onto new connections
 *
 * This function will blindly dequeue any requests in the specified state and get
 * them back to the unassigned state, cancelling any sent or partially sent requests.
 *
 * This function does not check that dequeuing a request in a particular state is a
 * sane or sensible thing to do, that's up to the caller!
 *
 * @param[out] out  A list to insert the newly dequeued and unassigned
 *      requests into.
 * @param[in] tconn to dequeue requests from.
 * @param[in] states  Dequeue request in these states.
 * @param[in] max The maximum number of requests to dequeue. 0 for unlimited.
 */
static uint64_t trunk_connection_requests_dequeue(fr_dlist_head_t *out, trunk_connection_t *tconn,
              int states, uint64_t max)
{
  trunk_request_t *treq;
  uint64_t  count = 0;

  if (max == 0) max = UINT64_MAX;

#define OVER_MAX_CHECK if (++count > max) return (count - 1)

#define DEQUEUE_ALL(_src_list, _state) do { \
  while ((treq = fr_dlist_head(_src_list))) { \
    OVER_MAX_CHECK; \
    fr_assert(treq->pub.state == (_state)); \
    trunk_request_enter_unassigned(treq); \
    fr_dlist_insert_tail(out, treq); \
  } } while (0)

  /*
   *  Don't need to do anything with
   *  cancellation requests.
   */
  if (states & TRUNK_REQUEST_STATE_CANCEL) DEQUEUE_ALL(&tconn->cancel,
                TRUNK_REQUEST_STATE_CANCEL);

  /*
   *  ...same with cancel inform
   */
  if (states & TRUNK_REQUEST_STATE_CANCEL_SENT) DEQUEUE_ALL(&tconn->cancel_sent,
                     TRUNK_REQUEST_STATE_CANCEL_SENT);

  /*
   *  ....same with cancel partial
   */
  if (states & TRUNK_REQUEST_STATE_CANCEL_PARTIAL) {
    OVER_MAX_CHECK;
    treq = tconn->cancel_partial;
    if (treq) {
      fr_assert(treq->pub.state == TRUNK_REQUEST_STATE_CANCEL_PARTIAL);
      trunk_request_enter_unassigned(treq);
      fr_dlist_insert_tail(out, treq);
    }
  }

  /*
   *  ...and pending.
   */
  if (states & TRUNK_REQUEST_STATE_PENDING) {
    while ((treq = fr_heap_peek(tconn->pending))) {
      OVER_MAX_CHECK;
      fr_assert(treq->pub.state == TRUNK_REQUEST_STATE_PENDING);
      trunk_request_enter_unassigned(treq);
      fr_dlist_insert_tail(out, treq);
    }
  }

  /*
   *  Cancel partially sent requests
   */
  if (states & TRUNK_REQUEST_STATE_PARTIAL) {
    OVER_MAX_CHECK;
    treq = tconn->partial;
    if (treq) {
      fr_assert(treq->pub.state == TRUNK_REQUEST_STATE_PARTIAL);

      /*
       *  Don't allow the connection to change state whilst
       *  we're draining requests from it.
       */
      connection_signals_pause(tconn->pub.conn);
      trunk_request_enter_cancel(treq, TRUNK_CANCEL_REASON_MOVE);
      trunk_request_enter_unassigned(treq);
      fr_dlist_insert_tail(out, treq);
      connection_signals_resume(tconn->pub.conn);
    }
  }

  /*
   *  Cancel sent requests
   */
  if (states & TRUNK_REQUEST_STATE_SENT) {
    /*
     *  Don't allow the connection to change state whilst
     *  we're draining requests from it.
     */
    connection_signals_pause(tconn->pub.conn);
    while ((count < max) && (treq = fr_dlist_head(&tconn->sent))) {
      fr_assert(treq->pub.state == TRUNK_REQUEST_STATE_SENT);

      trunk_request_enter_cancel(treq, TRUNK_CANCEL_REASON_MOVE);
      trunk_request_enter_unassigned(treq);
      fr_dlist_insert_tail(out, treq);
      count++;
    }
    connection_signals_resume(tconn->pub.conn);
  }

  return count;
}

/** Remove requests in specified states from a connection, attempting to distribute them to new connections
 *
 * @param[in] tconn   To remove requests from.
 * @param[in] states    One or more states or'd together.
 * @param[in] max   The maximum number of requests to dequeue.
 *        0 for unlimited.
 * @param[in] fail_bound  If true causes any requests bound to the connection to fail.
 *            If false bound requests will not be moved.
 *
 * @return the number of requests re-queued.
 */
static uint64_t trunk_connection_requests_requeue_priv(trunk_connection_t *tconn, int states, uint64_t max, bool fail_bound)
{
  trunk_t     *trunk = tconn->pub.trunk;
  fr_dlist_head_t   to_process;
  trunk_request_t   *treq = NULL;
  uint64_t    moved = 0;

  if (max == 0) max = UINT64_MAX;

  fr_dlist_talloc_init(&to_process, trunk_request_t, entry);

  /*
   *  Prevent the connection changing state whilst we're
   *  working with it.
   *
   *  There's a user callback that can be called by
   *  trunk_request_enqueue_existing which can reconnect
   *  the connection.
   */
  connection_signals_pause(tconn->pub.conn);

  /*
   *  Remove non-cancelled requests from the connection
   */
  moved += trunk_connection_requests_dequeue(&to_process, tconn, states & ~TRUNK_REQUEST_STATE_CANCEL_ALL, max);

  /*
   *  Prevent requests being requeued on the same trunk
   *  connection, which would break rebalancing.
   *
   *  This is a bit of a hack, but nothing should test
   *  for connection/list consistency in this code,
   *      and if something is added later, it'll be flagged
   *  by the tests.
   */
  if (tconn->pub.state == TRUNK_CONN_ACTIVE) {
    int ret;

    ret = fr_minmax_heap_extract(trunk->active, tconn);
    if (!fr_cond_assert_msg(ret == 0,
          "Failed extracting conn from active heap: %s", fr_strerror())) goto done;

  }

  /*
   *  Loop over all the requests we gathered and
   *  redistribute them to new connections.
   */
  while ((treq = fr_dlist_next(&to_process, treq))) {
    trunk_request_t *prev;

    prev = fr_dlist_remove(&to_process, treq);

    /*
     *  Attempts to re-queue a request
     *  that's bound to a connection
     *  results in a failure.
     */
    if (treq->bound_to_conn) {
      if (fail_bound || !IS_SERVICEABLE(tconn)) {
        trunk_request_enter_failed(treq);
      } else {
        trunk_request_enter_pending(treq, tconn, false);
      }
      goto next;
    }

    switch (trunk_request_enqueue_existing(treq)) {
    case TRUNK_ENQUEUE_OK:
      break;

    /*
     *  A connection failed, and
     *  there's no other connections
     *  available to deal with the
     *  load, it's been placed back
     *  in the backlog.
     */
    case TRUNK_ENQUEUE_IN_BACKLOG:
      break;

    /*
     *  If we fail to re-enqueue then
     *  there's nothing to do except
     *  fail the request.
     */
    case TRUNK_ENQUEUE_DST_UNAVAILABLE:
    case TRUNK_ENQUEUE_NO_CAPACITY:
    case TRUNK_ENQUEUE_FAIL:
      trunk_request_enter_failed(treq);
      break;
    }
  next:
    treq = prev;
  }

  /*
   *  Add the connection back into the active list
   */
  if (tconn->pub.state == TRUNK_CONN_ACTIVE) {
    int ret;

    ret = fr_minmax_heap_insert(trunk->active, tconn);
    if (!fr_cond_assert_msg(ret == 0,
                "Failed re-inserting conn into active heap: %s", fr_strerror())) goto done;
  }
  if (moved >= max) goto done;

  /*
   *  Deal with the cancelled requests specially we can't
   *      queue them up again as they were only valid on that
   *  specific connection.
   *
   *  We just need to run them to completion which, as
   *  they should already be in the unassigned state,
   *  just means freeing them.
   */
  moved += trunk_connection_requests_dequeue(&to_process, tconn,
               states & TRUNK_REQUEST_STATE_CANCEL_ALL, max - moved);
  while ((treq = fr_dlist_next(&to_process, treq))) {
    trunk_request_t *prev;

    prev = fr_dlist_remove(&to_process, treq);
    trunk_request_free(&treq);
    treq = prev;
  }

done:

  /*
   *  Always re-calculate the request/connection
   *  ratio at the end.
   *
   *  This avoids having the state transition
   *  functions do it.
   *
   *  The ratio would be wrong when they calculated
   *  it anyway, because a bunch of requests are
   *  dequeued from the connection and temporarily
   *  cease to exist from the perspective of the
   *  trunk_requests_per_connection code.
   */
  trunk_requests_per_connection(NULL, NULL, trunk, fr_time(), false);

  connection_signals_resume(tconn->pub.conn);
  return moved;
}

/** Move requests off of a connection and requeue elsewhere
 *
 * @note We don't re-queue on draining or draining to free, as requests should have already been
 *   moved off of the connection.  It's also dangerous as the trunk management code main
 *   clean up a connection in this state when it's run on re-queue, and then the caller
 *   may try and access a now freed connection.
 *
 * @param[in] tconn   to move requests off of.
 * @param[in] states    Only move requests in this state.
 * @param[in] max   The maximum number of requests to dequeue. 0 for unlimited.
 * @param[in] fail_bound  If true causes any requests bound to the connection to fail.
 *            If false bound requests will not be moved.
 * @return The number of requests requeued.
 */
uint64_t trunk_connection_requests_requeue(trunk_connection_t *tconn, int states, uint64_t max, bool fail_bound)
{
  switch (tconn->pub.state) {
  case TRUNK_CONN_ACTIVE:
  case TRUNK_CONN_FULL:
  case TRUNK_CONN_INACTIVE:
    return trunk_connection_requests_requeue_priv(tconn, states, max, fail_bound);

  default:
    return 0;
  }
}

/** Signal a partial write
 *
 * Where there's high load, and the outbound write buffer is full
 *
 * @param[in] treq  to signal state change for.
 */
void trunk_request_signal_partial(trunk_request_t *treq)
{
  if (!fr_cond_assert_msg(treq->pub.trunk, "treq not associated with trunk")) return;

  if (!fr_cond_assert_msg(IN_REQUEST_MUX(treq->pub.trunk),
        "%s can only be called from within request_mux handler", __FUNCTION__)) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_PENDING:
    trunk_request_enter_partial(treq);
    break;

  default:
    return;
  }
}

/** Signal that the request was written to a connection successfully
 *
 * @param[in] treq  to signal state change for.
 */
void trunk_request_signal_sent(trunk_request_t *treq)
{
  if (!fr_cond_assert_msg(treq->pub.trunk, "treq not associated with trunk")) return;

  if (!fr_cond_assert_msg(IN_REQUEST_MUX(treq->pub.trunk),
        "%s can only be called from within request_mux handler", __FUNCTION__)) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_PENDING:
  case TRUNK_REQUEST_STATE_PARTIAL:
    trunk_request_enter_sent(treq);
    break;

  default:
    return;
  }
}

/** Signal that the request was written to a connection successfully, but no response is expected
 *
 * @param[in] treq  to signal state change for.
 */
void trunk_request_signal_reapable(trunk_request_t *treq)
{
  if (!fr_cond_assert_msg(treq->pub.trunk, "treq not associated with trunk")) return;

  if (!fr_cond_assert_msg(IN_REQUEST_MUX(treq->pub.trunk),
        "%s can only be called from within request_mux handler", __FUNCTION__)) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_PENDING:
  case TRUNK_REQUEST_STATE_PARTIAL:
    trunk_request_enter_reapable(treq);
    break;

  default:
    return;
  }
}

/** Signal that a trunk request is complete
 *
 * The API client will be informed that the request is now complete.
 */
void trunk_request_signal_complete(trunk_request_t *treq)
{
  trunk_t *trunk = treq->pub.trunk;

  if (!fr_cond_assert_msg(trunk, "treq not associated with trunk")) return;

  /*
   *  We assume that if the request is being signalled
   *  as complete from the demux function, that it was
   *  a successful read.
   *
   *  If this assumption turns out to be incorrect
   *  then we need to add an argument to signal_complete
   *  to indicate if this is a successful read.
   */
  if (IN_REQUEST_DEMUX(trunk)) {
    trunk_connection_t *tconn = treq->pub.tconn;

    trunk->pub.last_read_success = fr_time();
    tconn->pub.last_read_success = trunk->pub.last_read_success;
  }

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_SENT:
  case TRUNK_REQUEST_STATE_PENDING: /* Got immediate response, i.e. cached */
  case TRUNK_REQUEST_STATE_REAPABLE:
    trunk_request_enter_complete(treq);
    break;

  default:
    return;
  }
}

/** Signal that a trunk request failed
 *
 * The API client will be informed that the request has failed.
 */
void trunk_request_signal_fail(trunk_request_t *treq)
{
  if (!fr_cond_assert_msg(treq->pub.trunk, "treq not associated with trunk")) return;

  trunk_request_enter_failed(treq);
}

/** Cancel a trunk request
 *
 * treq can be in any state, but requests to cancel if the treq is not in
 * the TRUNK_REQUEST_STATE_PARTIAL or TRUNK_REQUEST_STATE_SENT state will be ignored.
 *
 * The complete or failed callbacks will not be called here, as it's assumed the request_t *
 * is now inviable as it's being cancelled.
 *
 * The free function however, is called, and that should be used to perform necessary
 * cleanup.
 *
 * @param[in] treq  to signal state change for.
 */
void trunk_request_signal_cancel(trunk_request_t *treq)
{
  trunk_t *trunk;

  /*
   *  Ensure treq hasn't been freed
   */
  (void)talloc_get_type_abort(treq, trunk_request_t);

  if (!fr_cond_assert_msg(treq->pub.trunk, "treq not associated with trunk")) return;

  if (!fr_cond_assert_msg(!IN_HANDLER(treq->pub.trunk),
        "%s cannot be called within a handler", __FUNCTION__)) return;

  trunk = treq->pub.trunk;

  switch (treq->pub.state) {
  /*
   *  We don't call the complete or failed callbacks
   *  as the request and rctx are no longer viable.
   */
  case TRUNK_REQUEST_STATE_PARTIAL:
  case TRUNK_REQUEST_STATE_SENT:
  {
    trunk_connection_t *tconn = treq->pub.tconn;

    /*
     *  Don't allow connection state changes
     */
    connection_signals_pause(tconn->pub.conn);
    trunk_request_enter_cancel(treq, TRUNK_CANCEL_REASON_SIGNAL);
    if (!fr_cond_assert_msg(treq->pub.state == TRUNK_REQUEST_STATE_CANCEL,
          "Bad state %s after cancellation",
          fr_table_str_by_value(trunk_request_states, treq->pub.state, "<INVALID>"))) {
      connection_signals_resume(tconn->pub.conn);
      return;
    }
    /*
     *  No cancel muxer.  We're done.
     *
     *  If we do have a cancel mux function,
     *  the next time this connection becomes
     *  writable, we'll call the cancel mux
     *  function.
     *
     *  We don't run the complete or failed
     *  callbacks here as the request is
     *  being cancelled.
     */
    if (!trunk->funcs.request_cancel_mux) {
      trunk_request_enter_unassigned(treq);
      trunk_request_free(&treq);
    }
    connection_signals_resume(tconn->pub.conn);
  }
    break;

  /*
   *  We're already in the process of cancelling a
   *  request, so ignore duplicate signals.
   */
  case TRUNK_REQUEST_STATE_CANCEL:
  case TRUNK_REQUEST_STATE_CANCEL_PARTIAL:
  case TRUNK_REQUEST_STATE_CANCEL_SENT:
  case TRUNK_REQUEST_STATE_CANCEL_COMPLETE:
    break;

  /*
   *  For any other state, we just release the request
   *  from its current connection and free it.
   */
  default:
    trunk_request_enter_unassigned(treq);
    trunk_request_free(&treq);
    break;
  }
}

/** Signal a partial cancel write
 *
 * Where there's high load, and the outbound write buffer is full
 *
 * @param[in] treq  to signal state change for.
 */
void trunk_request_signal_cancel_partial(trunk_request_t *treq)
{
  if (!fr_cond_assert_msg(treq->pub.trunk, "treq not associated with trunk")) return;

  if (!fr_cond_assert_msg(IN_REQUEST_CANCEL_MUX(treq->pub.trunk),
        "%s can only be called from within request_cancel_mux handler", __FUNCTION__)) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_CANCEL:
    trunk_request_enter_cancel_partial(treq);
    break;

  default:
    return;
  }
}

/** Signal that a remote server has been notified of the cancellation
 *
 * Called from request_cancel_mux to indicate that the datastore has been informed
 * that the response is no longer needed.
 *
 * @param[in] treq  to signal state change for.
 */
void trunk_request_signal_cancel_sent(trunk_request_t *treq)
{
  if (!fr_cond_assert_msg(treq->pub.trunk, "treq not associated with trunk")) return;

  if (!fr_cond_assert_msg(IN_REQUEST_CANCEL_MUX(treq->pub.trunk),
        "%s can only be called from within request_cancel_mux handler", __FUNCTION__)) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_CANCEL:
  case TRUNK_REQUEST_STATE_CANCEL_PARTIAL:
    trunk_request_enter_cancel_sent(treq);
    break;

  default:
    break;
  }
}

/** Signal that a remote server acked our cancellation
 *
 * Called from request_demux to indicate that it got an ack for the cancellation.
 *
 * @param[in] treq  to signal state change for.
 */
void trunk_request_signal_cancel_complete(trunk_request_t *treq)
{
  if (!fr_cond_assert_msg(treq->pub.trunk, "treq not associated with trunk")) return;

  if (!fr_cond_assert_msg(IN_REQUEST_DEMUX(treq->pub.trunk) || IN_REQUEST_CANCEL_MUX(treq->pub.trunk),
        "%s can only be called from within request_demux or request_cancel_mux handlers",
        __FUNCTION__)) return;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_CANCEL_SENT:
  /*
   *  This is allowed, as we may not need to wait
   *  for the database to ACK our cancellation
   *  request.
   *
   *  Note: TRUNK_REQUEST_STATE_CANCEL_PARTIAL
   *  is not allowed here, as that'd mean we'd half
   *  written the cancellation request out to the
   *  socket, and then decided to abandon it.
   *
   *  That'd leave the socket in an unusable state.
   */
  case TRUNK_REQUEST_STATE_CANCEL:
    trunk_request_enter_cancel_complete(treq);
    break;

  default:
    break;
  }
}

/** If the trunk request is freed then update the target requests
 *
 * gperftools showed calling the request free function directly was slightly faster
 * than using talloc_free.
 *
 * @param[in] treq_to_free  request.
 */
void trunk_request_free(trunk_request_t **treq_to_free)
{
  trunk_request_t *treq = *treq_to_free;
  trunk_t   *trunk;

  if (unlikely(!treq)) return;

  trunk = treq->pub.trunk;

  /*
   *  The only valid states a trunk request can be
   *  freed from.
   */
  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_INIT:
  case TRUNK_REQUEST_STATE_UNASSIGNED:
  case TRUNK_REQUEST_STATE_COMPLETE:
  case TRUNK_REQUEST_STATE_FAILED:
  case TRUNK_REQUEST_STATE_CANCEL_COMPLETE:
    break;

  default:
    if (!fr_cond_assert(0)) return;
  }

  /*
   *  Zero out the pointer to prevent double frees
   */
  *treq_to_free = NULL;

  /*
   *  Call the API client callback to free
   *  any associated memory.
   */
  DO_REQUEST_FREE(treq);

  /*
   *  Update the last above/below target stats
   *  We only do this when we alloc or free
   *  connections, or on connection
   *      state changes.
   */
  trunk_requests_per_connection(NULL, NULL, treq->pub.trunk, fr_time(), false);

  /*
   *  This tracks the total number of requests
   *  allocated and not freed or returned to
   *  the free list.
   */
  if (fr_cond_assert(trunk->pub.req_alloc > 0)) trunk->pub.req_alloc--;

  /*
   *  No cleanup delay, means cleanup immediately
   */
  if (!fr_time_delta_ispos(trunk->conf.req_cleanup_delay)) {
    treq->pub.state = TRUNK_REQUEST_STATE_INIT;

#ifndef NDEBUG
    /*
     *  Ensure anything parented off the treq
     *  is freed.  We do this to trigger
     *  the destructors for the log entries.
     */
    talloc_free_children(treq);

    /*
     *  State log should now be empty as entries
     *  remove themselves from the dlist
     *  on free.
     */
    fr_assert_msg(fr_dlist_num_elements(&treq->log) == 0,
            "Should have 0 remaining log entries, have %u", fr_dlist_num_elements(&treq->log));
#endif

    talloc_free(treq);
    return;
  }

  /*
   *  Ensure anything parented off the treq
   *  is freed.
   */
  talloc_free_children(treq);

#ifndef NDEBUG
  /*
   *  State log should now be empty as entries
   *  remove themselves from the dlist
   *  on free.
   */
  fr_assert_msg(fr_dlist_num_elements(&treq->log) == 0,
          "Should have 0 remaining log entries, have %u", fr_dlist_num_elements(&treq->log));
#endif

  /*
   *
   *  Return the trunk request back to the init state.
   */
  *treq = (trunk_request_t){
    .pub = {
      .state = TRUNK_REQUEST_STATE_INIT,
      .trunk = treq->pub.trunk,
    },
    .cancel_reason = TRUNK_CANCEL_REASON_NONE,
    .last_freed = fr_time(),
#ifndef NDEBUG
    .log = treq->log        /* Keep the list head, to save reinitialisation */
#endif
  };


  /*
   *  Insert at the head, so that we can free
   *  requests that have been unused for N
   *  seconds from the tail.
   */
  trunk_list_free_requests_add(trunk, treq);

}

/** Actually free the trunk request
 *
 */
static int _trunk_request_free(trunk_request_t *treq)
{
  trunk_t *trunk = treq->pub.trunk;

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_INIT:
  case TRUNK_REQUEST_STATE_UNASSIGNED:
    break;

  default:
    fr_assert(0);
    break;
  }

  trunk_list_free_requests_remove(trunk, treq);

  return 0;
}

/** (Pre-)Allocate a new trunk request
 *
 * If trunk->conf.req_pool_headers or trunk->conf.req_pool_size are not zero then the
 * request will be a talloc pool, which can be used to hold the preq.
 *
 * @note Do not use MEM to check the result of this allocated as it may fail for
 * non-fatal reasons.
 *
 * @param[in] trunk to add request to.
 * @param[in] request to wrap in a trunk request (treq).
 * @return
 *  - A newly allocated request.
 *  - NULL if too many requests are allocated.
 */
trunk_request_t *trunk_request_alloc(trunk_t *trunk, request_t *request)
{
  trunk_request_t *treq;

  /*
   *  The number of treqs currently allocated
   *  exceeds the maximum number allowed.
   */
  if (trunk->conf.max_req_per_conn && trunk->conf.max) {
    uint64_t limit;

    limit = (uint64_t) trunk->conf.max_req_per_conn * trunk->conf.max;
    if (trunk->pub.req_alloc >= (limit + trunk->conf.max_backlog)) {
      RATE_LIMIT_LOCAL_ROPTIONAL(&trunk->limit_max_requests_alloc_log,
               RWARN, WARN, "Refusing to alloc requests - "
               "Limit of %"PRIu64" (max = %u * per_connection_max = %u) "
               "plus %u backlog requests reached",
               limit, trunk->conf.max, trunk->conf.max_req_per_conn,
               trunk->conf.max_backlog);
      return NULL;
    }
  }

  /*
   *  Re-use a recently freed request, which might have some
   *  better cache locality than getting a request from the tail.
   *
   *  If we can't do that, just allocate a new one.
   */
  treq = trunk_list_free_requests_pop(trunk);
  if (treq) {
    fr_assert(treq->pub.state == TRUNK_REQUEST_STATE_INIT);
    fr_assert(treq->pub.trunk == trunk);
    fr_assert(treq->pub.tconn == NULL);
    fr_assert(treq->cancel_reason == TRUNK_CANCEL_REASON_NONE);
    fr_assert(fr_time_gt(treq->last_freed, fr_time_wrap(0)));
    trunk->pub.req_alloc_reused++;
  } else {
    MEM(treq = talloc_pooled_object(trunk, trunk_request_t,
            trunk->conf.req_pool_headers, trunk->conf.req_pool_size));
    talloc_set_destructor(treq, _trunk_request_free);

    *treq = (trunk_request_t){
      .pub = {
        .state = TRUNK_REQUEST_STATE_INIT,
        .trunk = trunk
      },
      .cancel_reason = TRUNK_CANCEL_REASON_NONE
    };
    trunk->pub.req_alloc_new++;
#ifndef NDEBUG
    fr_dlist_init(&treq->log, trunk_request_state_log_t, entry);
#endif
  }

  trunk->pub.req_alloc++;
  treq->id = atomic_fetch_add_explicit(&request_counter, 1, memory_order_relaxed);
  /* heap_id  - initialised when treq inserted into pending */
  /* list   - empty */
  /* preq   - populated later */
  /* rctx   - populated later */
  treq->pub.request = request;

  return treq;
}

/** Enqueue a request that needs data written to the trunk
 *
 * When a request_t * needs to make an asynchronous request to an external datastore
 * it should call this function, specifying a preq (protocol request) containing
 * the data necessary to request information from the external datastore, and an
 * rctx (resume ctx) used to hold the decoded response and/or any error codes.
 *
 * After a treq is successfully enqueued it will either be assigned immediately
 * to the pending queue of a connection, or if no connections are available,
 * (depending on the trunk configuration) the treq will be placed in the trunk's
 * global backlog.
 *
 * After receiving a positive return code from this function the caller should
 * immediately yield, to allow the various timers and I/O handlers that drive tconn
 * (trunk connection) and treq state changes to be called.
 *
 * When a tconn becomes writable (or the trunk is configured to be always writable)
 * the #trunk_request_mux_t callback will be called to dequeue, encode and
 * send any pending requests for that tconn.  The #trunk_request_mux_t callback
 * is also responsible for tracking the outbound requests to allow the
 * #trunk_request_demux_t callback to match inbound responses with the original
 * treq.  Once the #trunk_request_mux_t callback is done processing the treq
 * it signals what state the treq should enter next using one of the
 * trunk_request_signal_* functions.
 *
 * When a tconn becomes readable the user specified #trunk_request_demux_t
 * callback is called to process any responses, match them with the original treq.
 * and signal what state they should enter next using one of the
 * trunk_request_signal_* functions.
 *
 * @param[in,out] treq_out  A trunk request handle.  If the memory pointed to
 *        is NULL, a new treq will be allocated.
 *        Otherwise treq should point to memory allocated
 *        with trunk_request_alloc.
 * @param[in] trunk   to enqueue request on.
 * @param[in] request   to enqueue.
 * @param[in] preq    Protocol request to write out.  Will be freed when
 *        treq is freed. Should ideally be parented by the
 *        treq if possible.
 *        Use #trunk_request_alloc for pre-allocation of
 *        the treq.
 * @param[in] rctx    The resume context to write any result to.
 * @return
 *  - TRUNK_ENQUEUE_OK.
 *  - TRUNK_ENQUEUE_IN_BACKLOG.
 *  - TRUNK_ENQUEUE_NO_CAPACITY.
 *  - TRUNK_ENQUEUE_DST_UNAVAILABLE
 *  - TRUNK_ENQUEUE_FAIL
 */
trunk_enqueue_t trunk_request_enqueue(trunk_request_t **treq_out, trunk_t *trunk,
              request_t *request, void *preq, void *rctx)
{
  trunk_connection_t  *tconn = NULL;
  trunk_request_t   *treq;
  trunk_enqueue_t   ret;

  if (!fr_cond_assert_msg(!IN_HANDLER(trunk),
        "%s cannot be called within a handler", __FUNCTION__)) return TRUNK_ENQUEUE_FAIL;

  if (!fr_cond_assert_msg(!*treq_out || ((*treq_out)->pub.state == TRUNK_REQUEST_STATE_INIT),
        "%s requests must be in \"init\" state", __FUNCTION__)) return TRUNK_ENQUEUE_FAIL;

  /*
   *  If delay_start was set, we may need
   *  to insert the timer for the connection manager.
   */
  if (unlikely(!trunk->started)) {
    if (trunk_start(trunk) < 0) return TRUNK_ENQUEUE_FAIL;
  }

  ret = trunk_request_check_enqueue(&tconn, trunk, request);
  switch (ret) {
  case TRUNK_ENQUEUE_OK:
    if (*treq_out) {
      treq = *treq_out;
    } else {
      *treq_out = treq = trunk_request_alloc(trunk, request);
      if (!treq) return TRUNK_ENQUEUE_NO_CAPACITY;
    }
    treq->pub.preq = preq;
    treq->pub.rctx = rctx;
    if (trunk->conf.always_writable) {
      connection_signals_pause(tconn->pub.conn);
      trunk_request_enter_pending(treq, tconn, true);
      trunk_connection_writable(tconn);
      connection_signals_resume(tconn->pub.conn);
    } else {
      trunk_request_enter_pending(treq, tconn, true);
    }
    break;

  case TRUNK_ENQUEUE_IN_BACKLOG:
    if (*treq_out) {
      treq = *treq_out;
    } else {
      *treq_out = treq = trunk_request_alloc(trunk, request);
      if (!treq) return TRUNK_ENQUEUE_NO_CAPACITY;
    }
    treq->pub.preq = preq;
    treq->pub.rctx = rctx;
    trunk_request_enter_backlog(treq, true);
    break;

  default:
    /*
     *  If a trunk request was provided
     *  populate the preq and rctx fields
     *  so that if it's freed with
     *  trunk_request_free, the free
     *  function works as intended.
     */
    if (*treq_out) {
      treq = *treq_out;
      treq->pub.preq = preq;
      treq->pub.rctx = rctx;
    }
    return ret;
  }

  return ret;
}

/** Re-enqueue a request on the same connection
 *
 * If the treq has been sent, we assume that we're being signalled to requeue
 * because something outside of the trunk API has determined that a retransmission
 * is required.  The easiest way to perform that retransmission is to clean up
 * any tracking information for the request, and the requeue it for transmission.
 *
 * IF re-queueing fails, the request will enter the fail state.  It should not be
 * accessed if this occurs.
 *
 * @param[in] treq  to requeue (retransmit).
 * @return
 *  - TRUNK_ENQUEUE_OK.
 *  - TRUNK_ENQUEUE_DST_UNAVAILABLE - Connection cannot service requests.
 *  - TRUNK_ENQUEUE_FAIL - Request isn't in a valid state to be reassigned.
 */
trunk_enqueue_t trunk_request_requeue(trunk_request_t *treq)
{
  trunk_connection_t  *tconn = treq->pub.tconn; /* Existing conn */

  if (!tconn) return TRUNK_ENQUEUE_FAIL;

  if (!IS_PROCESSING(tconn)) {
    trunk_request_enter_failed(treq);
    return TRUNK_ENQUEUE_DST_UNAVAILABLE;
  }

  switch (treq->pub.state) {
  case TRUNK_REQUEST_STATE_PARTIAL:
  case TRUNK_REQUEST_STATE_SENT:
  case TRUNK_REQUEST_STATE_REAPABLE:
    connection_signals_pause(tconn->pub.conn);
    trunk_request_enter_cancel(treq, TRUNK_CANCEL_REASON_REQUEUE);
    trunk_request_enter_pending(treq, tconn, false);
    if (treq->pub.trunk->conf.always_writable) {
      trunk_connection_writable(tconn);
    }
    connection_signals_resume(tconn->pub.conn);
    break;

  case TRUNK_REQUEST_STATE_BACKLOG: /* Do nothing.... */
  case TRUNK_REQUEST_STATE_PENDING: /* Do nothing.... */
    break;

  default:
    trunk_request_enter_failed(treq);
    return TRUNK_ENQUEUE_FAIL;
  }

  return TRUNK_ENQUEUE_OK;
}

/** Enqueue additional requests on a specific connection
 *
 * This may be used to create a series of requests on a single connection, or to generate
 * in-band status checks.
 *
 * @note If conf->always_writable, then the muxer will be called immediately.  The caller
 *   must be able to handle multiple calls to its muxer gracefully.
 *
 * @param[in,out] treq_out  A trunk request handle.  If the memory pointed to
 *        is NULL, a new treq will be allocated.
 *        Otherwise treq should point to memory allocated
 *        with trunk_request_alloc.
 * @param[in] tconn   to enqueue request on.
 * @param[in] request   to enqueue.
 * @param[in] preq    Protocol request to write out.  Will be freed when
 *        treq is freed. Should ideally be parented by the
 *        treq if possible.
 *        Use #trunk_request_alloc for pre-allocation of
 *        the treq.
 * @param[in] rctx    The resume context to write any result to.
 * @param[in] ignore_limits Ignore max_req_per_conn.  Useful to force status
 *        checks through even if the connection is at capacity.
 *        Will also allow enqueuing on "inactive", "draining",
 *        "draining-to-free" connections.
 * @return
 *  - TRUNK_ENQUEUE_OK.
 *  - TRUNK_ENQUEUE_NO_CAPACITY - At max_req_per_conn_limit
 *  - TRUNK_ENQUEUE_DST_UNAVAILABLE - Connection cannot service requests.
 */
trunk_enqueue_t trunk_request_enqueue_on_conn(trunk_request_t **treq_out, trunk_connection_t *tconn,
                request_t *request, void *preq, void *rctx,
                bool ignore_limits)
{
  trunk_request_t *treq;
  trunk_t   *trunk = tconn->pub.trunk;

  if (!fr_cond_assert_msg(!*treq_out || ((*treq_out)->pub.state == TRUNK_REQUEST_STATE_INIT),
        "%s requests must be in \"init\" state", __FUNCTION__)) return TRUNK_ENQUEUE_FAIL;

  if (!IS_SERVICEABLE(tconn)) return TRUNK_ENQUEUE_DST_UNAVAILABLE;

  /*
   *  Limits check
   */
  if (!ignore_limits) {
    if (trunk->conf.max_req_per_conn &&
        (trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL) >=
         trunk->conf.max_req_per_conn)) return TRUNK_ENQUEUE_NO_CAPACITY;

    if (tconn->pub.state != TRUNK_CONN_ACTIVE) return TRUNK_ENQUEUE_NO_CAPACITY;
  }

  if (*treq_out) {
    treq = *treq_out;
  } else {
    *treq_out = treq = trunk_request_alloc(trunk, request);
    if (!treq) return TRUNK_ENQUEUE_NO_CAPACITY;
  }

  treq->pub.preq = preq;
  treq->pub.rctx = rctx;
  treq->bound_to_conn = true; /* Don't let the request be transferred */

  if (trunk->conf.always_writable) {
    connection_signals_pause(tconn->pub.conn);
    trunk_request_enter_pending(treq, tconn, true);
    trunk_connection_writable(tconn);
    connection_signals_resume(tconn->pub.conn);
  } else {
    trunk_request_enter_pending(treq, tconn, true);
  }

  return TRUNK_ENQUEUE_OK;
}

#ifndef NDEBUG
/** Used for sanity checks to ensure all log entries have been freed
 *
 */
static int _state_log_entry_free(trunk_request_state_log_t *slog)
{
  fr_dlist_remove(slog->log_head, slog);

  return 0;
}

void trunk_request_state_log_entry_add(char const *function, int line,
               trunk_request_t *treq, trunk_request_state_t new)
{
  trunk_request_state_log_t *slog = NULL;

  if (fr_dlist_num_elements(&treq->log) >= TRUNK_REQUEST_STATE_LOG_MAX) {
    slog = fr_dlist_head(&treq->log);
    fr_assert_msg(slog, "slog list head NULL but element counter was %u",
            fr_dlist_num_elements(&treq->log));
    (void)fr_dlist_remove(&treq->log, slog);  /* Returns NULL when removing the list head */
    memset(slog, 0, sizeof(*slog));
  } else {
    MEM(slog = talloc_zero(treq, trunk_request_state_log_t));
    talloc_set_destructor(slog, _state_log_entry_free);
  }

  slog->log_head = &treq->log;
  slog->from = treq->pub.state;
  slog->to = new;
  slog->function = function;
  slog->line = line;
  if (treq->pub.tconn) {
    slog->tconn = treq->pub.tconn;
    slog->tconn_id = treq->pub.tconn->pub.conn->id;
    slog->tconn_state = treq->pub.tconn->pub.state;
  }

  fr_dlist_insert_tail(&treq->log, slog);

}

void trunk_request_state_log(fr_log_t const *log, fr_log_type_t log_type, char const *file, int line,
        trunk_request_t const *treq)
{
  trunk_request_state_log_t *slog = NULL;

  int i;

  for (slog = fr_dlist_head(&treq->log), i = 0;
       slog;
       slog = fr_dlist_next(&treq->log, slog), i++) {
    fr_log(log, log_type, file, line, "[%u] %s:%i - in conn %"PRIu64" in state %s - %s -> %s",
           i, slog->function, slog->line,
           slog->tconn_id,
           slog->tconn ? fr_table_str_by_value(trunk_connection_states,
                        slog->tconn_state, "<INVALID>") : "none",
           fr_table_str_by_value(trunk_request_states, slog->from, "<INVALID>"),
           fr_table_str_by_value(trunk_request_states, slog->to, "<INVALID>"));
  }
}
#endif

/** Return the count number of connections in the specified states
 *
 * @param[in] trunk   to retrieve counts for.
 * @param[in] conn_state  One or more #trunk_connection_state_t states or'd together.
 * @return The number of connections in the specified states.
 */
uint16_t trunk_connection_count_by_state(trunk_t *trunk, int conn_state)
{
  uint16_t count = 0;

  if (conn_state & TRUNK_CONN_INIT) count += fr_dlist_num_elements(&trunk->init);
  if (conn_state & TRUNK_CONN_CONNECTING) count += fr_dlist_num_elements(&trunk->connecting);
  if (conn_state & TRUNK_CONN_ACTIVE) count += fr_minmax_heap_num_elements(trunk->active);
  if (conn_state & TRUNK_CONN_FULL) count += fr_dlist_num_elements(&trunk->full);
  if (conn_state & TRUNK_CONN_INACTIVE) count += fr_dlist_num_elements(&trunk->inactive);
  if (conn_state & TRUNK_CONN_INACTIVE_DRAINING) count += fr_dlist_num_elements(&trunk->inactive_draining);
  if (conn_state & TRUNK_CONN_CLOSED) count += fr_dlist_num_elements(&trunk->closed);
  if (conn_state & TRUNK_CONN_DRAINING) count += fr_dlist_num_elements(&trunk->draining);
  if (conn_state & TRUNK_CONN_DRAINING_TO_FREE) count += fr_dlist_num_elements(&trunk->draining_to_free);

  return count;
}

/** Return the count number of requests associated with a trunk connection
 *
 * @param[in] tconn   to return request count for.
 * @param[in] req_state   One or more request states or'd together.
 *
 * @return The number of requests in the specified states, associated with a tconn.
 */
uint32_t trunk_request_count_by_connection(trunk_connection_t const *tconn, int req_state)
{
  uint32_t count = 0;

  if (req_state & TRUNK_REQUEST_STATE_PENDING) count += fr_heap_num_elements(tconn->pending);
  if (req_state & TRUNK_REQUEST_STATE_PARTIAL) count += tconn->partial ? 1 : 0;
  if (req_state & TRUNK_REQUEST_STATE_SENT) count += fr_dlist_num_elements(&tconn->sent);
  if (req_state & TRUNK_REQUEST_STATE_REAPABLE) count += fr_dlist_num_elements(&tconn->reapable);
  if (req_state & TRUNK_REQUEST_STATE_CANCEL) count += fr_dlist_num_elements(&tconn->cancel);
  if (req_state & TRUNK_REQUEST_STATE_CANCEL_PARTIAL) count += tconn->cancel_partial ? 1 : 0;
  if (req_state & TRUNK_REQUEST_STATE_CANCEL_SENT) count += fr_dlist_num_elements(&tconn->cancel_sent);

  return count;
}

/** Automatically mark a connection as full
 *
 * @param[in] tconn to potentially mark as full.
 */
static inline void trunk_connection_auto_full(trunk_connection_t *tconn)
{
  trunk_t *trunk = tconn->pub.trunk;
  uint32_t  count;

  if (tconn->pub.state != TRUNK_CONN_ACTIVE) return;

  /*
   *  Enforces max_req_per_conn
   */
  if (trunk->conf.max_req_per_conn > 0) {
    count = trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL);
    if (count >= trunk->conf.max_req_per_conn) trunk_connection_enter_full(tconn);
  }
}

/** Return whether a trunk connection should currently be considered full
 *
 * @param[in] tconn to check.
 * @return
 *  - true if the connection is full.
 *  - false if the connection is not full.
 */
static inline bool trunk_connection_is_full(trunk_connection_t *tconn)
{
  trunk_t   *trunk = tconn->pub.trunk;
  uint32_t  count;

  /*
   *  Enforces max_req_per_conn
   */
  count = trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL);
  if ((trunk->conf.max_req_per_conn == 0) || (count < trunk->conf.max_req_per_conn)) return false;

  return true;
}

/** Automatically mark a connection as active or reconnect it
 *
 * @param[in] tconn to potentially mark as active or reconnect.
 */
static inline void trunk_connection_auto_unfull(trunk_connection_t *tconn)
{
  if (tconn->pub.state != TRUNK_CONN_FULL) return;

  /*
   *  Enforces max_req_per_conn
   */
  if (!trunk_connection_is_full(tconn)) trunk_connection_enter_active(tconn);
}

/** A connection is readable.  Call the request_demux function to read pending requests
 *
 */
static inline void trunk_connection_readable(trunk_connection_t *tconn)
{
  trunk_t *trunk = tconn->pub.trunk;

  DO_REQUEST_DEMUX(tconn);
}

/** A connection is writable.  Call the request_mux function to write pending requests
 *
 */
static inline void trunk_connection_writable(trunk_connection_t *tconn)
{
  trunk_t *trunk = tconn->pub.trunk;

  /*
   *  Call the cancel_sent function (if we have one)
   *      to inform a backend datastore we no longer
   *  care about the result
   */
  if (trunk->funcs.request_cancel_mux && trunk_request_count_by_connection(tconn,
                        TRUNK_REQUEST_STATE_CANCEL |
                        TRUNK_REQUEST_STATE_CANCEL_PARTIAL)) {
    DO_REQUEST_CANCEL_MUX(tconn);
  }
  if (!trunk_request_count_by_connection(tconn,
              TRUNK_REQUEST_STATE_PENDING |
              TRUNK_REQUEST_STATE_PARTIAL)) return;
  DO_REQUEST_MUX(tconn);
}

/** Update the registrations for I/O events we're interested in
 *
 */
static void trunk_connection_event_update(trunk_connection_t *tconn)
{
  trunk_t       *trunk = tconn->pub.trunk;
  trunk_connection_event_t  events = TRUNK_CONN_EVENT_NONE;

  switch (tconn->pub.state) {
  /*
   *  We only register I/O events if the trunk connection is
   *  in one of these states.
   *
   *  For the other states the trunk shouldn't be processing
   *  requests.
   */
  case TRUNK_CONN_ACTIVE:
  case TRUNK_CONN_FULL:
  case TRUNK_CONN_INACTIVE:
  case TRUNK_CONN_INACTIVE_DRAINING:
  case TRUNK_CONN_DRAINING:
  case TRUNK_CONN_DRAINING_TO_FREE:
    /*
     *  If the connection is always writable,
     *  then we don't care about write events.
     */
    if (!trunk->conf.always_writable &&
        trunk_request_count_by_connection(tconn,
               TRUNK_REQUEST_STATE_PARTIAL |
                     TRUNK_REQUEST_STATE_PENDING |
               (trunk->funcs.request_cancel_mux ?
               TRUNK_REQUEST_STATE_CANCEL |
               TRUNK_REQUEST_STATE_CANCEL_PARTIAL : 0)) > 0) {
      events |= TRUNK_CONN_EVENT_WRITE;
    }

    if (trunk_request_count_by_connection(tconn,
               TRUNK_REQUEST_STATE_SENT |
               (trunk->funcs.request_cancel_mux ?
               TRUNK_REQUEST_STATE_CANCEL_SENT : 0)) > 0) {
      events |= TRUNK_CONN_EVENT_READ;
    }
    break;

  default:
    break;
  }

  if (tconn->events != events) {
    /*
     *  There may be a fatal error which results
     *  in the connection being freed.
     *
     *  Stop that from happening until after
     *  we're done using it.
     */
    connection_signals_pause(tconn->pub.conn);
    DO_CONNECTION_NOTIFY(tconn, events);
    tconn->events = events;
    connection_signals_resume(tconn->pub.conn);
  }
}

/** Remove a trunk connection from whichever list it's currently in
 *
 * @param[in] tconn to remove.
 */
static void trunk_connection_remove(trunk_connection_t *tconn)
{
  trunk_t *trunk = tconn->pub.trunk;

  switch (tconn->pub.state) {
  case TRUNK_CONN_ACTIVE:
  {
    int ret;

    ret = fr_minmax_heap_extract(trunk->active, tconn);
    if (!fr_cond_assert_msg(ret == 0, "Failed extracting conn from active heap: %s", fr_strerror())) return;
  }
    return;

  case TRUNK_CONN_INIT:
    fr_dlist_remove(&trunk->init, tconn);
    break;

  case TRUNK_CONN_CONNECTING:
    fr_dlist_remove(&trunk->connecting, tconn);
    return;

  case TRUNK_CONN_CLOSED:
    fr_dlist_remove(&trunk->closed, tconn);
    return;

  case TRUNK_CONN_FULL:
    trunk_list_full_remove(trunk, tconn);
    return;

  case TRUNK_CONN_INACTIVE:
    trunk_list_inactive_remove(trunk, tconn);
    return;

  case TRUNK_CONN_INACTIVE_DRAINING:
    trunk_list_inactive_draining_remove(trunk, tconn);
    return;

  case TRUNK_CONN_DRAINING:
    trunk_list_draining_remove(trunk, tconn);
    return;

  case TRUNK_CONN_DRAINING_TO_FREE:
    fr_dlist_remove(&trunk->draining_to_free, tconn);
    return;

  case TRUNK_CONN_HALTED:
    return;
  }
}

/** Transition a connection to the full state
 *
 * Called whenever a trunk connection is at the maximum number of requests.
 * Removes the connection from the connected heap, and places it in the full list.
 */
static void trunk_connection_enter_full(trunk_connection_t *tconn)
{
  trunk_t   *trunk = tconn->pub.trunk;

  switch (tconn->pub.state) {
  case TRUNK_CONN_ACTIVE:
    trunk_connection_remove(tconn);
    break;

  default:
    CONN_BAD_STATE_TRANSITION(TRUNK_CONN_FULL);
  }

  trunk_list_full_add(trunk, tconn);
  CONN_STATE_TRANSITION(TRUNK_CONN_FULL, DEBUG2);
}

/** Transition a connection to the inactive state
 *
 * Called whenever the API client wants to stop new requests being enqueued
 * on a trunk connection.
 */
static void trunk_connection_enter_inactive(trunk_connection_t *tconn)
{
  trunk_t   *trunk = tconn->pub.trunk;

  switch (tconn->pub.state) {
  case TRUNK_CONN_ACTIVE:
  case TRUNK_CONN_FULL:
    trunk_connection_remove(tconn);
    break;

  default:
    CONN_BAD_STATE_TRANSITION(TRUNK_CONN_INACTIVE);
  }

  trunk_list_inactive_add(trunk, tconn);
  CONN_STATE_TRANSITION(TRUNK_CONN_INACTIVE, DEBUG2);
}

/** Transition a connection to the inactive-draining state
 *
 * Called whenever the trunk manager wants to drain an inactive connection
 * of its requests.
 */
static void trunk_connection_enter_inactive_draining(trunk_connection_t *tconn)
{
  trunk_t   *trunk = tconn->pub.trunk;

  switch (tconn->pub.state) {
  case TRUNK_CONN_INACTIVE:
  case TRUNK_CONN_DRAINING:
    trunk_connection_remove(tconn);
    break;

  default:
    CONN_BAD_STATE_TRANSITION(TRUNK_CONN_INACTIVE_DRAINING);
  }

  trunk_list_inactive_draining_add(trunk, tconn);
  CONN_STATE_TRANSITION(TRUNK_CONN_INACTIVE_DRAINING, INFO);

  /*
   *  Immediately re-enqueue all pending
   *  requests, so the connection is drained
   *  quicker.
   */
  trunk_connection_requests_requeue_priv(tconn, TRUNK_REQUEST_STATE_PENDING, 0, false);
}

/** Transition a connection to the draining state
 *
 * Removes the connection from the active heap so it won't be assigned any new
 * connections.
 */
static void trunk_connection_enter_draining(trunk_connection_t *tconn)
{
  trunk_t   *trunk = tconn->pub.trunk;

  switch (tconn->pub.state) {
  case TRUNK_CONN_ACTIVE:
  case TRUNK_CONN_FULL:
  case TRUNK_CONN_INACTIVE:
  case TRUNK_CONN_INACTIVE_DRAINING:
    trunk_connection_remove(tconn);
    break;

  default:
    CONN_BAD_STATE_TRANSITION(TRUNK_CONN_DRAINING);
  }

  trunk_list_draining_add(trunk, tconn);
  CONN_STATE_TRANSITION(TRUNK_CONN_DRAINING, INFO);

  /*
   *  Immediately re-enqueue all pending
   *  requests, so the connection is drained
   *  quicker.
   */
  trunk_connection_requests_requeue_priv(tconn, TRUNK_REQUEST_STATE_PENDING, 0, false);
}

/** Transition a connection to the draining-to-reconnect state
 *
 * Removes the connection from the active heap so it won't be assigned any new
 * connections.
 */
static void trunk_connection_enter_draining_to_free(trunk_connection_t *tconn)
{
  trunk_t   *trunk = tconn->pub.trunk;

  FR_TIMER_DISARM(tconn->lifetime_ev);

  switch (tconn->pub.state) {
  case TRUNK_CONN_ACTIVE:
  case TRUNK_CONN_FULL:
  case TRUNK_CONN_INACTIVE:
  case TRUNK_CONN_INACTIVE_DRAINING:
  case TRUNK_CONN_DRAINING:
    trunk_connection_remove(tconn);
    break;

  default:
    CONN_BAD_STATE_TRANSITION(TRUNK_CONN_DRAINING_TO_FREE);
  }

  fr_dlist_insert_head(&trunk->draining_to_free, tconn);
  CONN_STATE_TRANSITION(TRUNK_CONN_DRAINING_TO_FREE, INFO);

  /*
   *  Immediately re-enqueue all pending
   *  requests, so the connection is drained
   *  quicker.
   */
  trunk_connection_requests_requeue_priv(tconn, TRUNK_REQUEST_STATE_PENDING, 0, false);
}


/** Transition a connection back to the active state
 *
 * This should only be called on a connection which is in the full state,
 * inactive state, draining state or connecting state.
 */
static void trunk_connection_enter_active(trunk_connection_t *tconn)
{
  trunk_t   *trunk = tconn->pub.trunk;
  int   ret;

  switch (tconn->pub.state) {
  case TRUNK_CONN_FULL:
  case TRUNK_CONN_INACTIVE:
  case TRUNK_CONN_INACTIVE_DRAINING:
  case TRUNK_CONN_DRAINING:
    trunk_connection_remove(tconn);
    break;

  case TRUNK_CONN_INIT:
  case TRUNK_CONN_CONNECTING:
    trunk_connection_remove(tconn);
    fr_assert(trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL) == trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_PENDING));
    break;

  default:
    CONN_BAD_STATE_TRANSITION(TRUNK_CONN_ACTIVE);
  }

  ret = fr_minmax_heap_insert(trunk->active, tconn);  /* re-insert into the active heap*/
  if (!fr_cond_assert_msg(ret == 0, "Failed inserting connection into active heap: %s", fr_strerror())) {
    trunk_connection_enter_inactive_draining(tconn);
    return;
  }

  CONN_STATE_TRANSITION(TRUNK_CONN_ACTIVE, DEBUG2);

  /*
   *  Reorder the connections
   */
  CONN_REORDER(tconn);

  /*
   *  Rebalance requests
   */
  trunk_rebalance(trunk);

  /*
   *  We place requests into the backlog
   *      because there were no connections
   *  available to handle them.
   *
   *  If a connection has become active
   *  chances are those backlogged requests
   *      can now be enqueued, so try and do
   *  that now.
   *
   *  If there's requests sitting in the
   *  backlog indefinitely, it's because
   *  they were inserted there erroneously
   *  when there were active connections
   *  which could have handled them.
   */
  trunk_backlog_drain(trunk);
}

/** Connection transitioned to the init state
 *
 * Reflect the connection state change in the lists we use to track connections.
 *
 * @note This function is only called from the connection API as a watcher.
 *
 * @param[in] conn  The connection which changes state.
 * @param[in] prev  The connection is was in.
 * @param[in] state The connection is now in.
 * @param[in] uctx  The trunk_connection_t wrapping the connection.
 */
static void _trunk_connection_on_init(UNUSED connection_t *conn,
              UNUSED connection_state_t prev,
              UNUSED connection_state_t state,
              void *uctx)
{
  trunk_connection_t  *tconn = talloc_get_type_abort(uctx, trunk_connection_t);
  trunk_t   *trunk = tconn->pub.trunk;

  switch (tconn->pub.state) {
  case TRUNK_CONN_HALTED:
    break;

  case TRUNK_CONN_CLOSED:
    trunk_connection_remove(tconn);
    break;

  default:
    CONN_BAD_STATE_TRANSITION(TRUNK_CONN_INIT);
  }

  fr_dlist_insert_head(&trunk->init, tconn);
  CONN_STATE_TRANSITION(TRUNK_CONN_INIT, DEBUG2);
}

/** Connection transitioned to the connecting state
 *
 * Reflect the connection state change in the lists we use to track connections.
 *
 * @note This function is only called from the connection API as a watcher.
 *
 * @param[in] conn  The connection which changes state.
 * @param[in] prev  The connection is was in.
 * @param[in] state The connection is now in.
 * @param[in] uctx  The trunk_connection_t wrapping the connection.
 */
static void _trunk_connection_on_connecting(UNUSED connection_t *conn,
              UNUSED connection_state_t prev,
              UNUSED connection_state_t state,
              void *uctx)
{
  trunk_connection_t  *tconn = talloc_get_type_abort(uctx, trunk_connection_t);
  trunk_t   *trunk = tconn->pub.trunk;

  switch (tconn->pub.state) {
  case TRUNK_CONN_INIT:
  case TRUNK_CONN_CLOSED:
    trunk_connection_remove(tconn);
    break;

  default:
    CONN_BAD_STATE_TRANSITION(TRUNK_CONN_CONNECTING);
  }

  /*
   *  If a connection just entered the
   *  connecting state, it should have
   *  no requests associated with it.
   */
  fr_assert(trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL) == 0);

  fr_dlist_insert_head(&trunk->connecting, tconn);  /* MUST remain a head insertion for reconnect logic */
  CONN_STATE_TRANSITION(TRUNK_CONN_CONNECTING, INFO);
}

/** Connection transitioned to the shutdown state
 *
 * If we're not already in the draining-to-free state, transition there now.
 *
 * The idea is that if something signalled the connection to shutdown, we need
 * to reflect that by dequeuing any pending requests, not accepting new ones,
 * and waiting for the existing requests to complete.
 *
 * @note This function is only called from the connection API as a watcher.
 *
 * @param[in] conn  The connection which changes state.
 * @param[in] prev  The connection is was in.
 * @param[in] state The connection is now in.
 * @param[in] uctx  The trunk_connection_t wrapping the connection.
 */
static void _trunk_connection_on_shutdown(UNUSED connection_t *conn,
            UNUSED connection_state_t prev,
            UNUSED connection_state_t state,
            void *uctx)
{
  trunk_connection_t  *tconn = talloc_get_type_abort(uctx, trunk_connection_t);

  switch (tconn->pub.state) {
  case TRUNK_CONN_DRAINING_TO_FREE:
    /*
     *  Shutdown from draining-to-free means no outstanding requests.
     *  Now signal to halt.
     */
    connection_signal_halt(conn);
    return;

  case TRUNK_CONN_ACTIVE:   /* Transition to draining-to-free */
  case TRUNK_CONN_FULL:
  case TRUNK_CONN_INACTIVE:
  case TRUNK_CONN_INACTIVE_DRAINING:
  case TRUNK_CONN_DRAINING:
    break;

  case TRUNK_CONN_INIT:
  case TRUNK_CONN_CONNECTING:
  case TRUNK_CONN_CLOSED:
  case TRUNK_CONN_HALTED:
    CONN_BAD_STATE_TRANSITION(TRUNK_CONN_DRAINING_TO_FREE);
  }

  trunk_connection_enter_draining_to_free(tconn);
}

/** Trigger a reconnection of the trunk connection
 *
 * @param[in] tl  timer list the timer was inserted into.
 * @param[in] now Current time.
 * @param[in] uctx  The tconn.
 */
static void  _trunk_connection_lifetime_expire(UNUSED fr_timer_list_t *tl, UNUSED fr_time_t now, void *uctx)
{
  trunk_connection_t  *tconn = talloc_get_type_abort(uctx, trunk_connection_t);

  trunk_connection_enter_draining_to_free(tconn);
}

/** Connection transitioned to the connected state
 *
 * Reflect the connection state change in the lists we use to track connections.
 *
 * @note This function is only called from the connection API as a watcher.
 *
 * @param[in] conn  The connection which changes state.
 * @param[in] prev  The connection is was in.
 * @param[in] state The connection is now in.
 * @param[in] uctx  The trunk_connection_t wrapping the connection.
 */
static void _trunk_connection_on_connected(UNUSED connection_t *conn,
             UNUSED connection_state_t prev,
             UNUSED connection_state_t state,
             void *uctx)
{
  trunk_connection_t  *tconn = talloc_get_type_abort(uctx, trunk_connection_t);
  trunk_t     *trunk = tconn->pub.trunk;

  /*
   *  If a connection was just connected, it should only
   *  have a pending list of requests.  This state is found
   *  in the rlm_radius module, which starts a new trunk,
   *  and then immediately enqueues a request onto it.  The
   *  alternative for rlm_radius is to keep it's own queue
   *  of pending requests before the trunk is fully
   *  initialized.  And then enqueue them onto the trunk
   *  when the trunk is connected.
   *
   *  It's instead easier (and makes more sense) to allow
   *  the trunk to accept packets into its queue.  If there
   *  are no connections within a period of time, then the
   *  requests will retry, or will time out.
   */
  fr_assert(trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL) == trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_PENDING));

  /*
   *  Set here, as the active state can
   *  be transitioned to from full and
   *  draining too.
   */
  trunk->pub.last_connected = fr_time();

  /*
   *  Set last_write_success so that idle timeout checks will run
   *  from when the connection has connected if they fire before
   *  any requests are written, rather than from server start time.
   */
  tconn->pub.last_write_success = fr_time();

  /*
   *  Insert a timer to reconnect the
   *  connection periodically.
   */
  if (fr_time_delta_ispos(trunk->conf.lifetime)) {
    if (fr_timer_in(tconn, trunk->el->tl, &tconn->lifetime_ev,
        trunk->conf.lifetime, false, _trunk_connection_lifetime_expire, tconn) < 0) {
      PERROR("Failed inserting connection reconnection timer event, halting connection");
      connection_signal_shutdown(tconn->pub.conn);
      return;
    }
  }

  trunk_connection_enter_active(tconn);
}

/** Connection failed after it was connected
 *
 * Reflect the connection state change in the lists we use to track connections.
 *
 * @note This function is only called from the connection API as a watcher.
 *
 * @param[in] conn  The connection which changes state.
 * @param[in] prev  The connection is was in.
 * @param[in] state The connection is now in.
 * @param[in] uctx  The trunk_connection_t wrapping the connection.
 */
static void _trunk_connection_on_closed(UNUSED connection_t *conn,
            UNUSED connection_state_t prev,
          UNUSED connection_state_t state,
          void *uctx)
{
  trunk_connection_t  *tconn = talloc_get_type_abort(uctx, trunk_connection_t);
  trunk_t     *trunk = tconn->pub.trunk;
  bool      need_requeue = false;

  switch (tconn->pub.state) {
  case TRUNK_CONN_ACTIVE:
  case TRUNK_CONN_FULL:
  case TRUNK_CONN_INACTIVE:
  case TRUNK_CONN_INACTIVE_DRAINING:
  case TRUNK_CONN_DRAINING:
  case TRUNK_CONN_DRAINING_TO_FREE:
    need_requeue = true;
    trunk_connection_remove(tconn);
    break;

  case TRUNK_CONN_INIT:     /* Initialisation failed */
  case TRUNK_CONN_CONNECTING:
    trunk_connection_remove(tconn);
    fr_assert(trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL) == 0);
    break;

  case TRUNK_CONN_CLOSED:
  case TRUNK_CONN_HALTED: /* Can't move backwards? */
    CONN_BAD_STATE_TRANSITION(TRUNK_CONN_CLOSED);
  }

  fr_dlist_insert_head(&trunk->closed, tconn);  /* MUST remain a head insertion for reconnect logic */
  CONN_STATE_TRANSITION(TRUNK_CONN_CLOSED, INFO);

  /*
   *  Now *AFTER* the connection has been
   *  removed from the active, pool
   *  re-enqueue the requests.
   */
  if (need_requeue) trunk_connection_requests_requeue_priv(tconn, TRUNK_REQUEST_STATE_ALL, 0, true);

  /*
   *  There should be no requests left on this
   *  connection.  They should have all been
   *  moved off or failed.
   */
  fr_assert(trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL) == 0);

  /*
   *  Clear statistics and flags
   */
  tconn->sent_count = 0;

  /*
   *  Remove the reconnect event
   */
  if (fr_time_delta_ispos(trunk->conf.lifetime)) FR_TIMER_DELETE(&tconn->lifetime_ev);

  /*
   *  Remove the I/O events
   */
  trunk_connection_event_update(tconn);
}

/** Connection failed
 *
 * @param[in] conn  The connection which changes state.
 * @param[in] prev  The connection is was in.
 * @param[in] state The connection is now in.
 * @param[in] uctx  The trunk_connection_t wrapping the connection.
 */
static void _trunk_connection_on_failed(connection_t *conn,
          connection_state_t prev,
          connection_state_t state,
          void *uctx)
{
  trunk_connection_t  *tconn = talloc_get_type_abort(uctx, trunk_connection_t);
  trunk_t   *trunk = tconn->pub.trunk;

  /*
   *  Need to set this first as it
   *  determines whether requests are
   *  re-queued or fail outright.
   */
  trunk->pub.last_failed = fr_time();

  /*
   *  Failed in the init state, transition the
   *  connection to closed, else we get an
   *  INIT -> INIT transition which triggers
   *  an assert.
   */
  if (prev == CONNECTION_STATE_INIT) _trunk_connection_on_closed(conn, prev, state, uctx);

  /*
   *  See what the state of the trunk is
   *  if there are no connections that could
   *  potentially accept requests in the near
   *  future, then fail all the requests in the
   *  trunk backlog.
   */
  if ((prev == CONNECTION_STATE_CONNECTED) &&
      (trunk_connection_count_by_state(trunk,
            (TRUNK_CONN_ACTIVE |
             TRUNK_CONN_FULL |
             TRUNK_CONN_DRAINING)) == 0)) trunk_backlog_drain(trunk);
}

/** Connection transitioned to the halted state
 *
 * Remove the connection remove all lists, as it's likely about to be freed.
 *
 * Setting the trunk back to the init state ensures that if the code is ever
 * refactored and #connection_signal_reconnect is used after a connection
 * is halted, then everything is maintained in a valid state.
 *
 * @note This function is only called from the connection API as a watcher.
 *
 * @param[in] conn  The connection which changes state.
 * @param[in] prev  The connection is was in.
 * @param[in] state The connection is now in.
 * @param[in] uctx  The trunk_connection_t wrapping the connection.
 */
static void _trunk_connection_on_halted(UNUSED connection_t *conn,
          UNUSED connection_state_t prev,
          UNUSED connection_state_t state,
          void *uctx)
{
  trunk_connection_t  *tconn = talloc_get_type_abort(uctx, trunk_connection_t);
  trunk_t     *trunk = tconn->pub.trunk;

  switch (tconn->pub.state) {
  case TRUNK_CONN_INIT:
  case TRUNK_CONN_CLOSED:
    trunk_connection_remove(tconn);
    break;

  default:
    CONN_BAD_STATE_TRANSITION(TRUNK_CONN_HALTED);
  }

  /*
   *  It began life in the halted state,
   *  and will end life in the halted state.
   */
  CONN_STATE_TRANSITION(TRUNK_CONN_HALTED, DEBUG2);

  /*
   *  There should be no requests left on this
   *  connection.  They should have all been
   *  moved off or failed.
   */
  fr_assert(trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL) == 0);

  /*
   *  And free the connection...
   */
  if (trunk->in_handler) {
    /*
     *  ...later.
     */
    fr_dlist_insert_tail(&trunk->to_free, tconn);
    return;
  }
  talloc_free(tconn);
}

/** Free a connection
 *
 * Enforces orderly free order of children of the tconn
 */
static int _trunk_connection_free(trunk_connection_t *tconn)
{
  fr_assert(tconn->pub.state == TRUNK_CONN_HALTED);
  fr_assert(!fr_dlist_entry_in_list(&tconn->entry)); /* Should not be in a list */

  /*
   *  Loop over all the requests we gathered
   *  and transition them to the failed state,
   *  freeing them.
   *
   *  Usually, requests will be re-queued when
   *  a connection enters the closed state,
   *  but in this case because the whole trunk
   *  is being freed, we don't bother, and
   *  just signal to the API client that the
   *  requests failed.
   */
  if (tconn->pub.trunk->freeing) {
    fr_dlist_head_t to_fail;
    trunk_request_t *treq = NULL;

    fr_dlist_talloc_init(&to_fail, trunk_request_t, entry);

    /*
     *  Remove requests from this connection
     */
    trunk_connection_requests_dequeue(&to_fail, tconn, TRUNK_REQUEST_STATE_ALL, 0);
    while ((treq = fr_dlist_next(&to_fail, treq))) {
      trunk_request_t *prev;

      prev = fr_dlist_remove(&to_fail, treq);
      trunk_request_enter_failed(treq);
      treq = prev;
    }
  }

  /*
   *  Ensure we're not signalled by the connection
   *  as it processes its backlog of state changes,
   *  as we are about to be freed.
   */
  connection_del_watch_pre(tconn->pub.conn, CONNECTION_STATE_INIT, _trunk_connection_on_init);
  connection_del_watch_post(tconn->pub.conn, CONNECTION_STATE_CONNECTING, _trunk_connection_on_connecting);
  connection_del_watch_post(tconn->pub.conn, CONNECTION_STATE_CONNECTED, _trunk_connection_on_connected);
  connection_del_watch_pre(tconn->pub.conn, CONNECTION_STATE_CLOSED, _trunk_connection_on_closed);
  connection_del_watch_post(tconn->pub.conn, CONNECTION_STATE_SHUTDOWN, _trunk_connection_on_shutdown);
  connection_del_watch_pre(tconn->pub.conn, CONNECTION_STATE_FAILED, _trunk_connection_on_failed);
  connection_del_watch_post(tconn->pub.conn, CONNECTION_STATE_HALTED, _trunk_connection_on_halted);

  /*
   *  This may return -1, indicating the free was deferred
   *  this is fine.  It just means the conn will be freed
   *  after all the handlers have exited.
   */
  (void)talloc_free(tconn->pub.conn);
  tconn->pub.conn = NULL;

  return 0;
}

/** Attempt to spawn a new connection
 *
 * Calls the API client's alloc() callback to create a new connection_t,
 * then inserts the connection into the 'connecting' list.
 *
 * @param[in] trunk to spawn connection in.
 * @param[in] now The current time.
 */
static int trunk_connection_spawn(trunk_t *trunk, fr_time_t now)
{
  trunk_connection_t  *tconn;


  /*
   *  Call the API client's callback to create
   *  a new connection_t.
   */
  MEM(tconn = talloc_zero(trunk, trunk_connection_t));
  tconn->pub.trunk = trunk;
  tconn->pub.state = TRUNK_CONN_HALTED; /* All connections start in the halted state */

  /*
   *  Allocate a new connection_t or fail.
   */
  DO_CONNECTION_ALLOC(tconn);

  MEM(tconn->pending = fr_heap_talloc_alloc(tconn, _trunk_request_prioritise, trunk_request_t, heap_id, 0));
  fr_dlist_talloc_init(&tconn->sent, trunk_request_t, entry);
  fr_dlist_talloc_init(&tconn->reapable, trunk_request_t, entry);
  fr_dlist_talloc_init(&tconn->cancel, trunk_request_t, entry);
  fr_dlist_talloc_init(&tconn->cancel_sent, trunk_request_t, entry);

  /*
   *  OK, we have the connection, now setup watch
   *  points so we know when it changes state.
   *
   *  This lets us automatically move the tconn
   *  between the different lists in the trunk
   *  with minimum extra code.
   */
  connection_add_watch_pre(tconn->pub.conn, CONNECTION_STATE_INIT,
            _trunk_connection_on_init, false, tconn);   /* Before init() has been called */

  connection_add_watch_post(tconn->pub.conn, CONNECTION_STATE_CONNECTING,
             _trunk_connection_on_connecting, false, tconn);  /* After init() has been called */

  connection_add_watch_post(tconn->pub.conn, CONNECTION_STATE_CONNECTED,
             _trunk_connection_on_connected, false, tconn); /* After open() has been called */

  connection_add_watch_pre(tconn->pub.conn, CONNECTION_STATE_CLOSED,
            _trunk_connection_on_closed, false, tconn);   /* Before close() has been called */

  connection_add_watch_pre(tconn->pub.conn, CONNECTION_STATE_FAILED,
            _trunk_connection_on_failed, false, tconn);   /* Before failed() has been called */

  connection_add_watch_post(tconn->pub.conn, CONNECTION_STATE_SHUTDOWN,
             _trunk_connection_on_shutdown, false, tconn);  /* After shutdown() has been called */

  connection_add_watch_post(tconn->pub.conn, CONNECTION_STATE_HALTED,
             _trunk_connection_on_halted, false, tconn);  /* About to be freed */

  talloc_set_destructor(tconn, _trunk_connection_free);

  connection_signal_init(tconn->pub.conn);  /* annnnd GO! */

  trunk->pub.last_open = now;

  return 0;
}

/** Pop a cancellation request off a connection's cancellation queue
 *
 * The request we return is advanced by the request moving out of the
 * cancel state and into the cancel_sent or cancel_complete state.
 *
 * One of these signalling functions must be called after the request
 * has been popped:
 *
 * - #trunk_request_signal_cancel_sent
 *   The remote datastore has been informed, but we need to wait for acknowledgement.
 *   The #trunk_request_demux_t callback must handle the acks calling
 *   #trunk_request_signal_cancel_complete when an ack is received.
 *
 * - #trunk_request_signal_cancel_complete
 *   The request was cancelled and we don't need to wait, clean it up immediately.
 *
 * @param[out] treq_out to process
 * @param[in] tconn Connection to drain cancellation request from.
 * @return
 *  - 1 if no more requests.
 *  - 0 if a new request was written to treq_out.
 *  - -1 if the connection was previously freed.  Caller *MUST NOT* touch any
 *    memory or requests associated with the connection.
 *  - -2 if called outside of the cancel muxer.
 */
int trunk_connection_pop_cancellation(trunk_request_t **treq_out, trunk_connection_t *tconn)
{
  if (unlikely(tconn->pub.state == TRUNK_CONN_HALTED)) return -1;

  if (!fr_cond_assert_msg(IN_REQUEST_CANCEL_MUX(tconn->pub.trunk),
        "%s can only be called from within request_cancel_mux handler",
        __FUNCTION__)) return -2;

  *treq_out = tconn->cancel_partial ? tconn->cancel_partial : fr_dlist_head(&tconn->cancel);
  if (!*treq_out) return 1;

  return 0;
}

/** Pop a request off a connection's pending queue
 *
 * The request we return is advanced by the request moving out of the partial or
 * pending states, when the mux function signals us.
 *
 * If the same request is returned again and again, it means the muxer isn't actually
 * doing anything with the request we returned, and it's and error in the muxer code.
 *
 * One of these signalling functions must be used after the request has been popped:
 *
 * - #trunk_request_signal_complete
 *   The request was completed. Either we got a synchronous response, or we knew the
 *   response without contacting an external server (cache).
 *
 * - #trunk_request_signal_fail
 *   Failed muxing the request due to a permanent issue, i.e. an invalid request.
 *
 * - #trunk_request_signal_partial
 *   Wrote part of a request.  This request will be returned on the next call to this
 *   function so that the request_mux function can finish writing it. Only useful
 *   for stream type connections.  Datagram type connections cannot have partial
 *   writes.
 *
 * - #trunk_request_signal_sent Successfully sent a request.
 *
 * @param[out] treq_out to process
 * @param[in] tconn to pop a request from.
 * @return
 *  - 1 if no more requests.
 *  - 0 if a new request was written to treq_out.
 *  - -1 if the connection was previously freed.  Caller *MUST NOT* touch any
 *    memory or requests associated with the connection.
 *  - -2 if called outside of the muxer.
 */
int trunk_connection_pop_request(trunk_request_t **treq_out, trunk_connection_t *tconn)
{
  if (unlikely(tconn->pub.state == TRUNK_CONN_HALTED)) return -1;

  if (!fr_cond_assert_msg(IN_REQUEST_MUX(tconn->pub.trunk),
        "%s can only be called from within request_mux handler",
        __FUNCTION__)) return -2;

  *treq_out = tconn->partial ? tconn->partial : fr_heap_peek(tconn->pending);
  if (!*treq_out) return 1;

  return 0;
}

/** Signal that a trunk connection is writable
 *
 * Should be called from the 'write' I/O handler to signal that requests can be enqueued.
 *
 * @param[in] tconn to signal.
 */
void trunk_connection_signal_writable(trunk_connection_t *tconn)
{
  trunk_t *trunk = tconn->pub.trunk;

  if (!fr_cond_assert_msg(!IN_HANDLER(tconn->pub.trunk),
        "%s cannot be called within a handler", __FUNCTION__)) return;

  DEBUG3("[%" PRIu64 "] Signalled writable", tconn->pub.conn->id);

  trunk_connection_writable(tconn);
}

/** Signal that a trunk connection is readable
 *
 * Should be called from the 'read' I/O handler to signal that requests should be dequeued.
 *
 * @param[in] tconn to signal.
 */
void trunk_connection_signal_readable(trunk_connection_t *tconn)
{
  trunk_t *trunk = tconn->pub.trunk;

  if (!fr_cond_assert_msg(!IN_HANDLER(tconn->pub.trunk),
        "%s cannot be called within a handler", __FUNCTION__)) return;

  DEBUG3("[%" PRIu64 "] Signalled readable", tconn->pub.conn->id);

  trunk_connection_readable(tconn);
}

/** Signal a trunk connection cannot accept more requests
 *
 * @param[in] tconn to signal.
 */
void trunk_connection_signal_inactive(trunk_connection_t *tconn)
{
  /* Can be called anywhere */

  switch (tconn->pub.state) {
  case TRUNK_CONN_ACTIVE:
  case TRUNK_CONN_FULL:
    trunk_connection_enter_inactive(tconn);
    break;

  case TRUNK_CONN_DRAINING:
    trunk_connection_enter_inactive_draining(tconn);
    break;

  default:
    return;
  }
}

/** Signal a trunk connection is no longer full
 *
 * @param[in] tconn to signal.
 */
void trunk_connection_signal_active(trunk_connection_t *tconn)
{
  switch (tconn->pub.state) {
  case TRUNK_CONN_FULL:
    trunk_connection_auto_unfull(tconn);  /* Mark as active if it should be active */
    break;

  case TRUNK_CONN_INACTIVE:
    /*
     *  Do the appropriate state transition based on
     *  how many requests the trunk connection is
     *  currently servicing.
     */
    if (trunk_connection_is_full(tconn)) {
      trunk_connection_enter_full(tconn);
      break;
    }
    trunk_connection_enter_active(tconn);
    break;

  /*
   *  Unsetting the active flag just moves
   *  the connection back to the normal
   *  draining state.
   */
  case TRUNK_CONN_INACTIVE_DRAINING:    /* Only an external signal can trigger this transition */
    trunk_connection_enter_draining(tconn);
    break;

  default:
    return;
  }
}

/** Signal a trunk connection is no longer viable
 *
 * @param[in] tconn to signal.
 * @param[in] reason  the connection is being reconnected.
 */
void trunk_connection_signal_reconnect(trunk_connection_t *tconn, connection_reason_t reason)
{
  connection_signal_reconnect(tconn->pub.conn, reason);
}

/** Standard I/O read function
 *
 * Underlying FD in now readable, so call the trunk to read any pending requests
 * from this connection.
 *
 * @param[in] el  The event list signalling.
 * @param[in] fd  that's now readable.
 * @param[in] flags describing the read event.
 * @param[in] uctx  The trunk connection handle (tconn).
 */
void trunk_connection_callback_readable(UNUSED fr_event_list_t *el, UNUSED int fd, UNUSED int flags, void *uctx)
{
  trunk_connection_t  *tconn = talloc_get_type_abort(uctx, trunk_connection_t);

  trunk_connection_signal_readable(tconn);
}

/** Standard I/O write function
 *
 * Underlying FD is now writable, so call the trunk to write any pending requests
 * to this connection.
 *
 * @param[in] el  The event list signalling.
 * @param[in] fd  that's now writable.
 * @param[in] flags describing the write event.
 * @param[in] uctx  The trunk connection handle (tcon).
 */
void trunk_connection_callback_writable(UNUSED fr_event_list_t *el, UNUSED int fd, UNUSED int flags, void *uctx)
{
  trunk_connection_t  *tconn = talloc_get_type_abort(uctx, trunk_connection_t);

  trunk_connection_signal_writable(tconn);
}


/** Returns true if the trunk connection is in one of the specified states
 *
 * @param[in] tconn To check state for.
 * @param[in] state to check
 * @return
 *  - True if trunk connection is in a particular state.
 *  - False if trunk connection is not in a particular state.
 */
bool trunk_connection_in_state(trunk_connection_t *tconn, int state)
{
  return (bool)(tconn->pub.state & state);
}

/** Close connections in a particular connection list if they have no requests associated with them
 *
 * @param[in] trunk containing connections we want to close.
 * @param[in] head  of list of connections to examine.
 */
static void trunk_connection_close_if_empty(trunk_t *trunk, fr_dlist_head_t *head)
{
  trunk_connection_t *tconn = NULL;

  while ((tconn = fr_dlist_next(head, tconn))) {
    trunk_connection_t *prev;

    if (trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL) != 0) continue;

    prev = fr_dlist_prev(head, tconn);

    DEBUG3("Closing %s connection with no requests",
           fr_table_str_by_value(trunk_connection_states, tconn->pub.state, "<INVALID>"));
    /*
     *  Close the connection as gracefully
     *  as possible by signalling it should
     *  shutdown.
     *
     *  The connection, should, if serviced
     *  correctly by the underlying library,
     *  automatically transition to halted after
     *  all pending reads/writes are
     *  complete at which point we'll be informed
     *  and free our tconn wrapper.
     */
    connection_signal_shutdown(tconn->pub.conn);
    tconn = prev;
  }
}

/** Rebalance connections across active trunk members when a new connection becomes active
 *
 * We don't have any visibility into the connection prioritisation algorithm
 * it's essentially a black box.
 *
 * We can however determine when the correct level of requests per connection
 * has been reached, by dequeuing and requeing  requests up until the point
 * where the connection that just had a request dequeued, receives the same
 * request back.
 *
 * @param[in] trunk The trunk to rebalance.
 */
static void trunk_rebalance(trunk_t *trunk)
{
  trunk_connection_t  *head;

  head = fr_minmax_heap_min_peek(trunk->active);

  /*
   *  Only rebalance if the top and bottom of
   *  the heap are not equal.
   */
  if (trunk->funcs.connection_prioritise(fr_minmax_heap_max_peek(trunk->active), head) == 0) return;

  DEBUG3("Rebalancing requests");

  /*
   *  Keep requeuing requests from the connection
   *  at the bottom of the heap until the
   *  connection at the top is shifted from that
   *  position.
   */
  while ((fr_minmax_heap_min_peek(trunk->active) == head) &&
         trunk_connection_requests_requeue_priv(fr_minmax_heap_max_peek(trunk->active),
                        TRUNK_REQUEST_STATE_PENDING, 1, false));
}

/** Recalculate the trunk's aggregate state from its connection counts
 *
 * Derives the global #trunk_state_t from the number of connections in each
 * connection state, and fires any registered state-change watchers (via
 * #TRUNK_STATE_TRANSITION) if the aggregate state has changed.
 *
 * @param[in] trunk to update.
 */
static void trunk_state_update(trunk_t *trunk)
{
  trunk_state_t new_state;

  /*
   *  Don't churn the state or fire watchers while the trunk is
   *  being torn down.
   */
  if (trunk->freeing) return;

  if (trunk_connection_count_by_state(trunk, TRUNK_CONN_ACTIVE)) {
    /*
     *  One or more connections are active and operational.  The trunk is ACTIVE.
     */
    new_state = TRUNK_STATE_ACTIVE;

  } else if (trunk_connection_count_by_state(trunk, TRUNK_CONN_INIT | TRUNK_CONN_CONNECTING)) {
    /*
     *  Connections are being opened, but none are usable yet.
     *
     *  This is checked before FULL.  If a connection is CONNECTING, then the trunk is by
     *  definition not full.
     */
    new_state = TRUNK_STATE_PENDING;

  } else if (trunk->conf.max &&
       (trunk_connection_count_by_state(trunk, TRUNK_CONN_FULL) == trunk->conf.max)) {
    /*
     *  No active or connecting connections, and every one of the maximum permitted
     *  connections is connected and full.  The backend is reachable, but the trunk has no
     *  spare capacity, and can accept no more traffic.
     */
    new_state = TRUNK_STATE_FULL;

  } else if (trunk_connection_count_by_state(trunk, TRUNK_CONN_CLOSED)) {
    /*
     *  Connections exist, but they have all failed and are
     *  closed / in reconnect backoff.  The backend is
     *  currently unreachable.
     */
    new_state = TRUNK_STATE_FAILED;

  } else {
    new_state = TRUNK_STATE_IDLE;
  }

  if (new_state == trunk->pub.state) return;

  /*
   *  This can be reached from within a state-change watcher.  A watcher may enqueue a request or
   *  reconnect a connection, which changes a connection's state and calls
   *  trunk_requests_per_connection() -> trunk_state_update().
   *
   *  Nested watcher calls are not allowed (see trunk_watch_call()).  If we're already inside one,
   *  leave the state unchanged and let the next non-nested update, or the periodic trunk_manage(),
   *  reconcile it.
   */
  if (trunk->next_watcher != NULL) return;

  TRUNK_STATE_TRANSITION(new_state);
}

/** Implements the algorithm we use to manage requests per connection levels
 *
 * This is executed periodically using a timer event, and opens/closes
 * connections.
 *
 * The aim is to try and keep the request per connection level in a sweet spot,
 * where there's enough outstanding work for the connection/pipelining to work
 * efficiently, but not so much so that we encounter increased latency.
 *
 * In the request enqueue and dequeue functions we record every time the
 * average number of requests per connection goes above the target count
 * and record every time the average number of requests per connection goes
 * below the target count.
 *
 * This may sound expensive, but in all cases we're just summing counters.
 * CPU time required does not increase with additional requests, only with
 * large numbers of connections.
 *
 * If we do encounter scaling issues, we can always maintain the counters
 * as aggregates as an optimisation later.
 *
 * If when the management function runs, the trunk was above the target
 * most recently, we:
 * - Return if we've been in this state for a shorter period than 'open_delay'.
 * - Return if we're at max.
 * - Return if opening a new connection will take us below the load target.
 * - Return if we last opened a connection within 'open_delay'.
 * - Otherwise we attempt to open a new connection.
 *
 * If the trunk we below the target most recently, we:
 * - Return if we've been in this state for a shorter period than 'close_delay'.
 * - Return if we're at min.
 * - Return if we have no connections.
 * - Close a connection if min is 0, and we have no outstanding
 *   requests.  Then return.
 * - Return if closing a new connection will take us above the load target.
 * - Return if we last closed a connection within 'closed_delay'.
 * - Otherwise we move a connection to draining state.
 */
static void trunk_manage(trunk_t *trunk, fr_time_t now)
{
  trunk_connection_t  *tconn = NULL;
  trunk_request_t   *treq;
  uint32_t    average = 0;
  uint32_t    req_count;
  uint16_t    conn_count;

  DEBUG4("Managing trunk");

  /*
   *  Cleanup requests in our request cache which
   *  have been reapable for too long.
   */
  while ((treq = trunk_list_free_requests_peek(trunk)) &&
      fr_time_lteq(fr_time_add(treq->last_freed, trunk->conf.req_cleanup_delay), now)) talloc_free(treq);

  /*
   *  If we have idle connections, then close them, maintaining "min" connections.
   */
  if (fr_time_delta_ispos(trunk->conf.idle_timeout) &&
      (fr_minmax_heap_num_elements(trunk->active) > trunk->conf.min)) {
    fr_minmax_heap_iter_t iter;
    fr_time_t idle_cutoff = fr_time_sub(now, trunk->conf.idle_timeout);

    for (tconn = fr_minmax_heap_iter_init(trunk->active, &iter);
         tconn;
         tconn = fr_minmax_heap_iter_next(trunk->active, &iter)) {
      /*
       *  The connection has outstanding requests without replies, don't do anything.
       */
      if (fr_heap_num_elements(tconn->pending) > 0) continue;

      /*
       *  The connection was last active after the idle cutoff time, don't do anything.
       */
      if (fr_time_gt(tconn->pub.last_write_success, idle_cutoff)) continue;

      /*
       *  This connection has been inactive since before the idle timeout.  Drain it,
       *  and free it.
       *
       *  This also extracts the connection from the minmax heap, which invalidates the
       *  iterator, so we stop iterating over it.
       */
      trunk_connection_enter_draining_to_free(tconn);
      break;
    }
  }

  /*
   *  Free any connections which have drained
   *  and we didn't reactivate during the last
   *  round of management.
   */
  trunk_connection_close_if_empty(trunk, &trunk->inactive_draining);
  trunk_connection_close_if_empty(trunk, &trunk->draining);
  trunk_connection_close_if_empty(trunk, &trunk->draining_to_free);

  /*
   *  Process deferred connection freeing
   */
  if (!trunk->in_handler) fr_dlist_talloc_free(&trunk->to_free);

  /*
   *  Update the state of the trunk
   */
  trunk_state_update(trunk);

  /*
   *  A trunk can be signalled to not proactively
   *  manage connections if a destination is known
   *  to be unreachable, and doing so would result
   *  in spurious connections still being opened.
   *
   *  We still run other connection management
   *  functions and just short circuit the function
   *  here.
   */
  if (!trunk->managing_connections) return;

  /*
   *  We're above the target requests per connection
   *  spawn more connections!
   */
  if (fr_time_gteq(trunk->pub.last_above_target, trunk->pub.last_below_target)) {
    /*
     *  If connecting is provided, check we
     *  wouldn't have too many connections in
     *  the connecting state.
     *
     *  This is a throttle in the case of transitory
     *  load spikes, or a backend becoming
     *  unavailable.
     */
    if ((trunk->conf.connecting > 0) &&
        (trunk_connection_count_by_state(trunk, TRUNK_CONN_CONNECTING) >=
         trunk->conf.connecting)) {
      DEBUG4("Not opening connection - Too many (%u) connections in the connecting state",
             trunk->conf.connecting);
      return;
    }

    trunk_requests_per_connection(&conn_count, &req_count, trunk, now, true);

    /*
     *  Only apply hysteresis if we have at least
     *  one available connection.
     */
    if (conn_count && fr_time_gt(fr_time_add(trunk->pub.last_above_target, trunk->conf.open_delay), now)) {
      DEBUG4("Not opening connection - Need to be above target for %pVs.  It's been %pVs",
             fr_box_time_delta(trunk->conf.open_delay),
             fr_box_time_delta(fr_time_sub(now, trunk->pub.last_above_target)));
      return; /* too soon */
    }

    /*
     *  We don't consider 'draining' connections
     *  in the max calculation, as if we do
     *  determine that we need to spawn a new
     *  request, then we'd move all 'draining'
     *  connections to active before spawning
     *  any new connections.
     */
    if ((trunk->conf.max > 0) && (conn_count >= trunk->conf.max)) {
      DEBUG4("Not opening connection - Have %u connections, need %u or below",
             conn_count, trunk->conf.max);
      return;
    }

    /*
     *  We consider requests pending on all connections
     *      and the trunk's backlog as that's the current count
     *  load.
     */
    if (!req_count) {
      DEBUG4("Not opening connection - No outstanding requests");
      return;
    }

    /*
     *  Do the n+1 check, i.e. if we open one connection
     *  will that take us below our target threshold.
     */
    if (conn_count > 0) {
      average = ROUND_UP_DIV(req_count, (conn_count + 1));
      if (average < trunk->conf.target_req_per_conn) {
        DEBUG4("Not opening connection - Would leave us below our target requests "
               "per connection (now %u, after open %u)",
               ROUND_UP_DIV(req_count, conn_count), average);
        return;
      }
    } else {
      (void)trunk_connection_spawn(trunk, now);
      return;
    }

    /*
     *  If we've got a connection in the draining list
     *      move it back into the active list if we've
     *      been requested to add a connection back in.
     */
    tconn = fr_dlist_head(&trunk->draining);
    if (tconn) {
      if (trunk_connection_is_full(tconn)) {
        trunk_connection_enter_full(tconn);
      } else {
        trunk_connection_enter_active(tconn);
      }
      return;
    }

    /*
     *  Implement delay if there's no connections that
     *  could be immediately re-activated.
     */
    if (fr_time_gt(fr_time_add(trunk->pub.last_open, trunk->conf.open_delay), now)) {
      DEBUG4("Not opening connection - Need to wait %pVs before opening another connection.  "
             "It's been %pVs",
             fr_box_time_delta(trunk->conf.open_delay),
             fr_box_time_delta(fr_time_sub(now, trunk->pub.last_open)));
      return;
    }

    DEBUG4("Opening connection - Above target requests per connection (now %u, target %u)",
           ROUND_UP_DIV(req_count, conn_count), trunk->conf.target_req_per_conn);
    /* last_open set by trunk_connection_spawn */
    (void)trunk_connection_spawn(trunk, now);
  }

  /*
   *  We're below the target requests per connection.
   *  Free some connections...
   */
  else if (fr_time_gt(trunk->pub.last_below_target, trunk->pub.last_above_target)) {
    if (fr_time_gt(fr_time_add(trunk->pub.last_below_target, trunk->conf.close_delay), now)) {
      DEBUG4("Not closing connection - Need to be below target for %pVs. It's been %pVs",
             fr_box_time_delta(trunk->conf.close_delay),
             fr_box_time_delta(fr_time_sub(now, trunk->pub.last_below_target)));
      return; /* too soon */
    }

    trunk_requests_per_connection(&conn_count, &req_count, trunk, now, true);

    if (!conn_count) {
      DEBUG4("Not closing connection - No connections to close!");
      return;
    }

    if ((trunk->conf.min > 0) && ((conn_count - 1) < trunk->conf.min)) {
      DEBUG4("Not closing connection - Have %u connections, need %u or above",
             conn_count, trunk->conf.min);
      return;
    }

    if (!req_count) {
      DEBUG4("Closing connection - No outstanding requests");
      goto close;
    }

    /*
     *  The minimum number of connections must be set
     *  to zero for this to work.
     *  min == 0, no requests, close all the connections.
     *      This is useful for backup databases, when
     *  maintaining the connection would lead to lots of
     *  log file churn.
     */
    if (conn_count == 1) {
      DEBUG4("Not closing connection - Would leave connections "
             "and there are still %u outstanding requests", req_count);
      return;
    }

    /*
     *  Do the n-1 check, i.e. if we close one connection
     *  will that take us above our target threshold.
     */
    average = ROUND_UP_DIV(req_count, (conn_count - 1));
    if (average > trunk->conf.target_req_per_conn) {
      DEBUG4("Not closing connection - Would leave us above our target requests per connection "
             "(now %u, after close %u)", ROUND_UP_DIV(req_count, conn_count), average);
      return;
    }

    DEBUG4("Closing connection - Below target requests per connection (now %u, target %u)",
           ROUND_UP_DIV(req_count, conn_count), trunk->conf.target_req_per_conn);

  close:
    if (fr_time_gt(fr_time_add(trunk->pub.last_closed, trunk->conf.close_delay), now)) {
      DEBUG4("Not closing connection - Need to wait %pVs before closing another connection.  "
             "It's been %pVs",
             fr_box_time_delta(trunk->conf.close_delay),
             fr_box_time_delta(fr_time_sub(now, trunk->pub.last_closed)));
      return;
    }

    /*
     *  If the last event on the trunk was a connection failure and
     *  there is only one connection, this may well be a reconnect
     *  attempt after a failure - and needs to persist otherwise
     *  the last event will be a failure and no new connection will
     *  be made, leading to no new requests being enqueued.
     */
    if (fr_time_gt(trunk->pub.last_failed, fr_time_wrap(0)) &&
        fr_time_lt(trunk->pub.last_connected, trunk->pub.last_failed) && (conn_count == 1)) {
      DEBUG4("Not closing remaining connection - last event was a failure");
      return;
    }

    /*
     *  Inactive connections get counted in the
     *  set of viable connections, but are likely
     *  to be congested or dead, so we drain
     *  (and possibly eventually free) those first.
     */
    if ((tconn = trunk_list_inactive_peek(trunk))) {
      /*
       *  If the connection has no requests associated
       *  with it then immediately free.
       */
      if (trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL) == 0) {
        connection_signal_halt(tconn->pub.conn);  /* Also frees the tconn */
      } else {
        trunk_connection_enter_inactive_draining(tconn);
      }
    /*
     *  It is possible to have too may connecting
     *  connections when the connections are
     *  taking a while to open and the number
     *  of requests decreases.
     */
    } else if ((tconn = fr_dlist_tail(&trunk->connecting))) {
      connection_signal_halt(tconn->pub.conn);  /* Also frees the tconn */

    /*
     *  Finally if there are no "connecting"
     *  connections to close, and no "inactive"
     *  connections, start draining "active"
     *  connections.
     */
    } else if ((tconn = fr_minmax_heap_max_peek(trunk->active))) {
      /*
       *  If the connection has no requests associated
       *  with it then immediately free.
       */
      if (trunk_request_count_by_connection(tconn, TRUNK_REQUEST_STATE_ALL) == 0) {
        connection_signal_halt(tconn->pub.conn);  /* Also frees the tconn */
      } else {
        trunk_connection_enter_draining(tconn);
      }
    }

    trunk->pub.last_closed = now;

    return;
  }
}

/** Event to periodically call the connection management function
 *
 * @param[in] tl  this event belongs to.
 * @param[in] now current time.
 * @param[in] uctx  The trunk.
 */
static void _trunk_timer(fr_timer_list_t *tl, fr_time_t now, void *uctx)
{
  trunk_t *trunk = talloc_get_type_abort(uctx, trunk_t);

  trunk_manage(trunk, now);

  if (fr_time_delta_ispos(trunk->conf.manage_interval)) {
    if (fr_timer_in(trunk, tl, &trunk->manage_ev, trunk->conf.manage_interval,
      false, _trunk_timer, trunk) < 0) {
      PERROR("Failed inserting trunk management event");
      /* Not much we can do, hopefully the trunk will be freed soon */
    }
  }
}

/** Return a count of requests on a connection in a specific state
 *
 * @param[in] trunk   to retrieve counts for.
 * @param[in] conn_state  One or more connection states or'd together.
 * @param[in] req_state   One or more request states or'd together.
 * @return The number of requests in a particular state, on connection in a particular state.
 */
uint64_t trunk_request_count_by_state(trunk_t *trunk, int conn_state, int req_state)
{
  uint64_t    count = 0;
  trunk_connection_t  *tconn = NULL;
  fr_minmax_heap_iter_t iter;

#define COUNT_BY_STATE(_state, _list) \
do { \
  if (conn_state & (_state)) { \
    tconn = NULL; \
    while ((tconn = fr_dlist_next(&trunk->_list, tconn))) { \
      count += trunk_request_count_by_connection(tconn, req_state); \
    } \
  } \
} while (0)

  if (conn_state & TRUNK_CONN_ACTIVE) {
    for (tconn = fr_minmax_heap_iter_init(trunk->active, &iter);
         tconn;
         tconn = fr_minmax_heap_iter_next(trunk->active, &iter)) {
      count += trunk_request_count_by_connection(tconn, req_state);
    }
  }

  COUNT_BY_STATE(TRUNK_CONN_FULL, full);
  COUNT_BY_STATE(TRUNK_CONN_INACTIVE, inactive);
  COUNT_BY_STATE(TRUNK_CONN_INACTIVE_DRAINING, inactive_draining);
  COUNT_BY_STATE(TRUNK_CONN_DRAINING, draining);
  COUNT_BY_STATE(TRUNK_CONN_DRAINING_TO_FREE, draining_to_free);

  if (req_state & TRUNK_REQUEST_STATE_BACKLOG) count += fr_heap_num_elements(trunk->backlog);

  return count;
}

/** Update timestamps for when we last had a transition from above target to below target or vice versa
 *
 * Should be called on every time a connection or request is allocated or freed.
 *
 * @param[out] conn_count_out How many connections we considered.
 * @param[out] req_count_out  How many requests we considered.
 * @param[in] trunk   to operate on.
 * @param[in] now   The current time.
 * @param[in] verify    if true (and this is a debug build), then assert if req_per_conn
 *        has changed.
 * @return
 *  - 0 if the average couldn't be calculated (no requests or no connections).
 *  - The average number of requests per connection.
 */
static uint64_t trunk_requests_per_connection(uint16_t *conn_count_out, uint32_t *req_count_out,
                trunk_t *trunk, fr_time_t now,
                NDEBUG_UNUSED bool verify)
{
  uint32_t req_count = 0;
  uint16_t conn_count = 0;
  uint64_t req_per_conn = 0;

  fr_assert(fr_time_gt(now, fr_time_wrap(0)));

  /*
   *  Recompute the trunk's aggregate state (and fire any state
   *  watchers) now that a connection's state may have changed.
   *  This is the authoritative, prompt trigger for the trunk
   *  entering / leaving states such as ACTIVE, FULL and FAILED.
   *  trunk_state_update() no-ops if the trunk is being freed.
   */
  trunk_state_update(trunk);

  /*
   *  No need to update these as the trunk is being freed
   */
  if (trunk->freeing) goto done;

  /*
   *  Count all connections except draining and draining to free.
   *
   *  Omitting these connection states artificially raises the
   *  request to connection ratio, so that we can preemptively spawn
   *  new connections.
   *
   *  In the case of TRUNK_CONN_DRAINING | TRUNK_CONN_INACTIVE_DRAINING
   *  the trunk management code has enough hysteresis to not
   *  immediately reactivate the connection.
   *
   *  In the case of TRUNK_CONN_DRAINING_TO_FREE the trunk
   *  management code should spawn a new connection to takes its place.
   *
   *  Connections placed in the DRAINING_TO_FREE state are being
   *  closed preemptively to deal with bugs on the server we're
   *  talking to, or misconfigured firewalls which are trashing
   *  TCP/UDP connection states.
   */
  conn_count = trunk_connection_count_by_state(trunk, TRUNK_CONN_ALL ^
              (TRUNK_CONN_DRAINING |
               TRUNK_CONN_INACTIVE_DRAINING |
               TRUNK_CONN_DRAINING_TO_FREE));

  /*
   *  Requests on all connections
   */
  req_count = trunk_request_count_by_state(trunk,
                TRUNK_CONN_ALL ^
                TRUNK_CONN_DRAINING_TO_FREE, TRUNK_REQUEST_STATE_ALL);

  /*
   *  No connections, but we do have requests
   */
  if (conn_count == 0) {
    if ((req_count > 0) && (trunk->conf.target_req_per_conn > 0)) goto above_target;
    goto done;
  }

  if (req_count == 0) {
    if (trunk->conf.target_req_per_conn > 0) goto below_target;
    goto done;
  }

  /*
   *  Calculate the req_per_conn
   */
  req_per_conn = ROUND_UP_DIV(req_count, conn_count);
  if (req_per_conn > trunk->conf.target_req_per_conn) {
  above_target:
    /*
     *  Edge - Below target to above target (too many requests per conn - spawn more)
     *
     *  The equality check is correct here as both values start at 0.
     */
    if (fr_time_lteq(trunk->pub.last_above_target, trunk->pub.last_below_target)) trunk->pub.last_above_target = now;
  } else if (req_per_conn < trunk->conf.target_req_per_conn) {
  below_target:
    /*
     *  Edge - Above target to below target (too few requests per conn - close some)
     *
     *  The equality check is correct here as both values start at 0.
     */
    if (fr_time_lteq(trunk->pub.last_below_target, trunk->pub.last_above_target)) trunk->pub.last_below_target = now;
  }

done:
  if (conn_count_out) *conn_count_out = conn_count;
  if (req_count_out) *req_count_out = req_count;

  /*
   *  Check we haven't missed a call to trunk_requests_per_connection
   */
  fr_assert(!verify || (trunk->last_req_per_conn == 0) || (req_per_conn == trunk->last_req_per_conn));

  trunk->last_req_per_conn = req_per_conn;

  return req_per_conn;
}

/** Drain the backlog of as many requests as possible
 *
 * @param[in] trunk To drain backlog requests for.
 */
static void trunk_backlog_drain(trunk_t *trunk)
{
  trunk_request_t *treq;

  if (fr_heap_num_elements(trunk->backlog) == 0) return;

  /*
   *  If it's always writable, this isn't
   *  really a noteworthy event.
   */
  if (!trunk->conf.always_writable) DEBUG3("Draining backlog of requests");

  /*
   *  Do *NOT* add an artificial limit
   *  here.  We rely on all available
   *  connections entering the full
   *  state and transitioning back to
   *  active in order to drain the
   *  backlog.
   */
  while ((treq = fr_heap_peek(trunk->backlog))) {
    switch (trunk_request_enqueue_existing(treq)) {
    case TRUNK_ENQUEUE_OK:
      continue;

    /*
     *  Signal to stop
     */
    case TRUNK_ENQUEUE_IN_BACKLOG:
      break;

    /*
     *  Failed enqueueing the request,
     *  have it enter the failed state
     *  which will free it and
     *  re-enliven the yielded request.
     */
    case TRUNK_ENQUEUE_DST_UNAVAILABLE:
    case TRUNK_ENQUEUE_FAIL:
      trunk_request_enter_failed(treq);
      continue;

    case TRUNK_ENQUEUE_NO_CAPACITY:
      fr_assert(fr_minmax_heap_num_elements(trunk->active) == 0);
      return;
    }
  }
}

/** Force the trunk to re-establish its connections
 *
 * @param[in] trunk   to signal.
 * @param[in] states    One or more states or'd together.
 * @param[in] reason    Why the connections are being signalled to reconnect.
 */
void trunk_reconnect(trunk_t *trunk, int states, connection_reason_t reason)
{

#define RECONNECT_BY_STATE(_state, _list) \
do { \
  if (states & (_state)) { \
    size_t i; \
    for (i = fr_dlist_num_elements(&trunk->_list); i > 0; i--) { \
      connection_signal_reconnect(((trunk_connection_t *)fr_dlist_tail(&trunk->_list))->pub.conn, reason); \
    } \
  } \
} while (0)

  /*
   *  Connections in the 'connecting' state
   *  may re-enter that state, so we need to
   *  be careful not to enter an infinite
   *  loop, as we iterate over the list
   *  again and again.
   */
  RECONNECT_BY_STATE(TRUNK_CONN_CONNECTING, connecting);

  if (states & TRUNK_CONN_ACTIVE) {
    trunk_connection_t *tconn;
    while ((tconn = fr_minmax_heap_min_peek(trunk->active))) connection_signal_reconnect(tconn->pub.conn, reason);
  }

  RECONNECT_BY_STATE(TRUNK_CONN_INIT, init);
  RECONNECT_BY_STATE(TRUNK_CONN_FULL, full);
  RECONNECT_BY_STATE(TRUNK_CONN_INACTIVE, inactive);
  RECONNECT_BY_STATE(TRUNK_CONN_INACTIVE_DRAINING, inactive_draining);
  RECONNECT_BY_STATE(TRUNK_CONN_CLOSED, closed);
  RECONNECT_BY_STATE(TRUNK_CONN_DRAINING, draining);
  RECONNECT_BY_STATE(TRUNK_CONN_DRAINING_TO_FREE, draining_to_free);
}

/** Start the trunk running
 *
 */
int trunk_start(trunk_t *trunk)
{
  uint16_t i;

  if (unlikely(trunk->started)) return 0;

  /*
   *  Spawn the initial set of connections
   */
  for (i = 0; i < trunk->conf.start; i++) {
    DEBUG("[%i] Starting initial connection", i);
    if (trunk_connection_spawn(trunk, fr_time()) != 0) return -1;
  }

  /*
   *  If the idle timeout is set, AND there's no management interval, OR the management interval is
   *  less than the idle timeout, update the management interval.
   */
  if (fr_time_delta_ispos(trunk->conf.idle_timeout) &&
      (!fr_time_delta_ispos(trunk->conf.manage_interval) ||
      fr_time_delta_gt(trunk->conf.manage_interval, trunk->conf.idle_timeout))) {
    trunk->conf.manage_interval = trunk->conf.idle_timeout;
  }

  if (fr_time_delta_ispos(trunk->conf.manage_interval)) {
    /*
     *  Insert the event timer to manage
     *  the interval between managing connections.
     */
    if (fr_timer_in(trunk, trunk->el->tl, &trunk->manage_ev, trunk->conf.manage_interval,
        false, _trunk_timer, trunk) < 0) {
      PERROR("Failed inserting trunk management event");
      return -1;
    }
  }
  trunk->started = true;
  trunk->managing_connections = true;

  return 0;
}

/** Allow the trunk to open and close connections in response to load
 *
 */
void trunk_connection_manage_start(trunk_t *trunk)
{
  if (!trunk->started || trunk->managing_connections) return;

  DEBUG3("Connection management enabled");
  trunk->managing_connections = true;
}

/** Stop the trunk from opening and closing connections in response to load
 *
 */
void trunk_connection_manage_stop(trunk_t *trunk)
{
  if (!trunk->started || !trunk->managing_connections) return;

  DEBUG3("Connection management disabled");
  trunk->managing_connections = false;
}

/** Schedule a trunk management event for the next time the event loop is executed
 */
int trunk_connection_manage_schedule(trunk_t *trunk)
{
  if (!trunk->started || !trunk->managing_connections) return 0;

  if (fr_timer_in(trunk, trunk->el->tl, &trunk->manage_ev, fr_time_delta_wrap(0),
      false, _trunk_timer, trunk) < 0) {
    PERROR("Failed inserting trunk management event");
    return -1;
  }

  return 0;
}

/** Order connections by queue depth
 *
 */
static int8_t _trunk_connection_order_by_shortest_queue(void const *one, void const *two)
{
  trunk_connection_t const  *a = talloc_get_type_abort_const(one, trunk_connection_t);
  trunk_connection_t const  *b = talloc_get_type_abort_const(two, trunk_connection_t);

  uint32_t      a_count = trunk_request_count_by_connection(a, TRUNK_REQUEST_STATE_ALL);
  uint32_t      b_count = trunk_request_count_by_connection(b, TRUNK_REQUEST_STATE_ALL);

  /*
   *  Add a fudge factor of 1 to reduce spurious rebalancing
   */
  return ((a_count > b_count) && ((a_count - b_count) > 1)) - ((b_count > a_count) && ((b_count - a_count) > 1));
}

/** Free a trunk, gracefully closing all connections.
 *
 */
static int _trunk_free(trunk_t *trunk)
{
  trunk_connection_t  *tconn;
  trunk_request_t *treq;
  trunk_watch_entry_t *watch;
  size_t      i;

  DEBUG4("Trunk free %p", trunk);

  trunk->freeing = true;  /* Prevent re-enqueuing */

  /*
   *  We really don't want this firing after
   *  we've freed everything.
   */
  FR_TIMER_DELETE_RETURN(&trunk->manage_ev);

  /*
   *  Now free the connections in each of the lists.
   *
   *  Each time a connection is freed it removes itself from the list
   *  its in, which means the head should keep advancing automatically.
   */
  while ((tconn = fr_minmax_heap_min_peek(trunk->active))) connection_signal_halt(tconn->pub.conn);
  while ((tconn = fr_dlist_head(&trunk->init))) connection_signal_halt(tconn->pub.conn);
  while ((tconn = fr_dlist_head(&trunk->connecting))) connection_signal_halt(tconn->pub.conn);
  while ((tconn = fr_dlist_head(&trunk->full))) connection_signal_halt(tconn->pub.conn);
  while ((tconn = fr_dlist_head(&trunk->inactive))) connection_signal_halt(tconn->pub.conn);
  while ((tconn = fr_dlist_head(&trunk->inactive_draining))) connection_signal_halt(tconn->pub.conn);
  while ((tconn = fr_dlist_head(&trunk->closed))) connection_signal_halt(tconn->pub.conn);
  while ((tconn = fr_dlist_head(&trunk->draining))) connection_signal_halt(tconn->pub.conn);
  while ((tconn = fr_dlist_head(&trunk->draining_to_free))) connection_signal_halt(tconn->pub.conn);

  /*
   *  Process any deferred connection frees
   */
  fr_dlist_talloc_free(&trunk->to_free);

  /*
   *  Free any requests left in the backlog
   */
  while ((treq = fr_heap_peek(trunk->backlog))) trunk_request_enter_failed(treq);

  /*
   *  Free any requests in our request cache
   */
  while ((treq = trunk_list_free_requests_peek(trunk))) talloc_free(treq);

  /*
   *  Free any entries in the watch lists
   */
  for (i = 0; i < NUM_ELEMENTS(trunk->watch); i++) {
    while ((watch = fr_dlist_pop_head(&trunk->watch[i]))) talloc_free(watch);
  }

  return 0;
}

/** Allocate a new collection of connections
 *
 * This function should be called first to allocate a new trunk connection.
 *
 * After the trunk has been allocated, #trunk_request_alloc and
 * #trunk_request_enqueue should be used to allocate memory for trunk
 * requests, and pass a preq (protocol request) to the trunk for
 * processing.
 *
 * The trunk will then asynchronously process the request, writing the result
 * to a specified rctx.  See #trunk_request_enqueue for more details.
 *
 * @note Trunks may not be shared between multiple threads under any circumstances.
 *
 * @param[in] ctx   To use for any memory allocations.  Must be thread local.
 * @param[in] el    to use for I/O and timer events.
 * @param[in] funcs   Callback functions.
 * @param[in] conf    Common user configurable parameters.
 * @param[in] log_prefix  To prepend to global messages.
 * @param[in] uctx    User data to pass to the alloc function.
 * @param[in] delay_start If true, then we will not spawn any connections
 *        until the first request is enqueued.
 * @param[in] trigger_args  Pairs to pass to trigger requests, if triggers are enabled.
 * @return
 *  - New trunk handle on success.
 *  - NULL on error.
 */
trunk_t *trunk_alloc(TALLOC_CTX *ctx, fr_event_list_t *el,
         trunk_io_funcs_t const *funcs, trunk_conf_t const *conf,
         char const *log_prefix, void const *uctx, bool delay_start, fr_pair_list_t *trigger_args)
{
  trunk_t *trunk;
  size_t    i;

  /*
   *  Check we have the functions we need
   */
  if (!fr_cond_assert(funcs->connection_alloc)) return NULL;

  MEM(trunk = talloc_zero(ctx, trunk_t));
  trunk->el = el;
  trunk->log_prefix = talloc_strdup(trunk, log_prefix);
  trunk->trigger_args = trigger_args;

  memcpy(&trunk->funcs, funcs, sizeof(trunk->funcs));
  if (!trunk->funcs.connection_prioritise) {
    trunk->funcs.connection_prioritise = _trunk_connection_order_by_shortest_queue;
  }
  if (!trunk->funcs.request_prioritise) trunk->funcs.request_prioritise = fr_pointer_cmp;

  memcpy(&trunk->conf, conf, sizeof(trunk->conf));

  memcpy(&trunk->uctx, &uctx, sizeof(trunk->uctx));
  talloc_set_destructor(trunk, _trunk_free);

  /*
   *  Free request list...
   */
  fr_dlist_talloc_init(&trunk->free_requests, trunk_request_t, entry);

  /*
   *  Request backlog queue
   */
  MEM(trunk->backlog = fr_heap_talloc_alloc(trunk, _trunk_request_prioritise,
               trunk_request_t, heap_id, 0));

  /*
   *  Connection queues and trees
   */
  MEM(trunk->active = fr_minmax_heap_talloc_alloc(trunk, trunk->funcs.connection_prioritise,
              trunk_connection_t, heap_id, 0));
  fr_dlist_talloc_init(&trunk->init, trunk_connection_t, entry);
  fr_dlist_talloc_init(&trunk->connecting, trunk_connection_t, entry);
  fr_dlist_talloc_init(&trunk->full, trunk_connection_t, entry);
  fr_dlist_talloc_init(&trunk->inactive, trunk_connection_t, entry);
  fr_dlist_talloc_init(&trunk->inactive_draining, trunk_connection_t, entry);
  fr_dlist_talloc_init(&trunk->closed, trunk_connection_t, entry);
  fr_dlist_talloc_init(&trunk->draining, trunk_connection_t, entry);
  fr_dlist_talloc_init(&trunk->draining_to_free, trunk_connection_t, entry);
  fr_dlist_talloc_init(&trunk->to_free, trunk_connection_t, entry);

  /*
   *  Watch lists
   */
  for (i = 0; i < NUM_ELEMENTS(trunk->watch); i++) {
    fr_dlist_talloc_init(&trunk->watch[i], trunk_watch_entry_t, entry);
  }

  DEBUG4("Trunk allocated %p", trunk);

  if (!delay_start) {
    if (trunk_start(trunk) < 0) {
      talloc_free(trunk);
      return NULL;
    }
  }

  return trunk;
}

/** Check for a module trigger section when parsing the `triggers` option.
 *
 */
int trunk_trigger_cf_parse(TALLOC_CTX *ctx, void *out, void *parent, CONF_ITEM *ci, conf_parser_t const *rule)
{
  trunk_conf_t  *conf = parent;
  CONF_SECTION  *cs = cf_item_to_section(cf_parent(ci));

  if (cf_pair_parse_value(ctx, out, parent, ci, rule)< 0) return -1;

  /*
   *  If the parent section of the `triggers` option contains a trigger
   *  section then store it as the module CONF SECTION for the appropriate
   *  trigger group.
   */
  if (cf_section_find(cs, "trigger", NULL)) {
    if (strcmp(cf_section_name(cs), "request") == 0) {
      conf->req_trigger_cs = cs;
    } else {
      conf->conn_trigger_cs = cs;
    }
  }

  return 0;
}

#ifndef TALLOC_GET_TYPE_ABORT_NOOP
/** Verify a trunk
 *
 * A trunk has some number of connections, which each have some number of requests. The connections and
 * requests are in differing kinds of containers depending on their state and how they are used, and may
 * have fields that can only be validated by comparison with a parent.  We had planned on passing a "context"
 * down with the ancestral values, but that breaks the foo_verify() API. Each foo_verify() will only verify the
 * foo's children.
 */
void trunk_verify(char const *file, int line, trunk_t *trunk)
{
  fr_fatal_assert_msg(trunk, "CONSISTENCY CHECK FAILED %s[%i]: trunk_t pointer was NULL", file, line);
  (void) talloc_get_type_abort(trunk, trunk_t);

  for (size_t i = 0; i < NUM_ELEMENTS(trunk->watch); i++) {
    _fr_dlist_verify(file, line, &trunk->watch[i]);
  }

#define IO_FUNC_VERIFY(_func) \
  fr_fatal_assert_msg(trunk->funcs._func, "CONSISTENCY_CHECK_FAILED %s[%i}: " #_func " was NULL", file, line)

  /*
   *  Only a few of the function pointers *must* be non-NULL..
   */
  IO_FUNC_VERIFY(connection_alloc);
  IO_FUNC_VERIFY(connection_prioritise);
  IO_FUNC_VERIFY(request_prioritise);

#define TRUNK_TCONN_CHECKS(_tconn, _state) \
do { \
  fr_fatal_assert_msg(trunk == _tconn->pub.trunk, \
          "CONSISTENCY_CHECK_FAILED %s[%i}: connection-trunk mismatch", file, line); \
  fr_fatal_assert_msg(_state == _tconn->pub.state, \
          "CONSISTENCY_CHECK_FAILED %s[%i}: connection-state mismatch", file, line); \
} while (0)

#define TCONN_DLIST_VERIFY(_dlist, _state) \
do { \
  _fr_dlist_verify(file, line, &(trunk->_dlist)); \
  fr_dlist_foreach(&(trunk->_dlist), trunk_connection_t, tconn)  { \
    trunk_connection_verify(file, line, tconn); \
    TRUNK_TCONN_CHECKS(tconn, _state); \
  } \
} while (0)

#define TCONN_MINMAX_HEAP_VERIFY(_heap, _state) \
do {\
  fr_minmax_heap_verify(file, line, trunk->_heap); \
  fr_minmax_heap_foreach(trunk->_heap, trunk_connection_t, tconn) { \
    trunk_connection_verify(file, line, tconn); \
    TRUNK_TCONN_CHECKS(tconn, _state); \
  }} \
} while (0)

  fr_dlist_verify(&(trunk->free_requests));
  FR_HEAP_VERIFY(trunk->backlog);

  TCONN_DLIST_VERIFY(init, TRUNK_CONN_INIT);
  TCONN_DLIST_VERIFY(connecting, TRUNK_CONN_CONNECTING);
  TCONN_MINMAX_HEAP_VERIFY(active, TRUNK_CONN_ACTIVE);
  TCONN_DLIST_VERIFY(full, TRUNK_CONN_FULL);
  TCONN_DLIST_VERIFY(inactive, TRUNK_CONN_INACTIVE);
  TCONN_DLIST_VERIFY(inactive_draining, TRUNK_CONN_INACTIVE_DRAINING);
  /* TCONN_DLIST_VERIFY(failed, ???); */
  TCONN_DLIST_VERIFY(closed, TRUNK_CONN_CLOSED);
  TCONN_DLIST_VERIFY(draining, TRUNK_CONN_DRAINING);
  TCONN_DLIST_VERIFY(draining_to_free, TRUNK_CONN_DRAINING_TO_FREE);
  TCONN_DLIST_VERIFY(to_free, TRUNK_CONN_HALTED);
}

void trunk_connection_verify(char const *file, int line, trunk_connection_t *tconn)
{
  fr_fatal_assert_msg(tconn, "CONSISTENCY CHECK FAILED %s[%i]: trunk_connection_t pointer was NULL", file, line);
  (void) talloc_get_type_abort(tconn, trunk_connection_t);

  (void) talloc_get_type_abort(tconn->pub.trunk, trunk_t);

  /*
   *  shouldn't be both in heap and on list--but it doesn't look like moves
   *  to active heap wipe the dlist pointers.
   */

#define TCONN_TREQ_CHECKS(_treq, _state) \
do { \
  fr_fatal_assert_msg(tconn == _treq->pub.tconn, \
          "CONSISTENCY_CHECK_FAILED %s[%i}: trunk request-tconn mismatch", file, line); \
  fr_fatal_assert_msg(tconn->pub.trunk == _treq->pub.trunk, \
          "CONSISTENCY_CHECK_FAILED %s[%i}: trunk request-trunk mismatch", file, line); \
  fr_fatal_assert_msg(_state == _treq->pub.state, \
          "CONSISTENCY_CHECK_FAILED %s[%i}: trunk request-state mismatch", file, line); \
} while (0)

#define TREQ_DLIST_VERIFY(_dlist, _state) \
do { \
  _fr_dlist_verify(file, line, &(tconn->_dlist)); \
  fr_dlist_foreach(&(tconn->_dlist), trunk_request_t, treq)  { \
    trunk_request_verify(file, line, treq); \
    TCONN_TREQ_CHECKS(treq, _state); \
  } \
} while (0)

#define TREQ_HEAP_VERIFY(_heap, _state) \
do { \
  fr_heap_iter_t _iter; \
  fr_heap_verify(file, line, tconn->_heap); \
  for (trunk_request_t *treq = fr_heap_iter_init(tconn->_heap, &_iter); \
       treq; \
       treq = fr_heap_iter_next(tconn->_heap, &_iter)) { \
    trunk_request_verify(file, line, treq); \
    TCONN_TREQ_CHECKS(treq, _state); \
  } \
} while (0)

#define TREQ_OPTION_VERIFY(_option, _state) \
do { \
  if (tconn->_option) { \
    trunk_request_verify(file, line, tconn->_option); \
    TCONN_TREQ_CHECKS(tconn->_option, _state); \
  } \
} while (0)

  /* verify associated requests */
  TREQ_HEAP_VERIFY(pending, TRUNK_REQUEST_STATE_PENDING);
  TREQ_DLIST_VERIFY(sent, TRUNK_REQUEST_STATE_SENT);
  TREQ_DLIST_VERIFY(cancel, TRUNK_REQUEST_STATE_CANCEL);
  TREQ_DLIST_VERIFY(cancel_sent, TRUNK_REQUEST_STATE_CANCEL_SENT);
  TREQ_OPTION_VERIFY(partial, TRUNK_REQUEST_STATE_PARTIAL);
  TREQ_OPTION_VERIFY(cancel_partial, TRUNK_REQUEST_STATE_CANCEL_PARTIAL);
}

void trunk_request_verify(char const *file, int line, trunk_request_t *treq)
{
  fr_fatal_assert_msg(treq, "CONSISTENCY CHECK FAILED %s[%i]: trunk_request_t pointer was NULL", file, line);
  (void) talloc_get_type_abort(treq, trunk_request_t);

#ifdef WITH_VERIFY_PTR
  if (treq->pub.request) request_verify(file, line, treq->pub.request);
#endif
}


bool trunk_search(trunk_t *trunk, void *ptr)
{
#define TCONN_DLIST_SEARCH(_dlist) \
do { \
  fr_dlist_foreach(&(trunk->_dlist), trunk_connection_t, tconn)  { \
    if (ptr == tconn) { \
      fr_fprintf(stderr, "trunk_search: tconn %p on " #_dlist "\n", ptr); \
      return true; \
    } \
    if (trunk_connection_search(tconn, ptr)) { \
      fr_fprintf(stderr, " in tconn %p on " #_dlist "\n", tconn); \
      return true; \
    } \
  } \
} while (0)

#define TCONN_MINMAX_HEAP_SEARCH(_heap) \
do { \
  fr_minmax_heap_foreach(trunk->_heap, trunk_connection_t, tconn) { \
    if (ptr == tconn) { \
      fr_fprintf(stderr, "trunk_search: tconn %p on " #_heap "\n", ptr); \
      return true; \
    } \
    if (trunk_connection_search(tconn, ptr)) { \
      fr_fprintf(stderr, " on tconn %p on " #_heap "\n", tconn); \
      return true; \
    } \
  }}\
} while (0)

  TCONN_DLIST_SEARCH(init);
  TCONN_DLIST_SEARCH(connecting);
  TCONN_MINMAX_HEAP_SEARCH(active);
  TCONN_DLIST_SEARCH(full);
  TCONN_DLIST_SEARCH(inactive);
  TCONN_DLIST_SEARCH(inactive_draining);
  TCONN_DLIST_SEARCH(failed);
  TCONN_DLIST_SEARCH(closed);
  TCONN_DLIST_SEARCH(draining);
  TCONN_DLIST_SEARCH(draining_to_free);
  TCONN_DLIST_SEARCH(to_free);

  return false;
}

bool trunk_connection_search(trunk_connection_t *tconn, void *ptr)
{
#define TREQ_DLIST_SEARCH(_dlist) \
do { \
  fr_dlist_foreach(&(tconn->_dlist), trunk_request_t, treq)  { \
    if (ptr == treq) { \
      fr_fprintf(stderr, "trunk_search: treq %p on " #_dlist "\n", ptr); \
      return true; \
    } \
    if (trunk_request_search(treq, ptr)) { \
      fr_fprintf(stderr, "trunk_search: preq %p found on " #_dlist, ptr); \
      return true; \
    } \
  } \
} while (0)

#define TREQ_HEAP_SEARCH(_heap) \
do { \
  fr_heap_iter_t _iter; \
  for (trunk_request_t *treq = fr_heap_iter_init(tconn->_heap, &_iter); \
       treq; \
       treq = fr_heap_iter_next(tconn->_heap, &_iter)) { \
    if (ptr == treq) { \
      fr_fprintf(stderr, "trunk_search: treq %p in " #_heap "\n", ptr); \
      return true; \
    } \
    if (trunk_request_search(treq, ptr)) { \
      fr_fprintf(stderr, "trunk_search: preq %p found in " #_heap, ptr); \
      return true; \
    } \
  } \
} while (0)

#define TREQ_OPTION_SEARCH(_option) \
do { \
  if (tconn->_option) { \
    if (ptr == tconn->_option) { \
      fr_fprintf(stderr, "trunk_search: treq %p is " #_option "\n", ptr); \
      return true; \
    } \
    if (trunk_request_search(tconn->_option, ptr)) { \
      fr_fprintf(stderr, "trunk_search: preq %p found in " #_option, ptr); \
      return true; \
    } \
  } \
} while (0)

  /* search associated requests */
  TREQ_HEAP_SEARCH(pending);
  TREQ_DLIST_SEARCH(sent);
  TREQ_DLIST_SEARCH(cancel);
  TREQ_DLIST_SEARCH(cancel_sent);
  TREQ_OPTION_SEARCH(partial);
  TREQ_OPTION_SEARCH(cancel_partial);

  return false;
}

bool trunk_request_search(trunk_request_t *treq, void *ptr)
{
  return treq->pub.preq == ptr;
}
#endif

Coverage Report

Created: 2026-08-08 07:10