path: root/fluent-bit/lib/librdkafka-2.1.0/src/rdkafka_partition.c

/*
 * librdkafka - The Apache Kafka C/C++ library
 *
 * Copyright (c) 2015 Magnus Edenhill
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice,
 *    this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 *    this list of conditions and the following disclaimer in the documentation
 *    and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 * POSSIBILITY OF SUCH DAMAGE.
 */
#include "rdkafka_int.h"
#include "rdkafka_topic.h"
#include "rdkafka_broker.h"
#include "rdkafka_request.h"
#include "rdkafka_offset.h"
#include "rdkafka_partition.h"
#include "rdkafka_fetcher.h"
#include "rdregex.h"
#include "rdports.h" /* rd_qsort_r() */

#include "rdunittest.h"

const char *rd_kafka_fetch_states[] = {"none",        "stopping",
                                       "stopped",     "offset-query",
                                       "offset-wait", "validate-epoch-wait",
                                       "active"};


static rd_kafka_op_res_t rd_kafka_toppar_op_serve(rd_kafka_t *rk,
                                                  rd_kafka_q_t *rkq,
                                                  rd_kafka_op_t *rko,
                                                  rd_kafka_q_cb_type_t cb_type,
                                                  void *opaque);

static void rd_kafka_toppar_offset_retry(rd_kafka_toppar_t *rktp,
                                         int backoff_ms,
                                         const char *reason);


static RD_INLINE int32_t
rd_kafka_toppar_version_new_barrier0(rd_kafka_toppar_t *rktp,
                                     const char *func,
                                     int line) {
        int32_t version = rd_atomic32_add(&rktp->rktp_version, 1);
        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "BARRIER",
                     "%s [%" PRId32 "]: %s:%d: new version barrier v%" PRId32,
                     rktp->rktp_rkt->rkt_topic->str, rktp->rktp_partition, func,
                     line, version);
        return version;
}

#define rd_kafka_toppar_version_new_barrier(rktp)                              \
        rd_kafka_toppar_version_new_barrier0(rktp, __FUNCTION__, __LINE__)


/**
 * Toppar based OffsetResponse handling.
 * This is used for updating the low water mark for consumer lag.
 */
static void rd_kafka_toppar_lag_handle_Offset(rd_kafka_t *rk,
                                              rd_kafka_broker_t *rkb,
                                              rd_kafka_resp_err_t err,
                                              rd_kafka_buf_t *rkbuf,
                                              rd_kafka_buf_t *request,
                                              void *opaque) {
        rd_kafka_toppar_t *rktp = opaque;
        rd_kafka_topic_partition_list_t *offsets;
        rd_kafka_topic_partition_t *rktpar;

        offsets = rd_kafka_topic_partition_list_new(1);

        /* Parse and return Offset */
        err = rd_kafka_handle_ListOffsets(rk, rkb, err, rkbuf, request, offsets,
                                          NULL);

        if (err == RD_KAFKA_RESP_ERR__IN_PROGRESS) {
                rd_kafka_topic_partition_list_destroy(offsets);
                return; /* Retrying */
        }

        if (!err && !(rktpar = rd_kafka_topic_partition_list_find(
                          offsets, rktp->rktp_rkt->rkt_topic->str,
                          rktp->rktp_partition)))
                err = RD_KAFKA_RESP_ERR__UNKNOWN_PARTITION;

        if (!err && !rktpar->err) {
                rd_kafka_toppar_lock(rktp);
                rktp->rktp_lo_offset = rktpar->offset;
                rd_kafka_toppar_unlock(rktp);
        }

        rd_kafka_topic_partition_list_destroy(offsets);

        rktp->rktp_wait_consumer_lag_resp = 0;

        rd_kafka_toppar_destroy(rktp); /* from request.opaque */
}



/**
 * Request information from broker to keep track of consumer lag.
 *
 * @locality toppar handle thread
 * @locks none
 */
static void rd_kafka_toppar_consumer_lag_req(rd_kafka_toppar_t *rktp) {
        rd_kafka_topic_partition_list_t *partitions;
        rd_kafka_topic_partition_t *rktpar;

        if (rktp->rktp_wait_consumer_lag_resp)
                return; /* Previous request not finished yet */

        rd_kafka_toppar_lock(rktp);

        /* Offset requests can only be sent to the leader replica.
         *
         * Note: If rktp is delegated to a preferred replica, it is
         * certain that FETCH >= v5 and so rktp_lo_offset will be
         * updated via LogStartOffset in the FETCH response.
         */
        if (!rktp->rktp_leader || (rktp->rktp_leader != rktp->rktp_broker)) {
                rd_kafka_toppar_unlock(rktp);
                return;
        }

        /* Also don't send a timed log start offset request if leader
         * broker supports FETCH >= v5, since this will be set when
         * doing fetch requests.
         */
        if (rd_kafka_broker_ApiVersion_supported(
                rktp->rktp_broker, RD_KAFKAP_Fetch, 0, 5, NULL) == 5) {
                rd_kafka_toppar_unlock(rktp);
                return;
        }

        rktp->rktp_wait_consumer_lag_resp = 1;

        partitions = rd_kafka_topic_partition_list_new(1);
        rktpar     = rd_kafka_topic_partition_list_add(
            partitions, rktp->rktp_rkt->rkt_topic->str, rktp->rktp_partition);
        rktpar->offset = RD_KAFKA_OFFSET_BEGINNING;
        rd_kafka_topic_partition_set_current_leader_epoch(
            rktpar, rktp->rktp_leader_epoch);

        /* Ask for oldest offset. The newest offset is automatically
         * propagated in FetchResponse.HighwaterMark. */
        rd_kafka_ListOffsetsRequest(
            rktp->rktp_broker, partitions, RD_KAFKA_REPLYQ(rktp->rktp_ops, 0),
            rd_kafka_toppar_lag_handle_Offset, rd_kafka_toppar_keep(rktp));

        rd_kafka_toppar_unlock(rktp);

        rd_kafka_topic_partition_list_destroy(partitions);
}



/**
 * Request earliest offset for a partition
 *
 * Locality: toppar handler thread
 */
static void rd_kafka_toppar_consumer_lag_tmr_cb(rd_kafka_timers_t *rkts,
                                                void *arg) {
        rd_kafka_toppar_t *rktp = arg;
        rd_kafka_toppar_consumer_lag_req(rktp);
}

/**
 * @brief Update rktp_op_version.
 *        Enqueue an RD_KAFKA_OP_BARRIER type of operation
 *        when the op_version is updated.
 *
 * @locks_required rd_kafka_toppar_lock() must be held.
 * @locality Toppar handler thread
 */
void rd_kafka_toppar_op_version_bump(rd_kafka_toppar_t *rktp, int32_t version) {
        rd_kafka_op_t *rko;

        rktp->rktp_op_version = version;
        rko                   = rd_kafka_op_new(RD_KAFKA_OP_BARRIER);
        rko->rko_version      = version;
        rko->rko_prio         = RD_KAFKA_PRIO_FLASH;
        rko->rko_rktp         = rd_kafka_toppar_keep(rktp);
        rd_kafka_q_enq(rktp->rktp_fetchq, rko);
}


/**
 * Add new partition to topic.
 *
 * Locks: rd_kafka_topic_wrlock() must be held.
 * Locks: rd_kafka_wrlock() must be held.
 */
rd_kafka_toppar_t *rd_kafka_toppar_new0(rd_kafka_topic_t *rkt,
                                        int32_t partition,
                                        const char *func,
                                        int line) {
        rd_kafka_toppar_t *rktp;

        rktp = rd_calloc(1, sizeof(*rktp));

        rktp->rktp_partition    = partition;
        rktp->rktp_rkt          = rkt;
        rktp->rktp_leader_id    = -1;
        rktp->rktp_broker_id    = -1;
        rktp->rktp_leader_epoch = -1;
        rd_interval_init(&rktp->rktp_lease_intvl);
        rd_interval_init(&rktp->rktp_new_lease_intvl);
        rd_interval_init(&rktp->rktp_new_lease_log_intvl);
        rd_interval_init(&rktp->rktp_metadata_intvl);
        /* Mark partition as unknown (does not exist) until we see the
         * partition in topic metadata. */
        if (partition != RD_KAFKA_PARTITION_UA)
                rktp->rktp_flags |= RD_KAFKA_TOPPAR_F_UNKNOWN;
        rktp->rktp_fetch_state = RD_KAFKA_TOPPAR_FETCH_NONE;
        rktp->rktp_fetch_msg_max_bytes =
            rkt->rkt_rk->rk_conf.fetch_msg_max_bytes;
        rktp->rktp_offset_fp = NULL;
        rd_kafka_offset_stats_reset(&rktp->rktp_offsets);
        rd_kafka_offset_stats_reset(&rktp->rktp_offsets_fin);
        rktp->rktp_ls_offset = RD_KAFKA_OFFSET_INVALID;
        rktp->rktp_hi_offset = RD_KAFKA_OFFSET_INVALID;
        rktp->rktp_lo_offset = RD_KAFKA_OFFSET_INVALID;
        rd_kafka_fetch_pos_init(&rktp->rktp_query_pos);
        rd_kafka_fetch_pos_init(&rktp->rktp_next_fetch_start);
        rd_kafka_fetch_pos_init(&rktp->rktp_last_next_fetch_start);
        rd_kafka_fetch_pos_init(&rktp->rktp_app_pos);
        rd_kafka_fetch_pos_init(&rktp->rktp_stored_pos);
        rd_kafka_fetch_pos_init(&rktp->rktp_committing_pos);
        rd_kafka_fetch_pos_init(&rktp->rktp_committed_pos);
        rd_kafka_msgq_init(&rktp->rktp_msgq);
        rd_kafka_msgq_init(&rktp->rktp_xmit_msgq);
        mtx_init(&rktp->rktp_lock, mtx_plain);

        rd_refcnt_init(&rktp->rktp_refcnt, 0);
        rktp->rktp_fetchq          = rd_kafka_q_new(rkt->rkt_rk);
        rktp->rktp_ops             = rd_kafka_q_new(rkt->rkt_rk);
        rktp->rktp_ops->rkq_serve  = rd_kafka_toppar_op_serve;
        rktp->rktp_ops->rkq_opaque = rktp;
        rd_atomic32_init(&rktp->rktp_version, 1);
        rktp->rktp_op_version = rd_atomic32_get(&rktp->rktp_version);

        rd_atomic32_init(&rktp->rktp_msgs_inflight, 0);
        rd_kafka_pid_reset(&rktp->rktp_eos.pid);

        /* Consumer: If statistics is available we query the log start offset
         * of each partition.
         * Since the oldest offset only moves on log retention, we cap this
         * value on the low end to a reasonable value to avoid flooding
         * the brokers with OffsetRequests when our statistics interval is low.
         * FIXME: Use a global timer to collect offsets for all partitions
         * FIXME: This timer is superfulous for FETCH >= v5 because the log
         *        start offset is included in fetch responses.
         * */
        if (rktp->rktp_rkt->rkt_rk->rk_conf.stats_interval_ms > 0 &&
            rkt->rkt_rk->rk_type == RD_KAFKA_CONSUMER &&
            rktp->rktp_partition != RD_KAFKA_PARTITION_UA) {
                int intvl = rkt->rkt_rk->rk_conf.stats_interval_ms;
                if (intvl < 10 * 1000 /* 10s */)
                        intvl = 10 * 1000;
                rd_kafka_timer_start(
                    &rkt->rkt_rk->rk_timers, &rktp->rktp_consumer_lag_tmr,
                    intvl * 1000ll, rd_kafka_toppar_consumer_lag_tmr_cb, rktp);
        }

        rktp->rktp_rkt = rd_kafka_topic_keep(rkt);

        rd_kafka_q_fwd_set(rktp->rktp_ops, rkt->rkt_rk->rk_ops);
        rd_kafka_dbg(rkt->rkt_rk, TOPIC, "TOPPARNEW",
                     "NEW %s [%" PRId32 "] %p refcnt %p (at %s:%d)",
                     rkt->rkt_topic->str, rktp->rktp_partition, rktp,
                     &rktp->rktp_refcnt, func, line);

        return rd_kafka_toppar_keep(rktp);
}



/**
 * Removes a toppar from its duties, global lists, etc.
 *
 * Locks: rd_kafka_toppar_lock() MUST be held
 */
static void rd_kafka_toppar_remove(rd_kafka_toppar_t *rktp) {
        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "TOPPARREMOVE",
                     "Removing toppar %s [%" PRId32 "] %p",
                     rktp->rktp_rkt->rkt_topic->str, rktp->rktp_partition,
                     rktp);

        rd_kafka_timer_stop(&rktp->rktp_rkt->rkt_rk->rk_timers,
                            &rktp->rktp_validate_tmr, 1 /*lock*/);
        rd_kafka_timer_stop(&rktp->rktp_rkt->rkt_rk->rk_timers,
                            &rktp->rktp_offset_query_tmr, 1 /*lock*/);
        rd_kafka_timer_stop(&rktp->rktp_rkt->rkt_rk->rk_timers,
                            &rktp->rktp_consumer_lag_tmr, 1 /*lock*/);

        rd_kafka_q_fwd_set(rktp->rktp_ops, NULL);
}


/**
 * Final destructor for partition.
 */
void rd_kafka_toppar_destroy_final(rd_kafka_toppar_t *rktp) {

        rd_kafka_toppar_remove(rktp);

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "DESTROY",
                     "%s [%" PRId32 "]: %p DESTROY_FINAL",
                     rktp->rktp_rkt->rkt_topic->str, rktp->rktp_partition,
                     rktp);

        /* Clear queues */
        rd_kafka_assert(rktp->rktp_rkt->rkt_rk,
                        rd_kafka_msgq_len(&rktp->rktp_xmit_msgq) == 0);
        rd_kafka_dr_msgq(rktp->rktp_rkt, &rktp->rktp_msgq,
                         RD_KAFKA_RESP_ERR__DESTROY);
        rd_kafka_q_destroy_owner(rktp->rktp_fetchq);
        rd_kafka_q_destroy_owner(rktp->rktp_ops);

        rd_kafka_replyq_destroy(&rktp->rktp_replyq);

        rd_kafka_topic_destroy0(rktp->rktp_rkt);

        mtx_destroy(&rktp->rktp_lock);

        if (rktp->rktp_leader)
                rd_kafka_broker_destroy(rktp->rktp_leader);

        rd_refcnt_destroy(&rktp->rktp_refcnt);

        rd_free(rktp);
}


/**
 * Set toppar fetching state.
 *
 * @locality any
 * @locks_required rd_kafka_toppar_lock() MUST be held.
 */
void rd_kafka_toppar_set_fetch_state(rd_kafka_toppar_t *rktp, int fetch_state) {
        rd_kafka_assert(NULL,
                        thrd_is_current(rktp->rktp_rkt->rkt_rk->rk_thread));

        if ((int)rktp->rktp_fetch_state == fetch_state)
                return;

        rd_kafka_dbg(
            rktp->rktp_rkt->rkt_rk, TOPIC, "PARTSTATE",
            "Partition %.*s [%" PRId32 "] changed fetch state %s -> %s",
            RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic), rktp->rktp_partition,
            rd_kafka_fetch_states[rktp->rktp_fetch_state],
            rd_kafka_fetch_states[fetch_state]);

        rktp->rktp_fetch_state = fetch_state;

        if (fetch_state == RD_KAFKA_TOPPAR_FETCH_ACTIVE)
                rd_kafka_dbg(
                    rktp->rktp_rkt->rkt_rk, CONSUMER | RD_KAFKA_DBG_TOPIC,
                    "FETCH",
                    "Partition %.*s [%" PRId32 "] start fetching at %s",
                    RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                    rktp->rktp_partition,
                    rd_kafka_fetch_pos2str(rktp->rktp_next_fetch_start));
}


/**
 * Returns the appropriate toppar for a given rkt and partition.
 * The returned toppar has increased refcnt and must be unreffed by calling
 *  rd_kafka_toppar_destroy().
 * May return NULL.
 *
 * If 'ua_on_miss' is true the UA (unassigned) toppar is returned if
 * 'partition' was not known locally, else NULL is returned.
 *
 * Locks: Caller must hold rd_kafka_topic_*lock()
 */
rd_kafka_toppar_t *rd_kafka_toppar_get0(const char *func,
                                        int line,
                                        const rd_kafka_topic_t *rkt,
                                        int32_t partition,
                                        int ua_on_miss) {
        rd_kafka_toppar_t *rktp;

        if (partition >= 0 && partition < rkt->rkt_partition_cnt)
                rktp = rkt->rkt_p[partition];
        else if (partition == RD_KAFKA_PARTITION_UA || ua_on_miss)
                rktp = rkt->rkt_ua;
        else
                return NULL;

        if (rktp)
                return rd_kafka_toppar_keep_fl(func, line, rktp);

        return NULL;
}


/**
 * Same as rd_kafka_toppar_get() but no need for locking and
 * looks up the topic first.
 *
 * Locality: any
 * Locks: none
 */
rd_kafka_toppar_t *rd_kafka_toppar_get2(rd_kafka_t *rk,
                                        const char *topic,
                                        int32_t partition,
                                        int ua_on_miss,
                                        int create_on_miss) {
        rd_kafka_topic_t *rkt;
        rd_kafka_toppar_t *rktp;

        rd_kafka_wrlock(rk);

        /* Find or create topic */
        if (unlikely(!(rkt = rd_kafka_topic_find(rk, topic, 0 /*no-lock*/)))) {
                if (!create_on_miss) {
                        rd_kafka_wrunlock(rk);
                        return NULL;
                }
                rkt = rd_kafka_topic_new0(rk, topic, NULL, NULL, 0 /*no-lock*/);
                if (!rkt) {
                        rd_kafka_wrunlock(rk);
                        rd_kafka_log(rk, LOG_ERR, "TOPIC",
                                     "Failed to create local topic \"%s\": %s",
                                     topic, rd_strerror(errno));
                        return NULL;
                }
        }

        rd_kafka_wrunlock(rk);

        rd_kafka_topic_wrlock(rkt);
        rktp = rd_kafka_toppar_desired_add(rkt, partition);
        rd_kafka_topic_wrunlock(rkt);

        rd_kafka_topic_destroy0(rkt);

        return rktp;
}


/**
 * Returns a toppar if it is available in the cluster.
 * '*errp' is set to the error-code if lookup fails.
 *
 * Locks: topic_*lock() MUST be held
 */
rd_kafka_toppar_t *rd_kafka_toppar_get_avail(const rd_kafka_topic_t *rkt,
                                             int32_t partition,
                                             int ua_on_miss,
                                             rd_kafka_resp_err_t *errp) {
        rd_kafka_toppar_t *rktp;

        switch (rkt->rkt_state) {
        case RD_KAFKA_TOPIC_S_UNKNOWN:
                /* No metadata received from cluster yet.
                 * Put message in UA partition and re-run partitioner when
                 * cluster comes up. */
                partition = RD_KAFKA_PARTITION_UA;
                break;

        case RD_KAFKA_TOPIC_S_NOTEXISTS:
                /* Topic not found in cluster.
                 * Fail message immediately. */
                *errp = RD_KAFKA_RESP_ERR__UNKNOWN_TOPIC;
                return NULL;

        case RD_KAFKA_TOPIC_S_ERROR:
                /* Permanent topic error. */
                *errp = rkt->rkt_err;
                return NULL;

        case RD_KAFKA_TOPIC_S_EXISTS:
                /* Topic exists in cluster. */

                /* Topic exists but has no partitions.
                 * This is usually an transient state following the
                 * auto-creation of a topic. */
                if (unlikely(rkt->rkt_partition_cnt == 0)) {
                        partition = RD_KAFKA_PARTITION_UA;
                        break;
                }

                /* Check that partition exists. */
                if (partition >= rkt->rkt_partition_cnt) {
                        *errp = RD_KAFKA_RESP_ERR__UNKNOWN_PARTITION;
                        return NULL;
                }
                break;

        default:
                rd_kafka_assert(rkt->rkt_rk, !*"NOTREACHED");
                break;
        }

        /* Get new partition */
        rktp = rd_kafka_toppar_get(rkt, partition, 0);

        if (unlikely(!rktp)) {
                /* Unknown topic or partition */
                if (rkt->rkt_state == RD_KAFKA_TOPIC_S_NOTEXISTS)
                        *errp = RD_KAFKA_RESP_ERR__UNKNOWN_TOPIC;
                else
                        *errp = RD_KAFKA_RESP_ERR__UNKNOWN_PARTITION;

                return NULL;
        }

        return rktp;
}


/**
 * Looks for partition 'i' in topic 'rkt's desired list.
 *
 * The desired partition list is the list of partitions that are desired
 * (e.g., by the consumer) but not yet seen on a broker.
 * As soon as the partition is seen on a broker the toppar is moved from
 * the desired list and onto the normal rkt_p array.
 * When the partition on the broker goes away a desired partition is put
 * back on the desired list.
 *
 * Locks: rd_kafka_topic_*lock() must be held.
 * Note: 'rktp' refcount is increased.
 */

rd_kafka_toppar_t *rd_kafka_toppar_desired_get(rd_kafka_topic_t *rkt,
                                               int32_t partition) {
        rd_kafka_toppar_t *rktp;
        int i;

        RD_LIST_FOREACH(rktp, &rkt->rkt_desp, i) {
                if (rktp->rktp_partition == partition)
                        return rd_kafka_toppar_keep(rktp);
        }

        return NULL;
}


/**
 * Link toppar on desired list.
 *
 * Locks: rd_kafka_topic_wrlock() and toppar_lock() must be held.
 */
void rd_kafka_toppar_desired_link(rd_kafka_toppar_t *rktp) {

        if (rktp->rktp_flags & RD_KAFKA_TOPPAR_F_ON_DESP)
                return; /* Already linked */

        rd_kafka_toppar_keep(rktp);
        rd_list_add(&rktp->rktp_rkt->rkt_desp, rktp);
        rd_interval_reset(&rktp->rktp_rkt->rkt_desp_refresh_intvl);
        rktp->rktp_flags |= RD_KAFKA_TOPPAR_F_ON_DESP;
}

/**
 * Unlink toppar from desired list.
 *
 * Locks: rd_kafka_topic_wrlock() and toppar_lock() must be held.
 */
void rd_kafka_toppar_desired_unlink(rd_kafka_toppar_t *rktp) {
        if (!(rktp->rktp_flags & RD_KAFKA_TOPPAR_F_ON_DESP))
                return; /* Not linked */

        rktp->rktp_flags &= ~RD_KAFKA_TOPPAR_F_ON_DESP;
        rd_list_remove(&rktp->rktp_rkt->rkt_desp, rktp);
        rd_interval_reset(&rktp->rktp_rkt->rkt_desp_refresh_intvl);
        rd_kafka_toppar_destroy(rktp);
}


/**
 * @brief If rktp is not already desired:
 *  - mark as DESIRED|~REMOVE
 *  - add to desired list if unknown
 *
 * @remark toppar_lock() MUST be held
 */
void rd_kafka_toppar_desired_add0(rd_kafka_toppar_t *rktp) {
        if ((rktp->rktp_flags & RD_KAFKA_TOPPAR_F_DESIRED))
                return;

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "DESIRED",
                     "%s [%" PRId32 "]: marking as DESIRED",
                     rktp->rktp_rkt->rkt_topic->str, rktp->rktp_partition);

        /* If toppar was marked for removal this is no longer
         * the case since the partition is now desired. */
        rktp->rktp_flags &= ~RD_KAFKA_TOPPAR_F_REMOVE;

        rktp->rktp_flags |= RD_KAFKA_TOPPAR_F_DESIRED;

        if (rktp->rktp_flags & RD_KAFKA_TOPPAR_F_UNKNOWN) {
                rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "DESIRED",
                             "%s [%" PRId32 "]: adding to DESIRED list",
                             rktp->rktp_rkt->rkt_topic->str,
                             rktp->rktp_partition);
                rd_kafka_toppar_desired_link(rktp);
        }
}


/**
 * Adds 'partition' as a desired partition to topic 'rkt', or updates
 * an existing partition to be desired.
 *
 * Locks: rd_kafka_topic_wrlock() must be held.
 */
rd_kafka_toppar_t *rd_kafka_toppar_desired_add(rd_kafka_topic_t *rkt,
                                               int32_t partition) {
        rd_kafka_toppar_t *rktp;

        rktp = rd_kafka_toppar_get(rkt, partition, 0 /*no_ua_on_miss*/);

        if (!rktp)
                rktp = rd_kafka_toppar_desired_get(rkt, partition);

        if (!rktp)
                rktp = rd_kafka_toppar_new(rkt, partition);

        rd_kafka_toppar_lock(rktp);
        rd_kafka_toppar_desired_add0(rktp);
        rd_kafka_toppar_unlock(rktp);

        return rktp; /* Callers refcount */
}



/**
 * Unmarks an 'rktp' as desired.
 *
 * Locks: rd_kafka_topic_wrlock() and rd_kafka_toppar_lock() MUST be held.
 */
void rd_kafka_toppar_desired_del(rd_kafka_toppar_t *rktp) {

        if (!(rktp->rktp_flags & RD_KAFKA_TOPPAR_F_DESIRED))
                return;

        rktp->rktp_flags &= ~RD_KAFKA_TOPPAR_F_DESIRED;
        rd_kafka_toppar_desired_unlink(rktp);

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "DESP",
                     "Removing (un)desired topic %s [%" PRId32 "]",
                     rktp->rktp_rkt->rkt_topic->str, rktp->rktp_partition);

        if (rktp->rktp_flags & RD_KAFKA_TOPPAR_F_UNKNOWN) {
                /* If this partition does not exist in the cluster
                 * and is no longer desired, remove it. */
                rd_kafka_toppar_broker_leave_for_remove(rktp);
        }
}



/**
 * Append message at tail of 'rktp' message queue.
 */
void rd_kafka_toppar_enq_msg(rd_kafka_toppar_t *rktp,
                             rd_kafka_msg_t *rkm,
                             rd_ts_t now) {
        rd_kafka_q_t *wakeup_q = NULL;

        rd_kafka_toppar_lock(rktp);

        if (!rkm->rkm_u.producer.msgid &&
            rktp->rktp_partition != RD_KAFKA_PARTITION_UA)
                rkm->rkm_u.producer.msgid = ++rktp->rktp_msgid;

        if (rktp->rktp_partition == RD_KAFKA_PARTITION_UA ||
            rktp->rktp_rkt->rkt_conf.queuing_strategy == RD_KAFKA_QUEUE_FIFO) {
                /* No need for enq_sorted(), this is the oldest message. */
                rd_kafka_msgq_enq(&rktp->rktp_msgq, rkm);
        } else {
                rd_kafka_msgq_enq_sorted(rktp->rktp_rkt, &rktp->rktp_msgq, rkm);
        }

        if (unlikely(rktp->rktp_partition != RD_KAFKA_PARTITION_UA &&
                     rd_kafka_msgq_may_wakeup(&rktp->rktp_msgq, now) &&
                     (wakeup_q = rktp->rktp_msgq_wakeup_q))) {
                /* Wake-up broker thread */
                rktp->rktp_msgq.rkmq_wakeup.signalled = rd_true;
                rd_kafka_q_keep(wakeup_q);
        }

        rd_kafka_toppar_unlock(rktp);

        if (unlikely(wakeup_q != NULL)) {
                rd_kafka_q_yield(wakeup_q);
                rd_kafka_q_destroy(wakeup_q);
        }
}


/**
 * @brief Insert \p srcq before \p insert_before in \p destq.
 *
 * If \p srcq and \p destq overlaps only part of the \p srcq will be inserted.
 *
 * Upon return \p srcq will contain any remaining messages that require
 * another insert position in \p destq.
 */
static void rd_kafka_msgq_insert_msgq_before(rd_kafka_msgq_t *destq,
                                             rd_kafka_msg_t *insert_before,
                                             rd_kafka_msgq_t *srcq,
                                             int (*cmp)(const void *a,
                                                        const void *b)) {
        rd_kafka_msg_t *slast;
        rd_kafka_msgq_t tmpq;

        if (!insert_before) {
                /* Append all of srcq to destq */
                rd_kafka_msgq_concat(destq, srcq);
                rd_kafka_msgq_verify_order(NULL, destq, 0, rd_false);
                return;
        }

        slast = rd_kafka_msgq_last(srcq);
        rd_dassert(slast);

        if (cmp(slast, insert_before) > 0) {
                rd_kafka_msg_t *new_sfirst;
                int cnt;
                int64_t bytes;

                /* destq insert_before resides somewhere between
                 * srcq.first and srcq.last, find the first message in
                 * srcq that is > insert_before and split srcq into
                 * a left part that contains the messages to insert before
                 * insert_before, and a right part that will need another
                 * insert position. */

                new_sfirst = rd_kafka_msgq_find_pos(srcq, NULL, insert_before,
                                                    cmp, &cnt, &bytes);
                rd_assert(new_sfirst);

                /* split srcq into two parts using the divider message */
                rd_kafka_msgq_split(srcq, &tmpq, new_sfirst, cnt, bytes);

                rd_kafka_msgq_verify_order(NULL, srcq, 0, rd_false);
                rd_kafka_msgq_verify_order(NULL, &tmpq, 0, rd_false);
        } else {
                rd_kafka_msgq_init(&tmpq);
        }

        /* srcq now contains messages up to the first message in destq,
         * insert srcq at insert_before in destq. */
        rd_dassert(!TAILQ_EMPTY(&destq->rkmq_msgs));
        rd_dassert(!TAILQ_EMPTY(&srcq->rkmq_msgs));
        TAILQ_INSERT_LIST_BEFORE(&destq->rkmq_msgs, insert_before,
                                 &srcq->rkmq_msgs, rd_kafka_msgs_head_s,
                                 rd_kafka_msg_t *, rkm_link);
        destq->rkmq_msg_cnt += srcq->rkmq_msg_cnt;
        destq->rkmq_msg_bytes += srcq->rkmq_msg_bytes;
        srcq->rkmq_msg_cnt   = 0;
        srcq->rkmq_msg_bytes = 0;

        rd_kafka_msgq_verify_order(NULL, destq, 0, rd_false);
        rd_kafka_msgq_verify_order(NULL, srcq, 0, rd_false);

        /* tmpq contains the remaining messages in srcq, move it over. */
        rd_kafka_msgq_move(srcq, &tmpq);

        rd_kafka_msgq_verify_order(NULL, srcq, 0, rd_false);
}


/**
 * @brief Insert all messages from \p srcq into \p destq in their sorted
 *        position (using \p cmp)
 */
void rd_kafka_msgq_insert_msgq(rd_kafka_msgq_t *destq,
                               rd_kafka_msgq_t *srcq,
                               int (*cmp)(const void *a, const void *b)) {
        rd_kafka_msg_t *sfirst, *dlast, *start_pos = NULL;

        if (unlikely(RD_KAFKA_MSGQ_EMPTY(srcq))) {
                /* srcq is empty */
                return;
        }

        if (unlikely(RD_KAFKA_MSGQ_EMPTY(destq))) {
                /* destq is empty, simply move the srcq. */
                rd_kafka_msgq_move(destq, srcq);
                rd_kafka_msgq_verify_order(NULL, destq, 0, rd_false);
                return;
        }

        /* Optimize insertion by bulk-moving messages in place.
         * We know that:
         *  - destq is sorted but might not be continous (1,2,3,7)
         *  - srcq is sorted but might not be continous (4,5,6,8)
         *  - there migt be (multiple) overlaps between the two, e.g:
         *     destq = (1,2,3,7), srcq = (4,5,6,8)
         *  - there may be millions of messages.
         */

        rd_kafka_msgq_verify_order(NULL, destq, 0, rd_false);
        rd_kafka_msgq_verify_order(NULL, srcq, 0, rd_false);

        dlast  = rd_kafka_msgq_last(destq);
        sfirst = rd_kafka_msgq_first(srcq);

        /* Most common case, all of srcq goes after destq */
        if (likely(cmp(dlast, sfirst) < 0)) {
                rd_kafka_msgq_concat(destq, srcq);

                rd_kafka_msgq_verify_order(NULL, destq, 0, rd_false);

                rd_assert(RD_KAFKA_MSGQ_EMPTY(srcq));
                return;
        }

        /* Insert messages from srcq into destq in non-overlapping
         * chunks until srcq is exhausted. */
        while (likely(sfirst != NULL)) {
                rd_kafka_msg_t *insert_before;

                /* Get insert position in destq of first element in srcq */
                insert_before = rd_kafka_msgq_find_pos(destq, start_pos, sfirst,
                                                       cmp, NULL, NULL);

                /* Insert as much of srcq as possible at insert_before */
                rd_kafka_msgq_insert_msgq_before(destq, insert_before, srcq,
                                                 cmp);

                /* Remember the current destq position so the next find_pos()
                 * does not have to re-scan destq and what was
                 * added from srcq. */
                start_pos = insert_before;

                /* For next iteration */
                sfirst = rd_kafka_msgq_first(srcq);

                rd_kafka_msgq_verify_order(NULL, destq, 0, rd_false);
                rd_kafka_msgq_verify_order(NULL, srcq, 0, rd_false);
        }

        rd_kafka_msgq_verify_order(NULL, destq, 0, rd_false);

        rd_assert(RD_KAFKA_MSGQ_EMPTY(srcq));
}


/**
 * @brief Inserts messages from \p srcq according to their sorted position
 *        into \p destq, filtering out messages that can not be retried.
 *
 * @param incr_retry Increment retry count for messages.
 * @param max_retries Maximum retries allowed per message.
 * @param backoff Absolute retry backoff for retried messages.
 *
 * @returns 0 if all messages were retried, or 1 if some messages
 *          could not be retried.
 */
int rd_kafka_retry_msgq(rd_kafka_msgq_t *destq,
                        rd_kafka_msgq_t *srcq,
                        int incr_retry,
                        int max_retries,
                        rd_ts_t backoff,
                        rd_kafka_msg_status_t status,
                        int (*cmp)(const void *a, const void *b)) {
        rd_kafka_msgq_t retryable = RD_KAFKA_MSGQ_INITIALIZER(retryable);
        rd_kafka_msg_t *rkm, *tmp;

        /* Scan through messages to see which ones are eligible for retry,
         * move the retryable ones to temporary queue and
         * set backoff time for first message and optionally
         * increase retry count for each message.
         * Sorted insert is not necessary since the original order
         * srcq order is maintained. */
        TAILQ_FOREACH_SAFE(rkm, &srcq->rkmq_msgs, rkm_link, tmp) {
                if (rkm->rkm_u.producer.retries + incr_retry > max_retries)
                        continue;

                rd_kafka_msgq_deq(srcq, rkm, 1);
                rd_kafka_msgq_enq(&retryable, rkm);

                rkm->rkm_u.producer.ts_backoff = backoff;
                rkm->rkm_u.producer.retries += incr_retry;

                /* Don't downgrade a message from any form of PERSISTED
                 * to NOT_PERSISTED, since the original cause of indicating
                 * PERSISTED can't be changed.
                 * E.g., a previous ack or in-flight timeout. */
                if (likely(!(status == RD_KAFKA_MSG_STATUS_NOT_PERSISTED &&
                             rkm->rkm_status !=
                                 RD_KAFKA_MSG_STATUS_NOT_PERSISTED)))
                        rkm->rkm_status = status;
        }

        /* No messages are retryable */
        if (RD_KAFKA_MSGQ_EMPTY(&retryable))
                return 0;

        /* Insert retryable list at sorted position */
        rd_kafka_msgq_insert_msgq(destq, &retryable, cmp);

        return 1;
}

/**
 * @brief Inserts messages from \p rkmq according to their sorted position
 *        into the partition's message queue.
 *
 * @param incr_retry Increment retry count for messages.
 * @param status Set status on each message.
 *
 * @returns 0 if all messages were retried, or 1 if some messages
 *          could not be retried.
 *
 * @locality Broker thread (but not necessarily the leader broker thread)
 */

int rd_kafka_toppar_retry_msgq(rd_kafka_toppar_t *rktp,
                               rd_kafka_msgq_t *rkmq,
                               int incr_retry,
                               rd_kafka_msg_status_t status) {
        rd_kafka_t *rk  = rktp->rktp_rkt->rkt_rk;
        rd_ts_t backoff = rd_clock() + (rk->rk_conf.retry_backoff_ms * 1000);
        int r;

        if (rd_kafka_terminating(rk))
                return 1;

        rd_kafka_toppar_lock(rktp);
        r = rd_kafka_retry_msgq(&rktp->rktp_msgq, rkmq, incr_retry,
                                rk->rk_conf.max_retries, backoff, status,
                                rktp->rktp_rkt->rkt_conf.msg_order_cmp);
        rd_kafka_toppar_unlock(rktp);

        return r;
}

/**
 * @brief Insert sorted message list \p rkmq at sorted position in \p rktp 's
 *        message queue. The queues must not overlap.
 * @remark \p rkmq will be cleared.
 */
void rd_kafka_toppar_insert_msgq(rd_kafka_toppar_t *rktp,
                                 rd_kafka_msgq_t *rkmq) {
        rd_kafka_toppar_lock(rktp);
        rd_kafka_msgq_insert_msgq(&rktp->rktp_msgq, rkmq,
                                  rktp->rktp_rkt->rkt_conf.msg_order_cmp);
        rd_kafka_toppar_unlock(rktp);
}



/**
 * Helper method for purging queues when removing a toppar.
 * Locks: rd_kafka_toppar_lock() MUST be held
 */
void rd_kafka_toppar_purge_and_disable_queues(rd_kafka_toppar_t *rktp) {
        rd_kafka_q_disable(rktp->rktp_fetchq);
        rd_kafka_q_purge(rktp->rktp_fetchq);
        rd_kafka_q_disable(rktp->rktp_ops);
        rd_kafka_q_purge(rktp->rktp_ops);
}


/**
 * @brief Migrate rktp from (optional) \p old_rkb to (optional) \p new_rkb,
 *        but at least one is required to be non-NULL.
 *
 * This is an async operation.
 *
 * @locks rd_kafka_toppar_lock() MUST be held
 */
static void rd_kafka_toppar_broker_migrate(rd_kafka_toppar_t *rktp,
                                           rd_kafka_broker_t *old_rkb,
                                           rd_kafka_broker_t *new_rkb) {
        rd_kafka_op_t *rko;
        rd_kafka_broker_t *dest_rkb;
        int had_next_broker = rktp->rktp_next_broker ? 1 : 0;

        rd_assert(old_rkb || new_rkb);

        /* Update next broker */
        if (new_rkb)
                rd_kafka_broker_keep(new_rkb);
        if (rktp->rktp_next_broker)
                rd_kafka_broker_destroy(rktp->rktp_next_broker);
        rktp->rktp_next_broker = new_rkb;

        /* If next_broker is set it means there is already an async
         * migration op going on and we should not send a new one
         * but simply change the next_broker (which we did above). */
        if (had_next_broker)
                return;

        /* Revert from offset-wait state back to offset-query
         * prior to leaving the broker to avoid stalling
         * on the new broker waiting for a offset reply from
         * this old broker (that might not come and thus need
         * to time out..slowly) */
        if (rktp->rktp_fetch_state == RD_KAFKA_TOPPAR_FETCH_OFFSET_WAIT)
                rd_kafka_toppar_offset_retry(rktp, 500,
                                             "migrating to new broker");

        if (old_rkb) {
                /* If there is an existing broker for this toppar we let it
                 * first handle its own leave and then trigger the join for
                 * the next broker, if any. */
                rko      = rd_kafka_op_new(RD_KAFKA_OP_PARTITION_LEAVE);
                dest_rkb = old_rkb;
        } else {
                /* No existing broker, send join op directly to new broker. */
                rko      = rd_kafka_op_new(RD_KAFKA_OP_PARTITION_JOIN);
                dest_rkb = new_rkb;
        }

        rko->rko_rktp = rd_kafka_toppar_keep(rktp);

        rd_kafka_dbg(
            rktp->rktp_rkt->rkt_rk, TOPIC, "BRKMIGR",
            "Migrating topic %.*s [%" PRId32
            "] %p from %s to %s "
            "(sending %s to %s)",
            RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic), rktp->rktp_partition,
            rktp, old_rkb ? rd_kafka_broker_name(old_rkb) : "(none)",
            new_rkb ? rd_kafka_broker_name(new_rkb) : "(none)",
            rd_kafka_op2str(rko->rko_type), rd_kafka_broker_name(dest_rkb));

        rd_kafka_q_enq(dest_rkb->rkb_ops, rko);
}


/**
 * Async toppar leave from broker.
 * Only use this when partitions are to be removed.
 *
 * Locks: rd_kafka_toppar_lock() MUST be held
 */
void rd_kafka_toppar_broker_leave_for_remove(rd_kafka_toppar_t *rktp) {
        rd_kafka_op_t *rko;
        rd_kafka_broker_t *dest_rkb;

        rktp->rktp_flags |= RD_KAFKA_TOPPAR_F_REMOVE;

        if (rktp->rktp_next_broker)
                dest_rkb = rktp->rktp_next_broker;
        else if (rktp->rktp_broker)
                dest_rkb = rktp->rktp_broker;
        else {
                rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "TOPPARDEL",
                             "%.*s [%" PRId32
                             "] %p not handled by any broker: "
                             "not sending LEAVE for remove",
                             RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                             rktp->rktp_partition, rktp);
                return;
        }


        /* Revert from offset-wait state back to offset-query
         * prior to leaving the broker to avoid stalling
         * on the new broker waiting for a offset reply from
         * this old broker (that might not come and thus need
         * to time out..slowly) */
        if (rktp->rktp_fetch_state == RD_KAFKA_TOPPAR_FETCH_OFFSET_WAIT)
                rd_kafka_toppar_set_fetch_state(
                    rktp, RD_KAFKA_TOPPAR_FETCH_OFFSET_QUERY);

        rko           = rd_kafka_op_new(RD_KAFKA_OP_PARTITION_LEAVE);
        rko->rko_rktp = rd_kafka_toppar_keep(rktp);

        rd_kafka_dbg(
            rktp->rktp_rkt->rkt_rk, TOPIC, "BRKMIGR",
            "%.*s [%" PRId32 "] %p sending final LEAVE for removal by %s",
            RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic), rktp->rktp_partition,
            rktp, rd_kafka_broker_name(dest_rkb));

        rd_kafka_q_enq(dest_rkb->rkb_ops, rko);
}


/**
 * @brief Delegates toppar 'rktp' to broker 'rkb'. 'rkb' may be NULL to
 *        undelegate broker.
 *
 * @locks Caller must have rd_kafka_toppar_lock(rktp) held.
 */
void rd_kafka_toppar_broker_delegate(rd_kafka_toppar_t *rktp,
                                     rd_kafka_broker_t *rkb) {
        rd_kafka_t *rk        = rktp->rktp_rkt->rkt_rk;
        int internal_fallback = 0;

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "BRKDELGT",
                     "%s [%" PRId32
                     "]: delegate to broker %s "
                     "(rktp %p, term %d, ref %d)",
                     rktp->rktp_rkt->rkt_topic->str, rktp->rktp_partition,
                     rkb ? rkb->rkb_name : "(none)", rktp,
                     rd_kafka_terminating(rk),
                     rd_refcnt_get(&rktp->rktp_refcnt));

        /* Undelegated toppars are delgated to the internal
         * broker for bookkeeping. */
        if (!rkb && !rd_kafka_terminating(rk)) {
                rkb               = rd_kafka_broker_internal(rk);
                internal_fallback = 1;
        }

        if (rktp->rktp_broker == rkb && !rktp->rktp_next_broker) {
                rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "BRKDELGT",
                             "%.*s [%" PRId32
                             "]: not updating broker: "
                             "already on correct broker %s",
                             RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                             rktp->rktp_partition,
                             rkb ? rd_kafka_broker_name(rkb) : "(none)");

                if (internal_fallback)
                        rd_kafka_broker_destroy(rkb);
                return;
        }

        if (rktp->rktp_broker)
                rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "BRKDELGT",
                             "%.*s [%" PRId32
                             "]: no longer delegated to "
                             "broker %s",
                             RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                             rktp->rktp_partition,
                             rd_kafka_broker_name(rktp->rktp_broker));


        if (rkb) {
                rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "BRKDELGT",
                             "%.*s [%" PRId32
                             "]: delegating to broker %s "
                             "for partition with %i messages "
                             "(%" PRIu64 " bytes) queued",
                             RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                             rktp->rktp_partition, rd_kafka_broker_name(rkb),
                             rktp->rktp_msgq.rkmq_msg_cnt,
                             rktp->rktp_msgq.rkmq_msg_bytes);


        } else {
                rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "BRKDELGT",
                             "%.*s [%" PRId32 "]: no broker delegated",
                             RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                             rktp->rktp_partition);
        }

        if (rktp->rktp_broker || rkb)
                rd_kafka_toppar_broker_migrate(rktp, rktp->rktp_broker, rkb);

        if (internal_fallback)
                rd_kafka_broker_destroy(rkb);
}



void rd_kafka_toppar_offset_commit_result(
    rd_kafka_toppar_t *rktp,
    rd_kafka_resp_err_t err,
    rd_kafka_topic_partition_list_t *offsets) {
        if (err)
                rd_kafka_consumer_err(
                    rktp->rktp_fetchq,
                    /* FIXME: propagate broker_id */
                    RD_KAFKA_NODEID_UA, err, 0 /* FIXME:VERSION*/, NULL, rktp,
                    RD_KAFKA_OFFSET_INVALID, "Offset commit failed: %s",
                    rd_kafka_err2str(err));

        rd_kafka_toppar_lock(rktp);
        if (!err)
                rktp->rktp_committed_pos =
                    rd_kafka_topic_partition_get_fetch_pos(&offsets->elems[0]);

        /* When stopping toppars:
         * Final commit is now done (or failed), propagate. */
        if (rktp->rktp_fetch_state == RD_KAFKA_TOPPAR_FETCH_STOPPING)
                rd_kafka_toppar_fetch_stopped(rktp, err);

        rd_kafka_toppar_unlock(rktp);
}



/**
 * Handle the next offset to consume for a toppar.
 * This is used during initial setup when trying to figure out what
 * offset to start consuming from.
 *
 * Locality: toppar handler thread.
 * Locks: toppar_lock(rktp) must be held
 */
void rd_kafka_toppar_next_offset_handle(rd_kafka_toppar_t *rktp,
                                        rd_kafka_fetch_pos_t next_pos) {

        if (RD_KAFKA_OFFSET_IS_LOGICAL(next_pos.offset)) {
                /* Offset storage returned logical offset (e.g. "end"),
                 * look it up. */

                /* Save next offset, even if logical, so that e.g.,
                 * assign(BEGINNING) survives a pause+resume, etc.
                 * See issue #2105. */
                rd_kafka_toppar_set_next_fetch_position(rktp, next_pos);

                rd_kafka_offset_reset(rktp, RD_KAFKA_NODEID_UA, next_pos,
                                      RD_KAFKA_RESP_ERR_NO_ERROR, "update");
                return;
        }

        /* Adjust by TAIL count if, if wanted */
        if (rktp->rktp_query_pos.offset <= RD_KAFKA_OFFSET_TAIL_BASE) {
                int64_t orig_offset = next_pos.offset;
                int64_t tail_cnt    = llabs(rktp->rktp_query_pos.offset -
                                         RD_KAFKA_OFFSET_TAIL_BASE);

                if (tail_cnt > next_pos.offset)
                        next_pos.offset = 0;
                else
                        next_pos.offset -= tail_cnt;

                rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "OFFSET",
                             "OffsetReply for topic %s [%" PRId32
                             "]: "
                             "offset %" PRId64
                             ": adjusting for "
                             "OFFSET_TAIL(%" PRId64 "): effective %s",
                             rktp->rktp_rkt->rkt_topic->str,
                             rktp->rktp_partition, orig_offset, tail_cnt,
                             rd_kafka_fetch_pos2str(next_pos));
        }

        rd_kafka_toppar_set_next_fetch_position(rktp, next_pos);

        rd_kafka_toppar_set_fetch_state(rktp, RD_KAFKA_TOPPAR_FETCH_ACTIVE);

        /* Wake-up broker thread which might be idling on IO */
        if (rktp->rktp_broker)
                rd_kafka_broker_wakeup(rktp->rktp_broker, "ready to fetch");
}



/**
 * Fetch committed offset for a single partition. (simple consumer)
 *
 * Locality: toppar thread
 */
void rd_kafka_toppar_offset_fetch(rd_kafka_toppar_t *rktp,
                                  rd_kafka_replyq_t replyq) {
        rd_kafka_t *rk = rktp->rktp_rkt->rkt_rk;
        rd_kafka_topic_partition_list_t *part;
        rd_kafka_op_t *rko;

        rd_kafka_dbg(rk, TOPIC, "OFFSETREQ",
                     "Partition %.*s [%" PRId32
                     "]: querying cgrp for "
                     "committed offset (opv %d)",
                     RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                     rktp->rktp_partition, replyq.version);

        part = rd_kafka_topic_partition_list_new(1);
        rd_kafka_topic_partition_list_add0(__FUNCTION__, __LINE__, part,
                                           rktp->rktp_rkt->rkt_topic->str,
                                           rktp->rktp_partition, rktp, NULL);

        rko             = rd_kafka_op_new(RD_KAFKA_OP_OFFSET_FETCH);
        rko->rko_rktp   = rd_kafka_toppar_keep(rktp);
        rko->rko_replyq = replyq;

        rko->rko_u.offset_fetch.partitions = part;
        rko->rko_u.offset_fetch.require_stable_offsets =
            rk->rk_conf.isolation_level == RD_KAFKA_READ_COMMITTED;
        rko->rko_u.offset_fetch.do_free = 1;

        rd_kafka_q_enq(rktp->rktp_cgrp->rkcg_ops, rko);
}



/**
 * Toppar based OffsetResponse handling.
 * This is used for finding the next offset to Fetch.
 *
 * Locality: toppar handler thread
 */
static void rd_kafka_toppar_handle_Offset(rd_kafka_t *rk,
                                          rd_kafka_broker_t *rkb,
                                          rd_kafka_resp_err_t err,
                                          rd_kafka_buf_t *rkbuf,
                                          rd_kafka_buf_t *request,
                                          void *opaque) {
        rd_kafka_toppar_t *rktp = opaque;
        rd_kafka_topic_partition_list_t *offsets;
        rd_kafka_topic_partition_t *rktpar;
        int actions = 0;

        rd_kafka_toppar_lock(rktp);
        /* Drop reply from previous partition leader */
        if (err != RD_KAFKA_RESP_ERR__DESTROY && rktp->rktp_broker != rkb)
                err = RD_KAFKA_RESP_ERR__OUTDATED;
        rd_kafka_toppar_unlock(rktp);

        offsets = rd_kafka_topic_partition_list_new(1);

        rd_rkb_dbg(rkb, TOPIC, "OFFSET",
                   "Offset reply for "
                   "topic %.*s [%" PRId32 "] (v%d vs v%d)",
                   RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                   rktp->rktp_partition, request->rkbuf_replyq.version,
                   rktp->rktp_op_version);

        rd_dassert(request->rkbuf_replyq.version > 0);
        if (err != RD_KAFKA_RESP_ERR__DESTROY &&
            rd_kafka_buf_version_outdated(request, rktp->rktp_op_version)) {
                /* Outdated request response, ignore. */
                err = RD_KAFKA_RESP_ERR__OUTDATED;
        }

        /* Parse and return Offset */
        if (err != RD_KAFKA_RESP_ERR__OUTDATED)
                err = rd_kafka_handle_ListOffsets(rk, rkb, err, rkbuf, request,
                                                  offsets, &actions);

        if (!err && !(rktpar = rd_kafka_topic_partition_list_find(
                          offsets, rktp->rktp_rkt->rkt_topic->str,
                          rktp->rktp_partition))) {
                /* Requested partition not found in response */
                err = RD_KAFKA_RESP_ERR__UNKNOWN_PARTITION;
                actions |= RD_KAFKA_ERR_ACTION_PERMANENT;
        }

        if (err) {
                rd_rkb_dbg(rkb, TOPIC, "OFFSET",
                           "Offset reply error for "
                           "topic %.*s [%" PRId32 "] (v%d, %s): %s",
                           RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                           rktp->rktp_partition, request->rkbuf_replyq.version,
                           rd_kafka_err2str(err),
                           rd_kafka_actions2str(actions));

                rd_kafka_topic_partition_list_destroy(offsets);

                if (err == RD_KAFKA_RESP_ERR__DESTROY ||
                    err == RD_KAFKA_RESP_ERR__OUTDATED) {
                        /* Termination or outdated, quick cleanup. */

                        if (err == RD_KAFKA_RESP_ERR__OUTDATED) {
                                rd_kafka_toppar_lock(rktp);
                                rd_kafka_toppar_offset_retry(
                                    rktp, 500, "outdated offset response");
                                rd_kafka_toppar_unlock(rktp);
                        }

                        /* from request.opaque */
                        rd_kafka_toppar_destroy(rktp);
                        return;

                } else if (err == RD_KAFKA_RESP_ERR__IN_PROGRESS)
                        return; /* Retry in progress */


                rd_kafka_toppar_lock(rktp);

                if (!(actions & (RD_KAFKA_ERR_ACTION_RETRY |
                                 RD_KAFKA_ERR_ACTION_REFRESH))) {
                        /* Permanent error. Trigger auto.offset.reset policy
                         * and signal error back to application. */

                        rd_kafka_offset_reset(rktp, rkb->rkb_nodeid,
                                              rktp->rktp_query_pos, err,
                                              "failed to query logical offset");

                        rd_kafka_consumer_err(
                            rktp->rktp_fetchq, rkb->rkb_nodeid, err, 0, NULL,
                            rktp,
                            (rktp->rktp_query_pos.offset <=
                                     RD_KAFKA_OFFSET_TAIL_BASE
                                 ? rktp->rktp_query_pos.offset -
                                       RD_KAFKA_OFFSET_TAIL_BASE
                                 : rktp->rktp_query_pos.offset),
                            "Failed to query logical offset %s: %s",
                            rd_kafka_offset2str(rktp->rktp_query_pos.offset),
                            rd_kafka_err2str(err));

                } else {
                        /* Temporary error. Schedule retry. */
                        char tmp[256];

                        rd_snprintf(
                            tmp, sizeof(tmp),
                            "failed to query logical offset %s: %s",
                            rd_kafka_offset2str(rktp->rktp_query_pos.offset),
                            rd_kafka_err2str(err));

                        rd_kafka_toppar_offset_retry(rktp, 500, tmp);
                }

                rd_kafka_toppar_unlock(rktp);

                rd_kafka_toppar_destroy(rktp); /* from request.opaque */
                return;
        }


        rd_kafka_toppar_lock(rktp);
        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "OFFSET",
                     "Offset %s request for %.*s [%" PRId32
                     "] "
                     "returned offset %s (%" PRId64 ") leader epoch %" PRId32,
                     rd_kafka_offset2str(rktp->rktp_query_pos.offset),
                     RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                     rktp->rktp_partition, rd_kafka_offset2str(rktpar->offset),
                     rktpar->offset,
                     rd_kafka_topic_partition_get_leader_epoch(rktpar));


        rd_kafka_toppar_next_offset_handle(
            rktp, RD_KAFKA_FETCH_POS(
                      rktpar->offset,
                      rd_kafka_topic_partition_get_leader_epoch(rktpar)));
        rd_kafka_toppar_unlock(rktp);

        rd_kafka_topic_partition_list_destroy(offsets);

        rd_kafka_toppar_destroy(rktp); /* from request.opaque */
}


/**
 * @brief An Offset fetch failed (for whatever reason) in
 *        the RD_KAFKA_TOPPAR_FETCH_OFFSET_WAIT state:
 *        set the state back to FETCH_OFFSET_QUERY and start the
 *        offset_query_tmr to trigger a new request eventually.
 *
 * @locality toppar handler thread
 * @locks toppar_lock() MUST be held
 */
static void rd_kafka_toppar_offset_retry(rd_kafka_toppar_t *rktp,
                                         int backoff_ms,
                                         const char *reason) {
        rd_ts_t tmr_next;
        int restart_tmr;

        /* (Re)start timer if not started or the current timeout
         * is larger than \p backoff_ms. */
        tmr_next = rd_kafka_timer_next(&rktp->rktp_rkt->rkt_rk->rk_timers,
                                       &rktp->rktp_offset_query_tmr, 1);

        restart_tmr =
            (tmr_next == -1 || tmr_next > rd_clock() + (backoff_ms * 1000ll));

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "OFFSET",
                     "%s [%" PRId32 "]: %s: %s for %s",
                     rktp->rktp_rkt->rkt_topic->str, rktp->rktp_partition,
                     reason,
                     restart_tmr ? "(re)starting offset query timer"
                                 : "offset query timer already scheduled",
                     rd_kafka_fetch_pos2str(rktp->rktp_query_pos));

        rd_kafka_toppar_set_fetch_state(rktp,
                                        RD_KAFKA_TOPPAR_FETCH_OFFSET_QUERY);

        if (restart_tmr)
                rd_kafka_timer_start(&rktp->rktp_rkt->rkt_rk->rk_timers,
                                     &rktp->rktp_offset_query_tmr,
                                     backoff_ms * 1000ll,
                                     rd_kafka_offset_query_tmr_cb, rktp);
}



/**
 * Send OffsetRequest for toppar.
 *
 * If \p backoff_ms is non-zero only the query timer is started,
 * otherwise a query is triggered directly.
 *
 * Locality: toppar handler thread
 * Locks: toppar_lock() must be held
 */
void rd_kafka_toppar_offset_request(rd_kafka_toppar_t *rktp,
                                    rd_kafka_fetch_pos_t query_pos,
                                    int backoff_ms) {
        rd_kafka_broker_t *rkb;

        rd_kafka_assert(NULL,
                        thrd_is_current(rktp->rktp_rkt->rkt_rk->rk_thread));

        rkb = rktp->rktp_broker;

        if (!backoff_ms && (!rkb || rkb->rkb_source == RD_KAFKA_INTERNAL))
                backoff_ms = 500;

        if (backoff_ms) {
                rd_kafka_toppar_offset_retry(
                    rktp, backoff_ms,
                    !rkb ? "no current leader for partition" : "backoff");
                return;
        }


        rd_kafka_timer_stop(&rktp->rktp_rkt->rkt_rk->rk_timers,
                            &rktp->rktp_offset_query_tmr, 1 /*lock*/);


        if (query_pos.offset == RD_KAFKA_OFFSET_STORED &&
            rktp->rktp_rkt->rkt_conf.offset_store_method ==
                RD_KAFKA_OFFSET_METHOD_BROKER) {
                /*
                 * Get stored offset from broker based storage:
                 * ask cgrp manager for offsets
                 */
                rd_kafka_toppar_offset_fetch(
                    rktp,
                    RD_KAFKA_REPLYQ(rktp->rktp_ops, rktp->rktp_op_version));

        } else {
                rd_kafka_topic_partition_list_t *offsets;
                rd_kafka_topic_partition_t *rktpar;

                /*
                 * Look up logical offset (end,beginning,tail,..)
                 */

                rd_rkb_dbg(rkb, TOPIC, "OFFREQ",
                           "Partition %.*s [%" PRId32
                           "]: querying for logical "
                           "offset %s (opv %d)",
                           RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                           rktp->rktp_partition,
                           rd_kafka_offset2str(query_pos.offset),
                           rktp->rktp_op_version);

                rd_kafka_toppar_keep(rktp); /* refcnt for OffsetRequest opaque*/

                if (query_pos.offset <= RD_KAFKA_OFFSET_TAIL_BASE)
                        query_pos.offset = RD_KAFKA_OFFSET_END;

                offsets = rd_kafka_topic_partition_list_new(1);
                rktpar  = rd_kafka_topic_partition_list_add(
                    offsets, rktp->rktp_rkt->rkt_topic->str,
                    rktp->rktp_partition);
                rd_kafka_topic_partition_set_from_fetch_pos(rktpar, query_pos);
                rd_kafka_topic_partition_set_current_leader_epoch(
                    rktpar, rktp->rktp_leader_epoch);

                rd_kafka_ListOffsetsRequest(
                    rkb, offsets,
                    RD_KAFKA_REPLYQ(rktp->rktp_ops, rktp->rktp_op_version),
                    rd_kafka_toppar_handle_Offset, rktp);

                rd_kafka_topic_partition_list_destroy(offsets);
        }

        rd_kafka_toppar_set_fetch_state(rktp,
                                        RD_KAFKA_TOPPAR_FETCH_OFFSET_WAIT);
}


/**
 * Start fetching toppar.
 *
 * Locality: toppar handler thread
 * Locks: none
 */
static void rd_kafka_toppar_fetch_start(rd_kafka_toppar_t *rktp,
                                        rd_kafka_fetch_pos_t pos,
                                        rd_kafka_op_t *rko_orig) {
        rd_kafka_cgrp_t *rkcg   = rko_orig->rko_u.fetch_start.rkcg;
        rd_kafka_resp_err_t err = 0;
        int32_t version         = rko_orig->rko_version;

        rd_kafka_toppar_lock(rktp);

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "FETCH",
                     "Start fetch for %.*s [%" PRId32
                     "] in "
                     "state %s at %s (v%" PRId32 ")",
                     RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                     rktp->rktp_partition,
                     rd_kafka_fetch_states[rktp->rktp_fetch_state],
                     rd_kafka_fetch_pos2str(pos), version);

        if (rktp->rktp_fetch_state == RD_KAFKA_TOPPAR_FETCH_STOPPING) {
                err = RD_KAFKA_RESP_ERR__PREV_IN_PROGRESS;
                rd_kafka_toppar_unlock(rktp);
                goto err_reply;
        }

        rd_kafka_toppar_op_version_bump(rktp, version);

        if (rkcg) {
                rd_kafka_assert(rktp->rktp_rkt->rkt_rk, !rktp->rktp_cgrp);
                /* Attach toppar to cgrp */
                rktp->rktp_cgrp = rkcg;
                rd_kafka_cgrp_op(rkcg, rktp, RD_KAFKA_NO_REPLYQ,
                                 RD_KAFKA_OP_PARTITION_JOIN, 0);
        }


        if (pos.offset == RD_KAFKA_OFFSET_BEGINNING ||
            pos.offset == RD_KAFKA_OFFSET_END ||
            pos.offset <= RD_KAFKA_OFFSET_TAIL_BASE) {
                rd_kafka_toppar_next_offset_handle(rktp, pos);

        } else if (pos.offset == RD_KAFKA_OFFSET_STORED) {
                rd_kafka_offset_store_init(rktp);

        } else if (pos.offset == RD_KAFKA_OFFSET_INVALID) {
                rd_kafka_offset_reset(rktp, RD_KAFKA_NODEID_UA, pos,
                                      RD_KAFKA_RESP_ERR__NO_OFFSET,
                                      "no previously committed offset "
                                      "available");

        } else {
                rd_kafka_toppar_set_next_fetch_position(rktp, pos);

                rd_kafka_toppar_set_fetch_state(rktp,
                                                RD_KAFKA_TOPPAR_FETCH_ACTIVE);

                /* Wake-up broker thread which might be idling on IO */
                if (rktp->rktp_broker)
                        rd_kafka_broker_wakeup(rktp->rktp_broker,
                                               "fetch start");
        }

        rktp->rktp_offsets_fin.eof_offset = RD_KAFKA_OFFSET_INVALID;

        rd_kafka_toppar_unlock(rktp);

        /* Signal back to caller thread that start has commenced, or err */
err_reply:
        if (rko_orig->rko_replyq.q) {
                rd_kafka_op_t *rko;

                rko = rd_kafka_op_new(RD_KAFKA_OP_FETCH_START);

                rko->rko_err  = err;
                rko->rko_rktp = rd_kafka_toppar_keep(rktp);

                rd_kafka_replyq_enq(&rko_orig->rko_replyq, rko, 0);
        }
}



/**
 * Mark toppar's fetch state as stopped (all decommissioning is done,
 * offsets are stored, etc).
 *
 * Locality: toppar handler thread
 * Locks: toppar_lock(rktp) MUST be held
 */
void rd_kafka_toppar_fetch_stopped(rd_kafka_toppar_t *rktp,
                                   rd_kafka_resp_err_t err) {


        rd_kafka_toppar_set_fetch_state(rktp, RD_KAFKA_TOPPAR_FETCH_STOPPED);

        rktp->rktp_app_pos.offset       = RD_KAFKA_OFFSET_INVALID;
        rktp->rktp_app_pos.leader_epoch = -1;

        if (rktp->rktp_cgrp) {
                /* Detach toppar from cgrp */
                rd_kafka_cgrp_op(rktp->rktp_cgrp, rktp, RD_KAFKA_NO_REPLYQ,
                                 RD_KAFKA_OP_PARTITION_LEAVE, 0);
                rktp->rktp_cgrp = NULL;
        }

        /* Signal back to application thread that stop is done. */
        if (rktp->rktp_replyq.q) {
                rd_kafka_op_t *rko;
                rko =
                    rd_kafka_op_new(RD_KAFKA_OP_FETCH_STOP | RD_KAFKA_OP_REPLY);
                rko->rko_err  = err;
                rko->rko_rktp = rd_kafka_toppar_keep(rktp);

                rd_kafka_replyq_enq(&rktp->rktp_replyq, rko, 0);
        }
}


/**
 * Stop toppar fetcher.
 * This is usually an async operation.
 *
 * Locality: toppar handler thread
 */
void rd_kafka_toppar_fetch_stop(rd_kafka_toppar_t *rktp,
                                rd_kafka_op_t *rko_orig) {
        int32_t version = rko_orig->rko_version;

        rd_kafka_toppar_lock(rktp);

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "FETCH",
                     "Stopping fetch for %.*s [%" PRId32 "] in state %s (v%d)",
                     RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                     rktp->rktp_partition,
                     rd_kafka_fetch_states[rktp->rktp_fetch_state], version);

        rd_kafka_toppar_op_version_bump(rktp, version);

        /* Abort pending offset lookups. */
        if (rktp->rktp_fetch_state == RD_KAFKA_TOPPAR_FETCH_OFFSET_QUERY)
                rd_kafka_timer_stop(&rktp->rktp_rkt->rkt_rk->rk_timers,
                                    &rktp->rktp_offset_query_tmr, 1 /*lock*/);

        /* Clear out the forwarding queue. */
        rd_kafka_q_fwd_set(rktp->rktp_fetchq, NULL);

        /* Assign the future replyq to propagate stop results. */
        rd_kafka_assert(rktp->rktp_rkt->rkt_rk, rktp->rktp_replyq.q == NULL);
        rktp->rktp_replyq = rko_orig->rko_replyq;
        rd_kafka_replyq_clear(&rko_orig->rko_replyq);

        rd_kafka_toppar_set_fetch_state(rktp, RD_KAFKA_TOPPAR_FETCH_STOPPING);

        /* Stop offset store (possibly async).
         * NOTE: will call .._stopped() if store finishes immediately,
         *       so no more operations after this call! */
        rd_kafka_offset_store_stop(rktp);

        rd_kafka_toppar_unlock(rktp);
}


/**
 * Update a toppars offset.
 * The toppar must have been previously FETCH_START:ed
 *
 * Locality: toppar handler thread
 */
void rd_kafka_toppar_seek(rd_kafka_toppar_t *rktp,
                          rd_kafka_fetch_pos_t pos,
                          rd_kafka_op_t *rko_orig) {
        rd_kafka_resp_err_t err = 0;
        int32_t version         = rko_orig->rko_version;

        rd_kafka_toppar_lock(rktp);

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "FETCH",
                     "Seek %.*s [%" PRId32 "] to %s in state %s (v%" PRId32 ")",
                     RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                     rktp->rktp_partition, rd_kafka_fetch_pos2str(pos),
                     rd_kafka_fetch_states[rktp->rktp_fetch_state], version);


        if (rktp->rktp_fetch_state == RD_KAFKA_TOPPAR_FETCH_STOPPING) {
                err = RD_KAFKA_RESP_ERR__PREV_IN_PROGRESS;
                goto err_reply;
        } else if (!RD_KAFKA_TOPPAR_FETCH_IS_STARTED(rktp->rktp_fetch_state)) {
                err = RD_KAFKA_RESP_ERR__STATE;
                goto err_reply;
        } else if (pos.offset == RD_KAFKA_OFFSET_STORED) {
                err = RD_KAFKA_RESP_ERR__INVALID_ARG;
                goto err_reply;
        }

        rd_kafka_toppar_op_version_bump(rktp, version);

        /* Reset app offsets since seek()ing is analogue to a (re)assign(),
         * and we want to avoid using the current app offset on resume()
         * following a seek (#3567). */
        rktp->rktp_app_pos.offset       = RD_KAFKA_OFFSET_INVALID;
        rktp->rktp_app_pos.leader_epoch = -1;

        /* Abort pending offset lookups. */
        if (rktp->rktp_fetch_state == RD_KAFKA_TOPPAR_FETCH_OFFSET_QUERY)
                rd_kafka_timer_stop(&rktp->rktp_rkt->rkt_rk->rk_timers,
                                    &rktp->rktp_offset_query_tmr, 1 /*lock*/);

        if (pos.offset <= 0 || pos.validated) {
                rd_kafka_toppar_next_offset_handle(rktp, pos);
        } else {
                rd_kafka_toppar_set_fetch_state(
                    rktp, RD_KAFKA_TOPPAR_FETCH_VALIDATE_EPOCH_WAIT);
                rd_kafka_toppar_set_next_fetch_position(rktp, pos);
                rd_kafka_offset_validate(rktp, "seek");
        }

        /* Signal back to caller thread that seek has commenced, or err */
err_reply:
        rd_kafka_toppar_unlock(rktp);

        if (rko_orig->rko_replyq.q) {
                rd_kafka_op_t *rko;

                rko = rd_kafka_op_new(RD_KAFKA_OP_SEEK | RD_KAFKA_OP_REPLY);

                rko->rko_err               = err;
                rko->rko_u.fetch_start.pos = rko_orig->rko_u.fetch_start.pos;
                rko->rko_rktp              = rd_kafka_toppar_keep(rktp);

                rd_kafka_replyq_enq(&rko_orig->rko_replyq, rko, 0);
        }
}


/**
 * @brief Pause/resume toppar.
 *
 * This is the internal handler of the pause/resume op.
 *
 * @locality toppar's handler thread
 */
static void rd_kafka_toppar_pause_resume(rd_kafka_toppar_t *rktp,
                                         rd_kafka_op_t *rko_orig) {
        rd_kafka_t *rk  = rktp->rktp_rkt->rkt_rk;
        int pause       = rko_orig->rko_u.pause.pause;
        int flag        = rko_orig->rko_u.pause.flag;
        int32_t version = rko_orig->rko_version;

        rd_kafka_toppar_lock(rktp);

        rd_kafka_toppar_op_version_bump(rktp, version);

        if (!pause && (rktp->rktp_flags & flag) != flag) {
                rd_kafka_dbg(rk, TOPIC, "RESUME",
                             "Not resuming %s [%" PRId32
                             "]: "
                             "partition is not paused by %s",
                             rktp->rktp_rkt->rkt_topic->str,
                             rktp->rktp_partition,
                             (flag & RD_KAFKA_TOPPAR_F_APP_PAUSE ? "application"
                                                                 : "library"));
                rd_kafka_toppar_unlock(rktp);
                return;
        }

        if (pause) {
                /* Pause partition by setting either
                 * RD_KAFKA_TOPPAR_F_APP_PAUSE or
                 * RD_KAFKA_TOPPAR_F_LIB_PAUSE */
                rktp->rktp_flags |= flag;

                if (rk->rk_type == RD_KAFKA_CONSUMER) {
                        /* Save offset of last consumed message+1 as the
                         * next message to fetch on resume. */
                        if (rktp->rktp_app_pos.offset !=
                            RD_KAFKA_OFFSET_INVALID)
                                rd_kafka_toppar_set_next_fetch_position(
                                    rktp, rktp->rktp_app_pos);

                        rd_kafka_dbg(
                            rk, TOPIC, pause ? "PAUSE" : "RESUME",
                            "%s %s [%" PRId32 "]: at %s (state %s, v%d)",
                            pause ? "Pause" : "Resume",
                            rktp->rktp_rkt->rkt_topic->str,
                            rktp->rktp_partition,
                            rd_kafka_fetch_pos2str(rktp->rktp_next_fetch_start),
                            rd_kafka_fetch_states[rktp->rktp_fetch_state],
                            version);
                } else {
                        rd_kafka_dbg(
                            rk, TOPIC, pause ? "PAUSE" : "RESUME",
                            "%s %s [%" PRId32 "] (state %s, v%d)",
                            pause ? "Pause" : "Resume",
                            rktp->rktp_rkt->rkt_topic->str,
                            rktp->rktp_partition,
                            rd_kafka_fetch_states[rktp->rktp_fetch_state],
                            version);
                }

        } else {
                /* Unset the RD_KAFKA_TOPPAR_F_APP_PAUSE or
                 * RD_KAFKA_TOPPAR_F_LIB_PAUSE flag */
                rktp->rktp_flags &= ~flag;

                if (rk->rk_type == RD_KAFKA_CONSUMER) {
                        rd_kafka_dbg(
                            rk, TOPIC, pause ? "PAUSE" : "RESUME",
                            "%s %s [%" PRId32 "]: at %s (state %s, v%d)",
                            rktp->rktp_fetch_state ==
                                    RD_KAFKA_TOPPAR_FETCH_ACTIVE
                                ? "Resuming"
                                : "Not resuming stopped",
                            rktp->rktp_rkt->rkt_topic->str,
                            rktp->rktp_partition,
                            rd_kafka_fetch_pos2str(rktp->rktp_next_fetch_start),
                            rd_kafka_fetch_states[rktp->rktp_fetch_state],
                            version);

                        /* If the resuming offset is logical we
                         * need to trigger a seek (that performs the
                         * logical->absolute lookup logic) to get
                         * things going.
                         * Typical case is when a partition is paused
                         * before anything has been consumed by app
                         * yet thus having rktp_app_offset=INVALID. */
                        if (!RD_KAFKA_TOPPAR_IS_PAUSED(rktp) &&
                            (rktp->rktp_fetch_state ==
                                 RD_KAFKA_TOPPAR_FETCH_ACTIVE ||
                             rktp->rktp_fetch_state ==
                                 RD_KAFKA_TOPPAR_FETCH_OFFSET_WAIT) &&
                            rktp->rktp_next_fetch_start.offset ==
                                RD_KAFKA_OFFSET_INVALID)
                                rd_kafka_toppar_next_offset_handle(
                                    rktp, rktp->rktp_next_fetch_start);

                } else
                        rd_kafka_dbg(
                            rk, TOPIC, pause ? "PAUSE" : "RESUME",
                            "%s %s [%" PRId32 "] (state %s, v%d)",
                            pause ? "Pause" : "Resume",
                            rktp->rktp_rkt->rkt_topic->str,
                            rktp->rktp_partition,
                            rd_kafka_fetch_states[rktp->rktp_fetch_state],
                            version);
        }
        rd_kafka_toppar_unlock(rktp);

        if (pause && rk->rk_type == RD_KAFKA_CONSUMER) {
                /* Flush partition's fetch queue */
                rd_kafka_q_purge_toppar_version(rktp->rktp_fetchq, rktp,
                                                rko_orig->rko_version);
        }
}



/**
 * @brief Serve a toppar in a consumer broker thread.
 *        This is considered the fast path and should be minimal,
 *        mostly focusing on fetch related mechanisms.
 *
 * @returns the partition's Fetch backoff timestamp, or 0 if no backoff.
 *
 * @locality broker thread
 * @locks none
 */
rd_ts_t rd_kafka_broker_consumer_toppar_serve(rd_kafka_broker_t *rkb,
                                              rd_kafka_toppar_t *rktp) {
        return rd_kafka_toppar_fetch_decide(rktp, rkb, 0);
}



/**
 * @brief Serve a toppar op
 *
 * @param rktp may be NULL for certain ops (OP_RECV_BUF)
 *
 * Will send an empty reply op if the request rko has a replyq set,
 * providing synchronous operation.
 *
 * @locality toppar handler thread
 */
static rd_kafka_op_res_t rd_kafka_toppar_op_serve(rd_kafka_t *rk,
                                                  rd_kafka_q_t *rkq,
                                                  rd_kafka_op_t *rko,
                                                  rd_kafka_q_cb_type_t cb_type,
                                                  void *opaque) {
        rd_kafka_toppar_t *rktp = NULL;
        int outdated            = 0;

        if (rko->rko_rktp)
                rktp = rko->rko_rktp;

        if (rktp) {
                outdated =
                    rd_kafka_op_version_outdated(rko, rktp->rktp_op_version);

                rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "OP",
                             "%.*s [%" PRId32
                             "] received %sop %s "
                             "(v%" PRId32 ") in fetch-state %s (opv%d)",
                             RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                             rktp->rktp_partition, outdated ? "outdated " : "",
                             rd_kafka_op2str(rko->rko_type), rko->rko_version,
                             rd_kafka_fetch_states[rktp->rktp_fetch_state],
                             rktp->rktp_op_version);

                if (outdated) {
#if ENABLE_DEVEL
                        rd_kafka_op_print(stdout, "PART_OUTDATED", rko);
#endif
                        rd_kafka_op_reply(rko, RD_KAFKA_RESP_ERR__OUTDATED);
                        return RD_KAFKA_OP_RES_HANDLED;
                }
        }

        switch ((int)rko->rko_type) {
        case RD_KAFKA_OP_FETCH_START:
                rd_kafka_toppar_fetch_start(rktp, rko->rko_u.fetch_start.pos,
                                            rko);
                break;

        case RD_KAFKA_OP_FETCH_STOP:
                rd_kafka_toppar_fetch_stop(rktp, rko);
                break;

        case RD_KAFKA_OP_SEEK:
                rd_kafka_toppar_seek(rktp, rko->rko_u.fetch_start.pos, rko);
                break;

        case RD_KAFKA_OP_PAUSE:
                rd_kafka_toppar_pause_resume(rktp, rko);
                break;

        case RD_KAFKA_OP_OFFSET_COMMIT | RD_KAFKA_OP_REPLY:
                rd_kafka_assert(NULL, rko->rko_u.offset_commit.cb);
                rko->rko_u.offset_commit.cb(rk, rko->rko_err,
                                            rko->rko_u.offset_commit.partitions,
                                            rko->rko_u.offset_commit.opaque);
                break;

        case RD_KAFKA_OP_OFFSET_FETCH | RD_KAFKA_OP_REPLY: {
                /* OffsetFetch reply */
                rd_kafka_topic_partition_list_t *offsets =
                    rko->rko_u.offset_fetch.partitions;
                rd_kafka_fetch_pos_t pos = {RD_KAFKA_OFFSET_INVALID, -1};

                rktp = rd_kafka_topic_partition_get_toppar(
                    rk, &offsets->elems[0], rd_true /*create-on-miss*/);

                if (!rko->rko_err) {
                        /* Request succeeded but per-partition might have failed
                         */
                        rko->rko_err = offsets->elems[0].err;
                        pos          = rd_kafka_topic_partition_get_fetch_pos(
                            &offsets->elems[0]);
                }

                rd_kafka_topic_partition_list_destroy(offsets);
                rko->rko_u.offset_fetch.partitions = NULL;

                rd_kafka_timer_stop(&rktp->rktp_rkt->rkt_rk->rk_timers,
                                    &rktp->rktp_offset_query_tmr, 1 /*lock*/);

                rd_kafka_toppar_lock(rktp);

                if (rko->rko_err) {
                        rd_kafka_dbg(
                            rktp->rktp_rkt->rkt_rk, TOPIC, "OFFSET",
                            "Failed to fetch offset for "
                            "%.*s [%" PRId32 "]: %s",
                            RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                            rktp->rktp_partition,
                            rd_kafka_err2str(rko->rko_err));

                        /* Keep on querying until we succeed. */
                        rd_kafka_toppar_offset_retry(rktp, 500,
                                                     "failed to fetch offsets");
                        rd_kafka_toppar_unlock(rktp);


                        /* Propagate error to application */
                        if (rko->rko_err != RD_KAFKA_RESP_ERR__WAIT_COORD &&
                            rko->rko_err !=
                                RD_KAFKA_RESP_ERR_UNSTABLE_OFFSET_COMMIT)
                                rd_kafka_consumer_err(
                                    rktp->rktp_fetchq, RD_KAFKA_NODEID_UA,
                                    rko->rko_err, 0, NULL, rktp,
                                    RD_KAFKA_OFFSET_INVALID,
                                    "Failed to fetch "
                                    "offsets from brokers: %s",
                                    rd_kafka_err2str(rko->rko_err));

                        /* Refcount from get_toppar() */
                        rd_kafka_toppar_destroy(rktp);

                        break;
                }

                rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "OFFSET",
                             "%.*s [%" PRId32 "]: OffsetFetch returned %s",
                             RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                             rktp->rktp_partition, rd_kafka_fetch_pos2str(pos));

                if (pos.offset > 0)
                        rktp->rktp_committed_pos = pos;

                if (pos.offset >= 0)
                        rd_kafka_toppar_next_offset_handle(rktp, pos);
                else
                        rd_kafka_offset_reset(rktp, RD_KAFKA_NODEID_UA, pos,
                                              RD_KAFKA_RESP_ERR__NO_OFFSET,
                                              "no previously committed offset "
                                              "available");
                rd_kafka_toppar_unlock(rktp);

                /* Refcount from get_toppar() */
                rd_kafka_toppar_destroy(rktp);
        } break;

        default:
                rd_kafka_assert(NULL, !*"unknown type");
                break;
        }

        rd_kafka_op_reply(rko, RD_KAFKA_RESP_ERR_NO_ERROR);

        return RD_KAFKA_OP_RES_HANDLED;
}



/**
 * Send command op to toppar (handled by toppar's thread).
 *
 * Locality: any thread
 */
static void rd_kafka_toppar_op0(rd_kafka_toppar_t *rktp,
                                rd_kafka_op_t *rko,
                                rd_kafka_replyq_t replyq) {
        rko->rko_rktp   = rd_kafka_toppar_keep(rktp);
        rko->rko_replyq = replyq;

        rd_kafka_q_enq(rktp->rktp_ops, rko);
}


/**
 * Send command op to toppar (handled by toppar's thread).
 *
 * Locality: any thread
 */
static void rd_kafka_toppar_op(rd_kafka_toppar_t *rktp,
                               rd_kafka_op_type_t type,
                               int32_t version,
                               rd_kafka_fetch_pos_t pos,
                               rd_kafka_cgrp_t *rkcg,
                               rd_kafka_replyq_t replyq) {
        rd_kafka_op_t *rko;

        rko              = rd_kafka_op_new(type);
        rko->rko_version = version;
        if (type == RD_KAFKA_OP_FETCH_START || type == RD_KAFKA_OP_SEEK) {
                if (rkcg)
                        rko->rko_u.fetch_start.rkcg = rkcg;
                rko->rko_u.fetch_start.pos = pos;
        }

        rd_kafka_toppar_op0(rktp, rko, replyq);
}



/**
 * Start consuming partition (async operation).
 *  'offset' is the initial offset
 *  'fwdq' is an optional queue to forward messages to, if this is NULL
 *  then messages will be enqueued on rktp_fetchq.
 *  'replyq' is an optional queue for handling the consume_start ack.
 *
 * This is the thread-safe interface that can be called from any thread.
 */
rd_kafka_resp_err_t rd_kafka_toppar_op_fetch_start(rd_kafka_toppar_t *rktp,
                                                   rd_kafka_fetch_pos_t pos,
                                                   rd_kafka_q_t *fwdq,
                                                   rd_kafka_replyq_t replyq) {
        int32_t version;

        rd_kafka_q_lock(rktp->rktp_fetchq);
        if (fwdq && !(rktp->rktp_fetchq->rkq_flags & RD_KAFKA_Q_F_FWD_APP))
                rd_kafka_q_fwd_set0(rktp->rktp_fetchq, fwdq, 0, /* no do_lock */
                                    0 /* no fwd_app */);
        rd_kafka_q_unlock(rktp->rktp_fetchq);

        /* Bump version barrier. */
        version = rd_kafka_toppar_version_new_barrier(rktp);

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "CONSUMER",
                     "Start consuming %.*s [%" PRId32 "] at %s (v%" PRId32 ")",
                     RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                     rktp->rktp_partition, rd_kafka_fetch_pos2str(pos),
                     version);

        rd_kafka_toppar_op(rktp, RD_KAFKA_OP_FETCH_START, version, pos,
                           rktp->rktp_rkt->rkt_rk->rk_cgrp, replyq);

        return RD_KAFKA_RESP_ERR_NO_ERROR;
}


/**
 * Stop consuming partition (async operatoin)
 * This is thread-safe interface that can be called from any thread.
 *
 * Locality: any thread
 */
rd_kafka_resp_err_t rd_kafka_toppar_op_fetch_stop(rd_kafka_toppar_t *rktp,
                                                  rd_kafka_replyq_t replyq) {
        int32_t version;

        /* Bump version barrier. */
        version = rd_kafka_toppar_version_new_barrier(rktp);

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "CONSUMER",
                     "Stop consuming %.*s [%" PRId32 "] (v%" PRId32 ")",
                     RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                     rktp->rktp_partition, version);

        rd_kafka_toppar_op(rktp, RD_KAFKA_OP_FETCH_STOP, version,
                           RD_KAFKA_FETCH_POS(-1, -1), NULL, replyq);

        return RD_KAFKA_RESP_ERR_NO_ERROR;
}


/**
 * @brief Set/Seek offset of a consumed partition (async operation).
 *
 * @param offset is the target offset.
 * @param leader_epoch is the partition leader epoch, or -1.
 * @param replyq is an optional queue for handling the ack.
 *
 * This is the thread-safe interface that can be called from any thread.
 */
rd_kafka_resp_err_t rd_kafka_toppar_op_seek(rd_kafka_toppar_t *rktp,
                                            rd_kafka_fetch_pos_t pos,
                                            rd_kafka_replyq_t replyq) {
        int32_t version;

        /* Bump version barrier. */
        version = rd_kafka_toppar_version_new_barrier(rktp);

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, "CONSUMER",
                     "Seek %.*s [%" PRId32 "] to %s (v%" PRId32 ")",
                     RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                     rktp->rktp_partition, rd_kafka_fetch_pos2str(pos),
                     version);

        rd_kafka_toppar_op(rktp, RD_KAFKA_OP_SEEK, version, pos, NULL, replyq);

        return RD_KAFKA_RESP_ERR_NO_ERROR;
}


/**
 * @brief Pause/resume partition (async operation).
 *
 * @param flag is either RD_KAFKA_TOPPAR_F_APP_PAUSE or .._F_LIB_PAUSE
 *             depending on if the app paused or librdkafka.
 * @param pause is 1 for pausing or 0 for resuming.
 *
 * @locality any
 */
rd_kafka_resp_err_t rd_kafka_toppar_op_pause_resume(rd_kafka_toppar_t *rktp,
                                                    int pause,
                                                    int flag,
                                                    rd_kafka_replyq_t replyq) {
        int32_t version;
        rd_kafka_op_t *rko;

        /* Bump version barrier. */
        version = rd_kafka_toppar_version_new_barrier(rktp);

        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC, pause ? "PAUSE" : "RESUME",
                     "%s %.*s [%" PRId32 "] (v%" PRId32 ")",
                     pause ? "Pause" : "Resume",
                     RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                     rktp->rktp_partition, version);

        rko                    = rd_kafka_op_new(RD_KAFKA_OP_PAUSE);
        rko->rko_version       = version;
        rko->rko_u.pause.pause = pause;
        rko->rko_u.pause.flag  = flag;

        rd_kafka_toppar_op0(rktp, rko, replyq);

        return RD_KAFKA_RESP_ERR_NO_ERROR;
}


/**
 * @brief Pause a toppar (asynchronous).
 *
 * @param flag is either RD_KAFKA_TOPPAR_F_APP_PAUSE or .._F_LIB_PAUSE
 *             depending on if the app paused or librdkafka.
 *
 * @locality any
 * @locks none needed
 */
void rd_kafka_toppar_pause(rd_kafka_toppar_t *rktp, int flag) {
        rd_kafka_toppar_op_pause_resume(rktp, 1 /*pause*/, flag,
                                        RD_KAFKA_NO_REPLYQ);
}

/**
 * @brief Resume a toppar (asynchronous).
 *
 * @param flag is either RD_KAFKA_TOPPAR_F_APP_PAUSE or .._F_LIB_PAUSE
 *             depending on if the app paused or librdkafka.
 *
 * @locality any
 * @locks none needed
 */
void rd_kafka_toppar_resume(rd_kafka_toppar_t *rktp, int flag) {
        rd_kafka_toppar_op_pause_resume(rktp, 1 /*pause*/, flag,
                                        RD_KAFKA_NO_REPLYQ);
}



/**
 * @brief Pause or resume a list of partitions.
 *
 * @param flag is either RD_KAFKA_TOPPAR_F_APP_PAUSE or .._F_LIB_PAUSE
 *             depending on if the app paused or librdkafka.
 * @param pause true for pausing, false for resuming.
 * @param async RD_SYNC to wait for background thread to handle op,
 *              RD_ASYNC for asynchronous operation.
 *
 * @locality any
 *
 * @remark This is an asynchronous call, the actual pause/resume is performed
 *         by toppar_pause() in the toppar's handler thread.
 */
rd_kafka_resp_err_t
rd_kafka_toppars_pause_resume(rd_kafka_t *rk,
                              rd_bool_t pause,
                              rd_async_t async,
                              int flag,
                              rd_kafka_topic_partition_list_t *partitions) {
        int i;
        int waitcnt        = 0;
        rd_kafka_q_t *tmpq = NULL;

        if (!async)
                tmpq = rd_kafka_q_new(rk);

        rd_kafka_dbg(
            rk, TOPIC, pause ? "PAUSE" : "RESUME", "%s %s %d partition(s)",
            flag & RD_KAFKA_TOPPAR_F_APP_PAUSE ? "Application" : "Library",
            pause ? "pausing" : "resuming", partitions->cnt);

        for (i = 0; i < partitions->cnt; i++) {
                rd_kafka_topic_partition_t *rktpar = &partitions->elems[i];
                rd_kafka_toppar_t *rktp;

                rktp =
                    rd_kafka_topic_partition_get_toppar(rk, rktpar, rd_false);
                if (!rktp) {
                        rd_kafka_dbg(rk, TOPIC, pause ? "PAUSE" : "RESUME",
                                     "%s %s [%" PRId32
                                     "]: skipped: "
                                     "unknown partition",
                                     pause ? "Pause" : "Resume", rktpar->topic,
                                     rktpar->partition);

                        rktpar->err = RD_KAFKA_RESP_ERR__UNKNOWN_PARTITION;
                        continue;
                }

                rd_kafka_toppar_op_pause_resume(rktp, pause, flag,
                                                RD_KAFKA_REPLYQ(tmpq, 0));

                if (!async)
                        waitcnt++;

                rd_kafka_toppar_destroy(rktp);

                rktpar->err = RD_KAFKA_RESP_ERR_NO_ERROR;
        }

        if (!async) {
                while (waitcnt-- > 0)
                        rd_kafka_q_wait_result(tmpq, RD_POLL_INFINITE);

                rd_kafka_q_destroy_owner(tmpq);
        }

        return RD_KAFKA_RESP_ERR_NO_ERROR;
}



/**
 * Propagate error for toppar
 */
void rd_kafka_toppar_enq_error(rd_kafka_toppar_t *rktp,
                               rd_kafka_resp_err_t err,
                               const char *reason) {
        rd_kafka_op_t *rko;
        char buf[512];

        rko           = rd_kafka_op_new(RD_KAFKA_OP_ERR);
        rko->rko_err  = err;
        rko->rko_rktp = rd_kafka_toppar_keep(rktp);

        rd_snprintf(buf, sizeof(buf), "%.*s [%" PRId32 "]: %s (%s)",
                    RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                    rktp->rktp_partition, reason, rd_kafka_err2str(err));

        rko->rko_u.err.errstr = rd_strdup(buf);

        rd_kafka_q_enq(rktp->rktp_fetchq, rko);
}



/**
 * Returns the currently delegated broker for this toppar.
 * If \p proper_broker is set NULL will be returned if current handler
 * is not a proper broker (INTERNAL broker).
 *
 * The returned broker has an increased refcount.
 *
 * Locks: none
 */
rd_kafka_broker_t *rd_kafka_toppar_broker(rd_kafka_toppar_t *rktp,
                                          int proper_broker) {
        rd_kafka_broker_t *rkb;
        rd_kafka_toppar_lock(rktp);
        rkb = rktp->rktp_broker;
        if (rkb) {
                if (proper_broker && rkb->rkb_source == RD_KAFKA_INTERNAL)
                        rkb = NULL;
                else
                        rd_kafka_broker_keep(rkb);
        }
        rd_kafka_toppar_unlock(rktp);

        return rkb;
}


/**
 * @brief Take action when partition broker becomes unavailable.
 *        This should be called when requests fail with
 *        NOT_LEADER_FOR.. or similar error codes, e.g. ProduceRequest.
 *
 * @locks none
 * @locality any
 */
void rd_kafka_toppar_leader_unavailable(rd_kafka_toppar_t *rktp,
                                        const char *reason,
                                        rd_kafka_resp_err_t err) {
        rd_kafka_topic_t *rkt = rktp->rktp_rkt;

        rd_kafka_dbg(rkt->rkt_rk, TOPIC, "BROKERUA",
                     "%s [%" PRId32 "]: broker unavailable: %s: %s",
                     rkt->rkt_topic->str, rktp->rktp_partition, reason,
                     rd_kafka_err2str(err));

        rd_kafka_topic_wrlock(rkt);
        rkt->rkt_flags |= RD_KAFKA_TOPIC_F_LEADER_UNAVAIL;
        rd_kafka_topic_wrunlock(rkt);

        rd_kafka_topic_fast_leader_query(rkt->rkt_rk);
}


const char *
rd_kafka_topic_partition_topic(const rd_kafka_topic_partition_t *rktpar) {
        const rd_kafka_toppar_t *rktp = (const rd_kafka_toppar_t *)rktpar;
        return rktp->rktp_rkt->rkt_topic->str;
}

int32_t
rd_kafka_topic_partition_partition(const rd_kafka_topic_partition_t *rktpar) {
        const rd_kafka_toppar_t *rktp = (const rd_kafka_toppar_t *)rktpar;
        return rktp->rktp_partition;
}

void rd_kafka_topic_partition_get(const rd_kafka_topic_partition_t *rktpar,
                                  const char **name,
                                  int32_t *partition) {
        const rd_kafka_toppar_t *rktp = (const rd_kafka_toppar_t *)rktpar;
        *name                         = rktp->rktp_rkt->rkt_topic->str;
        *partition                    = rktp->rktp_partition;
}



/**
 *
 * rd_kafka_topic_partition_t lists
 * Fixed-size non-growable list of partitions for propagation to application.
 *
 */


static void
rd_kafka_topic_partition_list_grow(rd_kafka_topic_partition_list_t *rktparlist,
                                   int add_size) {
        if (add_size < rktparlist->size)
                add_size = RD_MAX(rktparlist->size, 32);

        rktparlist->size += add_size;
        rktparlist->elems = rd_realloc(
            rktparlist->elems, sizeof(*rktparlist->elems) * rktparlist->size);
}


/**
 * @brief Initialize a list for fitting \p size partitions.
 */
void rd_kafka_topic_partition_list_init(
    rd_kafka_topic_partition_list_t *rktparlist,
    int size) {
        memset(rktparlist, 0, sizeof(*rktparlist));

        if (size > 0)
                rd_kafka_topic_partition_list_grow(rktparlist, size);
}


/**
 * Create a list for fitting 'size' topic_partitions (rktp).
 */
rd_kafka_topic_partition_list_t *rd_kafka_topic_partition_list_new(int size) {
        rd_kafka_topic_partition_list_t *rktparlist;

        rktparlist = rd_calloc(1, sizeof(*rktparlist));

        if (size > 0)
                rd_kafka_topic_partition_list_grow(rktparlist, size);

        return rktparlist;
}



rd_kafka_topic_partition_t *rd_kafka_topic_partition_new(const char *topic,
                                                         int32_t partition) {
        rd_kafka_topic_partition_t *rktpar = rd_calloc(1, sizeof(*rktpar));

        rktpar->topic     = rd_strdup(topic);
        rktpar->partition = partition;

        return rktpar;
}

/**
 * @brief Update \p dst with info from \p src.
 */
static void
rd_kafka_topic_partition_update(rd_kafka_topic_partition_t *dst,
                                const rd_kafka_topic_partition_t *src) {
        const rd_kafka_topic_partition_private_t *srcpriv;
        rd_kafka_topic_partition_private_t *dstpriv;

        rd_dassert(!strcmp(dst->topic, src->topic));
        rd_dassert(dst->partition == src->partition);
        rd_dassert(dst != src);

        dst->offset = src->offset;
        dst->opaque = src->opaque;
        dst->err    = src->err;

        if (src->metadata_size > 0) {
                dst->metadata      = rd_malloc(src->metadata_size);
                dst->metadata_size = src->metadata_size;
                ;
                memcpy(dst->metadata, src->metadata, dst->metadata_size);
        }

        if ((srcpriv = src->_private)) {
                dstpriv = rd_kafka_topic_partition_get_private(dst);
                if (srcpriv->rktp && !dstpriv->rktp)
                        dstpriv->rktp = rd_kafka_toppar_keep(srcpriv->rktp);

                rd_assert(dstpriv->rktp == srcpriv->rktp);

                dstpriv->leader_epoch = srcpriv->leader_epoch;

        } else if ((dstpriv = dst->_private)) {
                /* No private object in source, reset the leader epoch. */
                dstpriv->leader_epoch = -1;
        }
}


rd_kafka_topic_partition_t *
rd_kafka_topic_partition_copy(const rd_kafka_topic_partition_t *src) {
        rd_kafka_topic_partition_t *dst =
            rd_kafka_topic_partition_new(src->topic, src->partition);

        rd_kafka_topic_partition_update(dst, src);

        return dst;
}


/** Same as above but with generic void* signature */
void *rd_kafka_topic_partition_copy_void(const void *src) {
        return rd_kafka_topic_partition_copy(src);
}


rd_kafka_topic_partition_t *
rd_kafka_topic_partition_new_from_rktp(rd_kafka_toppar_t *rktp) {
        rd_kafka_topic_partition_t *rktpar = rd_calloc(1, sizeof(*rktpar));

        rktpar->topic     = RD_KAFKAP_STR_DUP(rktp->rktp_rkt->rkt_topic);
        rktpar->partition = rktp->rktp_partition;

        return rktpar;
}

/**
 * @brief Destroy a partition private glue object.
 */
static void rd_kafka_topic_partition_private_destroy(
    rd_kafka_topic_partition_private_t *parpriv) {
        if (parpriv->rktp)
                rd_kafka_toppar_destroy(parpriv->rktp);
        rd_free(parpriv);
}

static void
rd_kafka_topic_partition_destroy0(rd_kafka_topic_partition_t *rktpar,
                                  int do_free) {
        if (rktpar->topic)
                rd_free(rktpar->topic);
        if (rktpar->metadata)
                rd_free(rktpar->metadata);
        if (rktpar->_private)
                rd_kafka_topic_partition_private_destroy(
                    (rd_kafka_topic_partition_private_t *)rktpar->_private);

        if (do_free)
                rd_free(rktpar);
}


int32_t rd_kafka_topic_partition_get_leader_epoch(
    const rd_kafka_topic_partition_t *rktpar) {
        const rd_kafka_topic_partition_private_t *parpriv;

        if (!(parpriv = rktpar->_private))
                return -1;

        return parpriv->leader_epoch;
}

void rd_kafka_topic_partition_set_leader_epoch(
    rd_kafka_topic_partition_t *rktpar,
    int32_t leader_epoch) {
        rd_kafka_topic_partition_private_t *parpriv;

        /* Avoid allocating private_t if clearing the epoch */
        if (leader_epoch == -1 && !rktpar->_private)
                return;

        parpriv = rd_kafka_topic_partition_get_private(rktpar);

        parpriv->leader_epoch = leader_epoch;
}

int32_t rd_kafka_topic_partition_get_current_leader_epoch(
    const rd_kafka_topic_partition_t *rktpar) {
        const rd_kafka_topic_partition_private_t *parpriv;

        if (!(parpriv = rktpar->_private))
                return -1;

        return parpriv->current_leader_epoch;
}

void rd_kafka_topic_partition_set_current_leader_epoch(
    rd_kafka_topic_partition_t *rktpar,
    int32_t current_leader_epoch) {
        rd_kafka_topic_partition_private_t *parpriv;

        /* Avoid allocating private_t if clearing the epoch */
        if (current_leader_epoch == -1 && !rktpar->_private)
                return;

        parpriv = rd_kafka_topic_partition_get_private(rktpar);

        parpriv->current_leader_epoch = current_leader_epoch;
}

/**
 * @brief Set offset and leader epoch from a fetchpos.
 */
void rd_kafka_topic_partition_set_from_fetch_pos(
    rd_kafka_topic_partition_t *rktpar,
    const rd_kafka_fetch_pos_t fetchpos) {
        rktpar->offset = fetchpos.offset;
        rd_kafka_topic_partition_set_leader_epoch(rktpar,
                                                  fetchpos.leader_epoch);
}

/**
 * @brief Destroy all partitions in list.
 *
 * @remark The allocated size of the list will not shrink.
 */
void rd_kafka_topic_partition_list_clear(
    rd_kafka_topic_partition_list_t *rktparlist) {
        int i;

        for (i = 0; i < rktparlist->cnt; i++)
                rd_kafka_topic_partition_destroy0(&rktparlist->elems[i], 0);

        rktparlist->cnt = 0;
}


void rd_kafka_topic_partition_destroy_free(void *ptr) {
        rd_kafka_topic_partition_destroy0(ptr, rd_true /*do_free*/);
}

void rd_kafka_topic_partition_destroy(rd_kafka_topic_partition_t *rktpar) {
        rd_kafka_topic_partition_destroy0(rktpar, 1);
}


/**
 * Destroys a list previously created with .._list_new() and drops
 * any references to contained toppars.
 */
void rd_kafka_topic_partition_list_destroy(
    rd_kafka_topic_partition_list_t *rktparlist) {
        int i;

        for (i = 0; i < rktparlist->cnt; i++)
                rd_kafka_topic_partition_destroy0(&rktparlist->elems[i], 0);

        if (rktparlist->elems)
                rd_free(rktparlist->elems);

        rd_free(rktparlist);
}


/**
 * @brief Wrapper for rd_kafka_topic_partition_list_destroy() that
 *        matches the standard free(void *) signature, for callback use.
 */
void rd_kafka_topic_partition_list_destroy_free(void *ptr) {
        rd_kafka_topic_partition_list_destroy(
            (rd_kafka_topic_partition_list_t *)ptr);
}


/**
 * @brief Add a partition to an rktpar list.
 * The list must have enough room to fit it.
 *
 * @param rktp Optional partition object that will be stored on the
 * ._private object (with refcount increased).
 *
 * @returns a pointer to the added element.
 */
rd_kafka_topic_partition_t *rd_kafka_topic_partition_list_add0(
    const char *func,
    int line,
    rd_kafka_topic_partition_list_t *rktparlist,
    const char *topic,
    int32_t partition,
    rd_kafka_toppar_t *rktp,
    const rd_kafka_topic_partition_private_t *parpriv) {
        rd_kafka_topic_partition_t *rktpar;
        if (rktparlist->cnt == rktparlist->size)
                rd_kafka_topic_partition_list_grow(rktparlist, 1);
        rd_kafka_assert(NULL, rktparlist->cnt < rktparlist->size);

        rktpar = &rktparlist->elems[rktparlist->cnt++];
        memset(rktpar, 0, sizeof(*rktpar));
        rktpar->topic     = rd_strdup(topic);
        rktpar->partition = partition;
        rktpar->offset    = RD_KAFKA_OFFSET_INVALID;

        if (parpriv) {
                rd_kafka_topic_partition_private_t *parpriv_copy =
                    rd_kafka_topic_partition_get_private(rktpar);
                if (parpriv->rktp) {
                        parpriv_copy->rktp =
                            rd_kafka_toppar_keep_fl(func, line, parpriv->rktp);
                }
                parpriv_copy->leader_epoch         = parpriv->leader_epoch;
                parpriv_copy->current_leader_epoch = parpriv->leader_epoch;
        } else if (rktp) {
                rd_kafka_topic_partition_private_t *parpriv_copy =
                    rd_kafka_topic_partition_get_private(rktpar);
                parpriv_copy->rktp = rd_kafka_toppar_keep_fl(func, line, rktp);
        }

        return rktpar;
}


rd_kafka_topic_partition_t *
rd_kafka_topic_partition_list_add(rd_kafka_topic_partition_list_t *rktparlist,
                                  const char *topic,
                                  int32_t partition) {
        return rd_kafka_topic_partition_list_add0(
            __FUNCTION__, __LINE__, rktparlist, topic, partition, NULL, NULL);
}


/**
 * Adds a consecutive list of partitions to a list
 */
void rd_kafka_topic_partition_list_add_range(
    rd_kafka_topic_partition_list_t *rktparlist,
    const char *topic,
    int32_t start,
    int32_t stop) {

        for (; start <= stop; start++)
                rd_kafka_topic_partition_list_add(rktparlist, topic, start);
}


rd_kafka_topic_partition_t *rd_kafka_topic_partition_list_upsert(
    rd_kafka_topic_partition_list_t *rktparlist,
    const char *topic,
    int32_t partition) {
        rd_kafka_topic_partition_t *rktpar;

        if ((rktpar = rd_kafka_topic_partition_list_find(rktparlist, topic,
                                                         partition)))
                return rktpar;

        return rd_kafka_topic_partition_list_add(rktparlist, topic, partition);
}



/**
 * @brief Creates a copy of \p rktpar and adds it to \p rktparlist
 */
void rd_kafka_topic_partition_list_add_copy(
    rd_kafka_topic_partition_list_t *rktparlist,
    const rd_kafka_topic_partition_t *rktpar) {
        rd_kafka_topic_partition_t *dst;

        dst = rd_kafka_topic_partition_list_add0(
            __FUNCTION__, __LINE__, rktparlist, rktpar->topic,
            rktpar->partition, NULL, rktpar->_private);
        rd_kafka_topic_partition_update(dst, rktpar);
}



/**
 * Create and return a copy of list 'src'
 */
rd_kafka_topic_partition_list_t *
rd_kafka_topic_partition_list_copy(const rd_kafka_topic_partition_list_t *src) {
        rd_kafka_topic_partition_list_t *dst;
        int i;

        dst = rd_kafka_topic_partition_list_new(src->size);

        for (i = 0; i < src->cnt; i++)
                rd_kafka_topic_partition_list_add_copy(dst, &src->elems[i]);
        return dst;
}

/**
 * @brief Same as rd_kafka_topic_partition_list_copy() but suitable for
 *        rd_list_copy(). The \p opaque is ignored.
 */
void *rd_kafka_topic_partition_list_copy_opaque(const void *src, void *opaque) {
        return rd_kafka_topic_partition_list_copy(src);
}

/**
 * @brief Append copies of all elements in \p src to \p dst.
 *        No duplicate-checks are performed.
 */
void rd_kafka_topic_partition_list_add_list(
    rd_kafka_topic_partition_list_t *dst,
    const rd_kafka_topic_partition_list_t *src) {
        int i;

        if (src->cnt == 0)
                return;

        if (dst->size < dst->cnt + src->cnt)
                rd_kafka_topic_partition_list_grow(dst, src->cnt);

        for (i = 0; i < src->cnt; i++)
                rd_kafka_topic_partition_list_add_copy(dst, &src->elems[i]);
}


/**
 * @brief Compare two partition lists using partition comparator \p cmp.
 *
 * @warning This is an O(Na*Nb) operation.
 */
int rd_kafka_topic_partition_list_cmp(const void *_a,
                                      const void *_b,
                                      int (*cmp)(const void *, const void *)) {
        const rd_kafka_topic_partition_list_t *a = _a, *b = _b;
        int r;
        int i;

        r = a->cnt - b->cnt;
        if (r || a->cnt == 0)
                return r;

        /* Since the lists may not be sorted we need to scan all of B
         * for each element in A.
         * FIXME: If the list sizes are larger than X we could create a
         *        temporary hash map instead. */
        for (i = 0; i < a->cnt; i++) {
                int j;

                for (j = 0; j < b->cnt; j++) {
                        r = cmp(&a->elems[i], &b->elems[j]);
                        if (!r)
                                break;
                }

                if (j == b->cnt)
                        return 1;
        }

        return 0;
}


/**
 * @brief Ensures the \p rktpar has a toppar set in _private.
 *
 * @returns the toppar object (or possibly NULL if \p create_on_miss is true)
 *          WITHOUT refcnt increased.
 */
rd_kafka_toppar_t *
rd_kafka_topic_partition_ensure_toppar(rd_kafka_t *rk,
                                       rd_kafka_topic_partition_t *rktpar,
                                       rd_bool_t create_on_miss) {
        rd_kafka_topic_partition_private_t *parpriv;

        parpriv = rd_kafka_topic_partition_get_private(rktpar);

        if (!parpriv->rktp)
                parpriv->rktp = rd_kafka_toppar_get2(
                    rk, rktpar->topic, rktpar->partition,
                    0 /* not ua on miss */, create_on_miss);

        return parpriv->rktp;
}


int rd_kafka_topic_partition_cmp(const void *_a, const void *_b) {
        const rd_kafka_topic_partition_t *a = _a;
        const rd_kafka_topic_partition_t *b = _b;
        int r                               = strcmp(a->topic, b->topic);
        if (r)
                return r;
        else
                return RD_CMP(a->partition, b->partition);
}

/** @brief Compare only the topic */
int rd_kafka_topic_partition_cmp_topic(const void *_a, const void *_b) {
        const rd_kafka_topic_partition_t *a = _a;
        const rd_kafka_topic_partition_t *b = _b;
        return strcmp(a->topic, b->topic);
}

static int rd_kafka_topic_partition_cmp_opaque(const void *_a,
                                               const void *_b,
                                               void *opaque) {
        return rd_kafka_topic_partition_cmp(_a, _b);
}

/** @returns a hash of the topic and partition */
unsigned int rd_kafka_topic_partition_hash(const void *_a) {
        const rd_kafka_topic_partition_t *a = _a;
        int r                               = 31 * 17 + a->partition;
        return 31 * r + rd_string_hash(a->topic, -1);
}



/**
 * @brief Search 'rktparlist' for 'topic' and 'partition'.
 * @returns the elems[] index or -1 on miss.
 */
static int rd_kafka_topic_partition_list_find0(
    const rd_kafka_topic_partition_list_t *rktparlist,
    const char *topic,
    int32_t partition,
    int (*cmp)(const void *, const void *)) {
        rd_kafka_topic_partition_t skel;
        int i;

        skel.topic     = (char *)topic;
        skel.partition = partition;

        for (i = 0; i < rktparlist->cnt; i++) {
                if (!cmp(&skel, &rktparlist->elems[i]))
                        return i;
        }

        return -1;
}

rd_kafka_topic_partition_t *rd_kafka_topic_partition_list_find(
    const rd_kafka_topic_partition_list_t *rktparlist,
    const char *topic,
    int32_t partition) {
        int i = rd_kafka_topic_partition_list_find0(
            rktparlist, topic, partition, rd_kafka_topic_partition_cmp);
        if (i == -1)
                return NULL;
        else
                return &rktparlist->elems[i];
}


int rd_kafka_topic_partition_list_find_idx(
    const rd_kafka_topic_partition_list_t *rktparlist,
    const char *topic,
    int32_t partition) {
        return rd_kafka_topic_partition_list_find0(
            rktparlist, topic, partition, rd_kafka_topic_partition_cmp);
}


/**
 * @returns the first element that matches \p topic, regardless of partition.
 */
rd_kafka_topic_partition_t *rd_kafka_topic_partition_list_find_topic(
    const rd_kafka_topic_partition_list_t *rktparlist,
    const char *topic) {
        int i = rd_kafka_topic_partition_list_find0(
            rktparlist, topic, RD_KAFKA_PARTITION_UA,
            rd_kafka_topic_partition_cmp_topic);
        if (i == -1)
                return NULL;
        else
                return &rktparlist->elems[i];
}


int rd_kafka_topic_partition_list_del_by_idx(
    rd_kafka_topic_partition_list_t *rktparlist,
    int idx) {
        if (unlikely(idx < 0 || idx >= rktparlist->cnt))
                return 0;

        rd_kafka_topic_partition_destroy0(&rktparlist->elems[idx], 0);
        memmove(&rktparlist->elems[idx], &rktparlist->elems[idx + 1],
                (rktparlist->cnt - idx - 1) * sizeof(rktparlist->elems[idx]));
        rktparlist->cnt--;

        return 1;
}


int rd_kafka_topic_partition_list_del(
    rd_kafka_topic_partition_list_t *rktparlist,
    const char *topic,
    int32_t partition) {
        int i = rd_kafka_topic_partition_list_find0(
            rktparlist, topic, partition, rd_kafka_topic_partition_cmp);
        if (i == -1)
                return 0;

        return rd_kafka_topic_partition_list_del_by_idx(rktparlist, i);
}



/**
 * Returns true if 'topic' matches the 'rktpar', else false.
 * On match, if rktpar is a regex pattern then 'matched_by_regex' is set to 1.
 */
int rd_kafka_topic_partition_match(rd_kafka_t *rk,
                                   const rd_kafka_group_member_t *rkgm,
                                   const rd_kafka_topic_partition_t *rktpar,
                                   const char *topic,
                                   int *matched_by_regex) {
        int ret = 0;

        if (*rktpar->topic == '^') {
                char errstr[128];

                ret = rd_regex_match(rktpar->topic, topic, errstr,
                                     sizeof(errstr));
                if (ret == -1) {
                        rd_kafka_dbg(rk, CGRP, "SUBMATCH",
                                     "Invalid regex for member "
                                     "\"%.*s\" subscription \"%s\": %s",
                                     RD_KAFKAP_STR_PR(rkgm->rkgm_member_id),
                                     rktpar->topic, errstr);
                        return 0;
                }

                if (ret && matched_by_regex)
                        *matched_by_regex = 1;

        } else if (!strcmp(rktpar->topic, topic)) {

                if (matched_by_regex)
                        *matched_by_regex = 0;

                ret = 1;
        }

        return ret;
}



void rd_kafka_topic_partition_list_sort(
    rd_kafka_topic_partition_list_t *rktparlist,
    int (*cmp)(const void *, const void *, void *),
    void *opaque) {

        if (!cmp)
                cmp = rd_kafka_topic_partition_cmp_opaque;

        rd_qsort_r(rktparlist->elems, rktparlist->cnt,
                   sizeof(*rktparlist->elems), cmp, opaque);
}


void rd_kafka_topic_partition_list_sort_by_topic(
    rd_kafka_topic_partition_list_t *rktparlist) {
        rd_kafka_topic_partition_list_sort(
            rktparlist, rd_kafka_topic_partition_cmp_opaque, NULL);
}

rd_kafka_resp_err_t rd_kafka_topic_partition_list_set_offset(
    rd_kafka_topic_partition_list_t *rktparlist,
    const char *topic,
    int32_t partition,
    int64_t offset) {
        rd_kafka_topic_partition_t *rktpar;

        if (!(rktpar = rd_kafka_topic_partition_list_find(rktparlist, topic,
                                                          partition)))
                return RD_KAFKA_RESP_ERR__UNKNOWN_PARTITION;

        rktpar->offset = offset;

        return RD_KAFKA_RESP_ERR_NO_ERROR;
}


/**
 * @brief Reset all offsets to the provided value.
 */
void rd_kafka_topic_partition_list_reset_offsets(
    rd_kafka_topic_partition_list_t *rktparlist,
    int64_t offset) {

        int i;
        for (i = 0; i < rktparlist->cnt; i++)
                rktparlist->elems[i].offset = offset;
}


/**
 * Set offset values in partition list based on toppar's last stored offset.
 *
 *  from_rktp - true: set rktp's last stored offset, false: set def_value
 *  unless a concrete offset is set.
 *  is_commit: indicates that set offset is to be committed (for debug log)
 *
 * Returns the number of valid non-logical offsets (>=0).
 */
int rd_kafka_topic_partition_list_set_offsets(
    rd_kafka_t *rk,
    rd_kafka_topic_partition_list_t *rktparlist,
    int from_rktp,
    int64_t def_value,
    int is_commit) {
        int i;
        int valid_cnt = 0;

        for (i = 0; i < rktparlist->cnt; i++) {
                rd_kafka_topic_partition_t *rktpar = &rktparlist->elems[i];
                const char *verb                   = "setting";
                char preamble[128];

                *preamble = '\0'; /* Avoid warning */

                if (from_rktp) {
                        rd_kafka_toppar_t *rktp =
                            rd_kafka_topic_partition_ensure_toppar(rk, rktpar,
                                                                   rd_true);
                        rd_kafka_toppar_lock(rktp);

                        if (rk->rk_conf.debug &
                            (RD_KAFKA_DBG_CGRP | RD_KAFKA_DBG_TOPIC))
                                rd_snprintf(preamble, sizeof(preamble),
                                            "stored %s, committed %s: ",
                                            rd_kafka_fetch_pos2str(
                                                rktp->rktp_stored_pos),
                                            rd_kafka_fetch_pos2str(
                                                rktp->rktp_committed_pos));

                        if (rd_kafka_fetch_pos_cmp(&rktp->rktp_stored_pos,
                                                   &rktp->rktp_committed_pos) >
                            0) {
                                verb = "setting stored";
                                rd_kafka_topic_partition_set_from_fetch_pos(
                                    rktpar, rktp->rktp_stored_pos);
                        } else {
                                rktpar->offset = RD_KAFKA_OFFSET_INVALID;
                        }
                        rd_kafka_toppar_unlock(rktp);
                } else {
                        if (RD_KAFKA_OFFSET_IS_LOGICAL(rktpar->offset)) {
                                verb           = "setting default";
                                rktpar->offset = def_value;
                                rd_kafka_topic_partition_set_leader_epoch(
                                    rktpar, -1);
                        } else
                                verb = "keeping";
                }

                if (is_commit && rktpar->offset == RD_KAFKA_OFFSET_INVALID)
                        rd_kafka_dbg(rk, CGRP | RD_KAFKA_DBG_TOPIC, "OFFSET",
                                     "Topic %s [%" PRId32
                                     "]: "
                                     "%snot including in commit",
                                     rktpar->topic, rktpar->partition,
                                     preamble);
                else
                        rd_kafka_dbg(
                            rk, CGRP | RD_KAFKA_DBG_TOPIC, "OFFSET",
                            "Topic %s [%" PRId32
                            "]: "
                            "%s%s offset %s (leader epoch %" PRId32 ") %s",
                            rktpar->topic, rktpar->partition, preamble, verb,
                            rd_kafka_offset2str(rktpar->offset),
                            rd_kafka_topic_partition_get_leader_epoch(rktpar),
                            is_commit ? " for commit" : "");

                if (!RD_KAFKA_OFFSET_IS_LOGICAL(rktpar->offset))
                        valid_cnt++;
        }

        return valid_cnt;
}


/**
 * @returns the number of partitions with absolute (non-logical) offsets set.
 */
int rd_kafka_topic_partition_list_count_abs_offsets(
    const rd_kafka_topic_partition_list_t *rktparlist) {
        int i;
        int valid_cnt = 0;

        for (i = 0; i < rktparlist->cnt; i++)
                if (!RD_KAFKA_OFFSET_IS_LOGICAL(rktparlist->elems[i].offset))
                        valid_cnt++;

        return valid_cnt;
}


/**
 * @brief Update _private (toppar) field to point to valid rktp
 *        for each parition.
 *
 * @param create_on_miss Create partition (and topic_t object) if necessary.
 */
void rd_kafka_topic_partition_list_update_toppars(
    rd_kafka_t *rk,
    rd_kafka_topic_partition_list_t *rktparlist,
    rd_bool_t create_on_miss) {
        int i;
        for (i = 0; i < rktparlist->cnt; i++) {
                rd_kafka_topic_partition_t *rktpar = &rktparlist->elems[i];

                rd_kafka_topic_partition_ensure_toppar(rk, rktpar,
                                                       create_on_miss);
        }
}


/**
 * @brief Populate \p leaders with the leaders+partitions for the partitions in
 *        \p rktparlist. Duplicates are suppressed.
 *
 *        If no leader is found for a partition that element's \c .err will
 *        be set to RD_KAFKA_RESP_ERR_LEADER_NOT_AVAILABLE.
 *
 *        If the partition does not exist \c .err will be set to
 *        RD_KAFKA_RESP_ERR__UNKNOWN_PARTITION.
 *
 * @param rktparlist The partitions to look up leaders for, the .err field
 *                   will be set according to outcome, e.g., ERR_NO_ERROR,
 *                   ERR_UNKNOWN_TOPIC_OR_PART, etc.
 * @param leaders rd_list_t of allocated (struct rd_kafka_partition_leader *)
 * @param query_topics (optional) rd_list of strdupped (char *)
 * @param query_unknown Add unknown topics to \p query_topics.
 * @param eonce (optional) For triggering asynchronously on cache change
 *              in case not all leaders are known now.
 *
 * @remark This is based on the current topic_t and partition state
 *         which may lag behind the last metadata update due to internal
 *         threading and also the fact that no topic_t may have been created.
 *
 * @param leaders rd_list_t of type (struct rd_kafka_partition_leader *)
 *
 * @returns true if all partitions have leaders, else false.
 *
 * @sa rd_kafka_topic_partition_list_get_leaders_by_metadata
 *
 * @locks rd_kafka_*lock() MUST NOT be held
 */
static rd_bool_t rd_kafka_topic_partition_list_get_leaders(
    rd_kafka_t *rk,
    rd_kafka_topic_partition_list_t *rktparlist,
    rd_list_t *leaders,
    rd_list_t *query_topics,
    rd_bool_t query_unknown,
    rd_kafka_enq_once_t *eonce) {
        rd_bool_t complete;
        int cnt = 0;
        int i;

        if (eonce)
                rd_kafka_wrlock(rk);
        else
                rd_kafka_rdlock(rk);

        for (i = 0; i < rktparlist->cnt; i++) {
                rd_kafka_topic_partition_t *rktpar = &rktparlist->elems[i];
                rd_kafka_topic_partition_t *rktpar2;
                rd_kafka_broker_t *rkb = NULL;
                struct rd_kafka_partition_leader leader_skel;
                struct rd_kafka_partition_leader *leader;
                const rd_kafka_metadata_topic_t *mtopic;
                const rd_kafka_metadata_partition_t *mpart;
                rd_bool_t topic_wait_cache;

                rd_kafka_metadata_cache_topic_partition_get(
                    rk, &mtopic, &mpart, rktpar->topic, rktpar->partition,
                    0 /*negative entries too*/);

                topic_wait_cache =
                    !mtopic ||
                    RD_KAFKA_METADATA_CACHE_ERR_IS_TEMPORARY(mtopic->err);

                if (!topic_wait_cache && mtopic &&
                    mtopic->err != RD_KAFKA_RESP_ERR_NO_ERROR &&
                    mtopic->err != RD_KAFKA_RESP_ERR_LEADER_NOT_AVAILABLE) {
                        /* Topic permanently errored */
                        rktpar->err = mtopic->err;
                        continue;
                }

                if (mtopic && !mpart && mtopic->partition_cnt > 0) {
                        /* Topic exists but partition doesnt.
                         * This is a permanent error. */
                        rktpar->err = RD_KAFKA_RESP_ERR__UNKNOWN_PARTITION;
                        continue;
                }

                if (mpart &&
                    (mpart->leader == -1 ||
                     !(rkb = rd_kafka_broker_find_by_nodeid0(
                           rk, mpart->leader, -1 /*any state*/, rd_false)))) {
                        /* Partition has no (valid) leader.
                         * This is a permanent error. */
                        rktpar->err =
                            mtopic->err
                                ? mtopic->err
                                : RD_KAFKA_RESP_ERR_LEADER_NOT_AVAILABLE;
                        continue;
                }

                if (topic_wait_cache || !rkb) {
                        /* Topic unknown or no current leader for partition,
                         * add topic to query list. */
                        rktpar->err = RD_KAFKA_RESP_ERR__IN_PROGRESS;
                        if (query_topics &&
                            !rd_list_find(query_topics, rktpar->topic,
                                          (void *)strcmp))
                                rd_list_add(query_topics,
                                            rd_strdup(rktpar->topic));
                        continue;
                }

                /* Leader exists, add to leader list. */

                rktpar->err = RD_KAFKA_RESP_ERR_NO_ERROR;

                memset(&leader_skel, 0, sizeof(leader_skel));
                leader_skel.rkb = rkb;

                leader = rd_list_find(leaders, &leader_skel,
                                      rd_kafka_partition_leader_cmp);

                if (!leader) {
                        leader = rd_kafka_partition_leader_new(rkb);
                        rd_list_add(leaders, leader);
                }

                rktpar2 = rd_kafka_topic_partition_list_find(
                    leader->partitions, rktpar->topic, rktpar->partition);
                if (rktpar2) {
                        /* Already exists in partitions list, just update. */
                        rd_kafka_topic_partition_update(rktpar2, rktpar);
                } else {
                        /* Make a copy of rktpar and add to partitions list */
                        rd_kafka_topic_partition_list_add_copy(
                            leader->partitions, rktpar);
                }

                rktpar->err = RD_KAFKA_RESP_ERR_NO_ERROR;

                rd_kafka_broker_destroy(rkb); /* loose refcount */
                cnt++;
        }

        complete = cnt == rktparlist->cnt;

        if (!complete && eonce)
                /* Add eonce to cache observers */
                rd_kafka_metadata_cache_wait_state_change_async(rk, eonce);

        if (eonce)
                rd_kafka_wrunlock(rk);
        else
                rd_kafka_rdunlock(rk);

        return complete;
}


/**
 * @brief Timer timeout callback for query_leaders_async rko's eonce object.
 */
static void
rd_kafka_partition_leader_query_eonce_timeout_cb(rd_kafka_timers_t *rkts,
                                                 void *arg) {
        rd_kafka_enq_once_t *eonce = arg;
        rd_kafka_enq_once_trigger(eonce, RD_KAFKA_RESP_ERR__TIMED_OUT,
                                  "timeout timer");
}


/**
 * @brief Query timer callback for query_leaders_async rko's eonce object.
 */
static void
rd_kafka_partition_leader_query_eonce_timer_cb(rd_kafka_timers_t *rkts,
                                               void *arg) {
        rd_kafka_enq_once_t *eonce = arg;
        rd_kafka_enq_once_trigger(eonce, RD_KAFKA_RESP_ERR_NO_ERROR,
                                  "query timer");
}


/**
 * @brief Query metadata cache for partition leaders, or trigger metadata
 *        refresh if leaders not known.
 *
 * @locks_required none
 * @locality any
 */
static rd_kafka_op_res_t
rd_kafka_topic_partition_list_query_leaders_async_worker(rd_kafka_op_t *rko) {
        rd_kafka_t *rk = rko->rko_rk;
        rd_list_t query_topics, *leaders = NULL;
        rd_kafka_op_t *reply;

        RD_KAFKA_OP_TYPE_ASSERT(rko, RD_KAFKA_OP_LEADERS);

        if (rko->rko_err)
                goto reply; /* Timeout or ERR__DESTROY */

        /* Since we're iterating over get_leaders() until all partition leaders
         * are known we need to re-enable the eonce to be triggered again (which
         * is not necessary the first time we get here, but there
         * is no harm doing it then either). */
        rd_kafka_enq_once_reenable(rko->rko_u.leaders.eonce, rko,
                                   RD_KAFKA_REPLYQ(rk->rk_ops, 0));

        /* Look up the leaders in the metadata cache, if not all leaders
         * are known the eonce is registered for metadata cache changes
         * which will cause our function to be called
         * again on (any) metadata cache change.
         *
         * When we are called again we perform the cache lookup again and
         * hopefully get all leaders, otherwise defer a new async wait.
         * Repeat until success or timeout. */

        rd_list_init(&query_topics, 4 + rko->rko_u.leaders.partitions->cnt / 2,
                     rd_free);

        leaders = rd_list_new(1 + rko->rko_u.leaders.partitions->cnt / 2,
                              rd_kafka_partition_leader_destroy_free);

        if (rd_kafka_topic_partition_list_get_leaders(
                rk, rko->rko_u.leaders.partitions, leaders, &query_topics,
                /* Add unknown topics to query_topics only on the
                 * first query, after that we consider them permanently
                 * non-existent */
                rko->rko_u.leaders.query_cnt == 0, rko->rko_u.leaders.eonce)) {
                /* All leaders now known (or failed), reply to caller */
                rd_list_destroy(&query_topics);
                goto reply;
        }

        if (rd_list_empty(&query_topics)) {
                /* Not all leaders known but no topics left to query,
                 * reply to caller. */
                rd_list_destroy(&query_topics);
                goto reply;
        }

        /* Need to refresh topic metadata, but at most every interval. */
        if (!rd_kafka_timer_is_started(&rk->rk_timers,
                                       &rko->rko_u.leaders.query_tmr)) {

                rko->rko_u.leaders.query_cnt++;

                /* Add query interval timer. */
                rd_kafka_enq_once_add_source(rko->rko_u.leaders.eonce,
                                             "query timer");
                rd_kafka_timer_start_oneshot(
                    &rk->rk_timers, &rko->rko_u.leaders.query_tmr, rd_true,
                    3 * 1000 * 1000 /* 3s */,
                    rd_kafka_partition_leader_query_eonce_timer_cb,
                    rko->rko_u.leaders.eonce);

                /* Request metadata refresh */
                rd_kafka_metadata_refresh_topics(
                    rk, NULL, &query_topics, rd_true /*force*/,
                    rd_false /*!allow_auto_create*/, rd_false /*!cgrp_update*/,
                    "query partition leaders");
        }

        rd_list_destroy(leaders);
        rd_list_destroy(&query_topics);

        /* Wait for next eonce trigger */
        return RD_KAFKA_OP_RES_KEEP; /* rko is still used */

reply:
        /* Decommission worker state and reply to caller */

        if (rd_kafka_timer_stop(&rk->rk_timers, &rko->rko_u.leaders.query_tmr,
                                RD_DO_LOCK))
                rd_kafka_enq_once_del_source(rko->rko_u.leaders.eonce,
                                             "query timer");
        if (rd_kafka_timer_stop(&rk->rk_timers, &rko->rko_u.leaders.timeout_tmr,
                                RD_DO_LOCK))
                rd_kafka_enq_once_del_source(rko->rko_u.leaders.eonce,
                                             "timeout timer");

        if (rko->rko_u.leaders.eonce) {
                rd_kafka_enq_once_disable(rko->rko_u.leaders.eonce);
                rko->rko_u.leaders.eonce = NULL;
        }

        /* No leaders found, set a request-level error */
        if (leaders && rd_list_cnt(leaders) == 0) {
                if (!rko->rko_err)
                        rko->rko_err = RD_KAFKA_RESP_ERR__NOENT;
                rd_list_destroy(leaders);
                leaders = NULL;
        }

        /* Create and enqueue reply rko */
        if (rko->rko_u.leaders.replyq.q) {
                reply = rd_kafka_op_new_cb(rk, RD_KAFKA_OP_LEADERS,
                                           rko->rko_u.leaders.cb);
                rd_kafka_op_get_reply_version(reply, rko);
                reply->rko_err = rko->rko_err;
                reply->rko_u.leaders.partitions =
                    rko->rko_u.leaders.partitions; /* Transfer ownership for
                                                    * partition list that
                                                    * now contains
                                                    * per-partition errors*/
                rko->rko_u.leaders.partitions = NULL;
                reply->rko_u.leaders.leaders  = leaders; /* Possibly NULL */
                reply->rko_u.leaders.opaque   = rko->rko_u.leaders.opaque;

                rd_kafka_replyq_enq(&rko->rko_u.leaders.replyq, reply, 0);
        }

        return RD_KAFKA_OP_RES_HANDLED;
}


static rd_kafka_op_res_t
rd_kafka_topic_partition_list_query_leaders_async_worker_op_cb(
    rd_kafka_t *rk,
    rd_kafka_q_t *rkq,
    rd_kafka_op_t *rko) {
        return rd_kafka_topic_partition_list_query_leaders_async_worker(rko);
}

/**
 * @brief Async variant of rd_kafka_topic_partition_list_query_leaders().
 *
 * The reply rko op will contain:
 * - .leaders which is a list of leaders and their partitions, this may be
 *    NULL for overall errors (such as no leaders are found), or a
 *    partial or complete list of leaders.
 * - .partitions which is a copy of the input list of partitions with the
 *   .err field set to the outcome of the leader query, typically ERR_NO_ERROR
 *   or ERR_UNKNOWN_TOPIC_OR_PART.
 *
 * @locks_acquired rd_kafka_*lock()
 *
 * @remark rd_kafka_*lock() MUST NOT be held
 */
void rd_kafka_topic_partition_list_query_leaders_async(
    rd_kafka_t *rk,
    const rd_kafka_topic_partition_list_t *rktparlist,
    int timeout_ms,
    rd_kafka_replyq_t replyq,
    rd_kafka_op_cb_t *cb,
    void *opaque) {
        rd_kafka_op_t *rko;

        rd_assert(rktparlist && rktparlist->cnt > 0);
        rd_assert(replyq.q);

        rko = rd_kafka_op_new_cb(
            rk, RD_KAFKA_OP_LEADERS,
            rd_kafka_topic_partition_list_query_leaders_async_worker_op_cb);
        rko->rko_u.leaders.replyq = replyq;
        rko->rko_u.leaders.partitions =
            rd_kafka_topic_partition_list_copy(rktparlist);
        rko->rko_u.leaders.ts_timeout = rd_timeout_init(timeout_ms);
        rko->rko_u.leaders.cb         = cb;
        rko->rko_u.leaders.opaque     = opaque;

        /* Create an eonce to be triggered either by metadata cache update
         * (from refresh_topics()), query interval, or timeout. */
        rko->rko_u.leaders.eonce =
            rd_kafka_enq_once_new(rko, RD_KAFKA_REPLYQ(rk->rk_ops, 0));

        rd_kafka_enq_once_add_source(rko->rko_u.leaders.eonce, "timeout timer");
        rd_kafka_timer_start_oneshot(
            &rk->rk_timers, &rko->rko_u.leaders.timeout_tmr, rd_true,
            rd_timeout_remains_us(rko->rko_u.leaders.ts_timeout),
            rd_kafka_partition_leader_query_eonce_timeout_cb,
            rko->rko_u.leaders.eonce);

        if (rd_kafka_topic_partition_list_query_leaders_async_worker(rko) ==
            RD_KAFKA_OP_RES_HANDLED)
                rd_kafka_op_destroy(rko); /* Reply queue already disabled */
}


/**
 * @brief Get leaders for all partitions in \p rktparlist, querying metadata
 *        if needed.
 *
 * @param leaders is a pre-initialized (empty) list which will be populated
 *        with the leader brokers and their partitions
 *        (struct rd_kafka_partition_leader *)
 *
 * @remark Will not trigger topic auto creation (unless configured).
 *
 * @returns an error code on error.
 *
 * @locks rd_kafka_*lock() MUST NOT be held
 */
rd_kafka_resp_err_t rd_kafka_topic_partition_list_query_leaders(
    rd_kafka_t *rk,
    rd_kafka_topic_partition_list_t *rktparlist,
    rd_list_t *leaders,
    int timeout_ms) {
        rd_ts_t ts_end   = rd_timeout_init(timeout_ms);
        rd_ts_t ts_query = 0;
        rd_ts_t now;
        int query_cnt = 0;
        int i         = 0;

        /* Get all the partition leaders, try multiple times:
         * if there are no leaders after the first run fire off a leader
         * query and wait for broker state update before trying again,
         * keep trying and re-querying at increasing intervals until
         * success or timeout. */
        do {
                rd_list_t query_topics;
                int query_intvl;

                rd_list_init(&query_topics, rktparlist->cnt, rd_free);

                rd_kafka_topic_partition_list_get_leaders(
                    rk, rktparlist, leaders, &query_topics,
                    /* Add unknown topics to query_topics only on the
                     * first query, after that we consider them
                     * permanently non-existent */
                    query_cnt == 0, NULL);

                if (rd_list_empty(&query_topics)) {
                        /* No remaining topics to query: leader-list complete.*/
                        rd_list_destroy(&query_topics);

                        /* No leader(s) for partitions means all partitions
                         * are unknown. */
                        if (rd_list_empty(leaders))
                                return RD_KAFKA_RESP_ERR__UNKNOWN_PARTITION;

                        return RD_KAFKA_RESP_ERR_NO_ERROR;
                }

                now = rd_clock();

                /*
                 * Missing leader for some partitions
                 */
                query_intvl = (i + 1) * 100; /* add 100ms per iteration */
                if (query_intvl > 2 * 1000)
                        query_intvl = 2 * 1000; /* Cap to 2s */

                if (now >= ts_query + (query_intvl * 1000)) {
                        /* Query metadata for missing leaders,
                         * possibly creating the topic. */
                        rd_kafka_metadata_refresh_topics(
                            rk, NULL, &query_topics, rd_true /*force*/,
                            rd_false /*!allow_auto_create*/,
                            rd_false /*!cgrp_update*/,
                            "query partition leaders");
                        ts_query = now;
                        query_cnt++;

                } else {
                        /* Wait for broker ids to be updated from
                         * metadata refresh above. */
                        int wait_ms =
                            rd_timeout_remains_limit(ts_end, query_intvl);
                        rd_kafka_metadata_cache_wait_change(rk, wait_ms);
                }

                rd_list_destroy(&query_topics);

                i++;
        } while (ts_end == RD_POLL_INFINITE ||
                 now < ts_end); /* now is deliberately outdated here
                                 * since wait_change() will block.
                                 * This gives us one more chance to spin thru*/

        if (rd_atomic32_get(&rk->rk_broker_up_cnt) == 0)
                return RD_KAFKA_RESP_ERR__ALL_BROKERS_DOWN;

        return RD_KAFKA_RESP_ERR__TIMED_OUT;
}


/**
 * @brief Populate \p rkts with the rd_kafka_topic_t objects for the
 *        partitions in. Duplicates are suppressed.
 *
 * @returns the number of topics added.
 */
int rd_kafka_topic_partition_list_get_topics(
    rd_kafka_t *rk,
    rd_kafka_topic_partition_list_t *rktparlist,
    rd_list_t *rkts) {
        int cnt = 0;

        int i;
        for (i = 0; i < rktparlist->cnt; i++) {
                rd_kafka_topic_partition_t *rktpar = &rktparlist->elems[i];
                rd_kafka_toppar_t *rktp;

                rktp =
                    rd_kafka_topic_partition_get_toppar(rk, rktpar, rd_false);
                if (!rktp) {
                        rktpar->err = RD_KAFKA_RESP_ERR__UNKNOWN_PARTITION;
                        continue;
                }

                if (!rd_list_find(rkts, rktp->rktp_rkt,
                                  rd_kafka_topic_cmp_rkt)) {
                        rd_list_add(rkts, rd_kafka_topic_keep(rktp->rktp_rkt));
                        cnt++;
                }

                rd_kafka_toppar_destroy(rktp);
        }

        return cnt;
}


/**
 * @brief Populate \p topics with the strdupped topic names in \p rktparlist.
 *        Duplicates are suppressed.
 *
 * @param include_regex: include regex topics
 *
 * @returns the number of topics added.
 */
int rd_kafka_topic_partition_list_get_topic_names(
    const rd_kafka_topic_partition_list_t *rktparlist,
    rd_list_t *topics,
    int include_regex) {
        int cnt = 0;
        int i;

        for (i = 0; i < rktparlist->cnt; i++) {
                const rd_kafka_topic_partition_t *rktpar =
                    &rktparlist->elems[i];

                if (!include_regex && *rktpar->topic == '^')
                        continue;

                if (!rd_list_find(topics, rktpar->topic, (void *)strcmp)) {
                        rd_list_add(topics, rd_strdup(rktpar->topic));
                        cnt++;
                }
        }

        return cnt;
}


/**
 * @brief Create a copy of \p rktparlist only containing the partitions
 *        matched by \p match function.
 *
 * \p match shall return 1 for match, else 0.
 *
 * @returns a new list
 */
rd_kafka_topic_partition_list_t *rd_kafka_topic_partition_list_match(
    const rd_kafka_topic_partition_list_t *rktparlist,
    int (*match)(const void *elem, const void *opaque),
    void *opaque) {
        rd_kafka_topic_partition_list_t *newlist;
        int i;

        newlist = rd_kafka_topic_partition_list_new(0);

        for (i = 0; i < rktparlist->cnt; i++) {
                const rd_kafka_topic_partition_t *rktpar =
                    &rktparlist->elems[i];

                if (!match(rktpar, opaque))
                        continue;

                rd_kafka_topic_partition_list_add_copy(newlist, rktpar);
        }

        return newlist;
}

void rd_kafka_topic_partition_list_log(
    rd_kafka_t *rk,
    const char *fac,
    int dbg,
    const rd_kafka_topic_partition_list_t *rktparlist) {
        int i;

        rd_kafka_dbg(rk, NONE | dbg, fac,
                     "List with %d partition(s):", rktparlist->cnt);
        for (i = 0; i < rktparlist->cnt; i++) {
                const rd_kafka_topic_partition_t *rktpar =
                    &rktparlist->elems[i];
                rd_kafka_dbg(rk, NONE | dbg, fac,
                             " %s [%" PRId32 "] offset %s%s%s", rktpar->topic,
                             rktpar->partition,
                             rd_kafka_offset2str(rktpar->offset),
                             rktpar->err ? ": error: " : "",
                             rktpar->err ? rd_kafka_err2str(rktpar->err) : "");
        }
}

/**
 * @returns a comma-separated list of partitions.
 */
const char *rd_kafka_topic_partition_list_str(
    const rd_kafka_topic_partition_list_t *rktparlist,
    char *dest,
    size_t dest_size,
    int fmt_flags) {
        int i;
        size_t of = 0;

        for (i = 0; i < rktparlist->cnt; i++) {
                const rd_kafka_topic_partition_t *rktpar =
                    &rktparlist->elems[i];
                char errstr[128];
                char offsetstr[32];
                int r;

                if (!rktpar->err && (fmt_flags & RD_KAFKA_FMT_F_ONLY_ERR))
                        continue;

                if (rktpar->err && !(fmt_flags & RD_KAFKA_FMT_F_NO_ERR))
                        rd_snprintf(errstr, sizeof(errstr), "(%s)",
                                    rd_kafka_err2str(rktpar->err));
                else
                        errstr[0] = '\0';

                if (rktpar->offset != RD_KAFKA_OFFSET_INVALID)
                        rd_snprintf(offsetstr, sizeof(offsetstr), "@%" PRId64,
                                    rktpar->offset);
                else
                        offsetstr[0] = '\0';

                r = rd_snprintf(&dest[of], dest_size - of,
                                "%s"
                                "%s[%" PRId32
                                "]"
                                "%s"
                                "%s",
                                of == 0 ? "" : ", ", rktpar->topic,
                                rktpar->partition, offsetstr, errstr);

                if ((size_t)r >= dest_size - of) {
                        rd_snprintf(&dest[dest_size - 4], 4, "...");
                        break;
                }

                of += r;
        }

        return dest;
}



/**
 * @brief Update \p dst with info from \p src.
 *
 * Fields updated:
 *  - metadata
 *  - metadata_size
 *  - offset
 *  - offset leader epoch
 *  - err
 *
 * Will only update partitions that are in both dst and src, other partitions
 * will remain unchanged.
 */
void rd_kafka_topic_partition_list_update(
    rd_kafka_topic_partition_list_t *dst,
    const rd_kafka_topic_partition_list_t *src) {
        int i;

        for (i = 0; i < dst->cnt; i++) {
                rd_kafka_topic_partition_t *d = &dst->elems[i];
                rd_kafka_topic_partition_t *s;
                rd_kafka_topic_partition_private_t *s_priv, *d_priv;

                if (!(s = rd_kafka_topic_partition_list_find(
                          (rd_kafka_topic_partition_list_t *)src, d->topic,
                          d->partition)))
                        continue;

                d->offset = s->offset;
                d->err    = s->err;
                if (d->metadata) {
                        rd_free(d->metadata);
                        d->metadata      = NULL;
                        d->metadata_size = 0;
                }
                if (s->metadata_size > 0) {
                        d->metadata      = rd_malloc(s->metadata_size);
                        d->metadata_size = s->metadata_size;
                        memcpy((void *)d->metadata, s->metadata,
                               s->metadata_size);
                }

                s_priv               = rd_kafka_topic_partition_get_private(s);
                d_priv               = rd_kafka_topic_partition_get_private(d);
                d_priv->leader_epoch = s_priv->leader_epoch;
        }
}


/**
 * @returns the sum of \p cb called for each element.
 */
size_t rd_kafka_topic_partition_list_sum(
    const rd_kafka_topic_partition_list_t *rktparlist,
    size_t (*cb)(const rd_kafka_topic_partition_t *rktpar, void *opaque),
    void *opaque) {
        int i;
        size_t sum = 0;

        for (i = 0; i < rktparlist->cnt; i++) {
                const rd_kafka_topic_partition_t *rktpar =
                    &rktparlist->elems[i];
                sum += cb(rktpar, opaque);
        }

        return sum;
}


/**
 * @returns rd_true if there are duplicate topic/partitions in the list,
 *          rd_false if not.
 *
 * @remarks sorts the elements of the list.
 */
rd_bool_t rd_kafka_topic_partition_list_has_duplicates(
    rd_kafka_topic_partition_list_t *rktparlist,
    rd_bool_t ignore_partition) {

        int i;

        if (rktparlist->cnt <= 1)
                return rd_false;

        rd_kafka_topic_partition_list_sort_by_topic(rktparlist);

        for (i = 1; i < rktparlist->cnt; i++) {
                const rd_kafka_topic_partition_t *p1 =
                    &rktparlist->elems[i - 1];
                const rd_kafka_topic_partition_t *p2 = &rktparlist->elems[i];

                if (((p1->partition == p2->partition) || ignore_partition) &&
                    !strcmp(p1->topic, p2->topic)) {
                        return rd_true;
                }
        }

        return rd_false;
}


/**
 * @brief Set \c .err field \p err on all partitions in list.
 */
void rd_kafka_topic_partition_list_set_err(
    rd_kafka_topic_partition_list_t *rktparlist,
    rd_kafka_resp_err_t err) {
        int i;

        for (i = 0; i < rktparlist->cnt; i++)
                rktparlist->elems[i].err = err;
}

/**
 * @brief Get the first set error in the partition list.
 */
rd_kafka_resp_err_t rd_kafka_topic_partition_list_get_err(
    const rd_kafka_topic_partition_list_t *rktparlist) {
        int i;

        for (i = 0; i < rktparlist->cnt; i++)
                if (rktparlist->elems[i].err)
                        return rktparlist->elems[i].err;

        return RD_KAFKA_RESP_ERR_NO_ERROR;
}


/**
 * @returns the number of wildcard/regex topics
 */
int rd_kafka_topic_partition_list_regex_cnt(
    const rd_kafka_topic_partition_list_t *rktparlist) {
        int i;
        int cnt = 0;

        for (i = 0; i < rktparlist->cnt; i++) {
                const rd_kafka_topic_partition_t *rktpar =
                    &rktparlist->elems[i];
                cnt += *rktpar->topic == '^';
        }
        return cnt;
}


/**
 * @brief Reset base sequence for this toppar.
 *
 * See rd_kafka_toppar_pid_change() below.
 *
 * @warning Toppar must be completely drained.
 *
 * @locality toppar handler thread
 * @locks toppar_lock MUST be held.
 */
static void rd_kafka_toppar_reset_base_msgid(rd_kafka_toppar_t *rktp,
                                             uint64_t new_base_msgid) {
        rd_kafka_dbg(
            rktp->rktp_rkt->rkt_rk, TOPIC | RD_KAFKA_DBG_EOS, "RESETSEQ",
            "%.*s [%" PRId32
            "] "
            "resetting epoch base seq from %" PRIu64 " to %" PRIu64,
            RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic), rktp->rktp_partition,
            rktp->rktp_eos.epoch_base_msgid, new_base_msgid);

        rktp->rktp_eos.next_ack_seq     = 0;
        rktp->rktp_eos.next_err_seq     = 0;
        rktp->rktp_eos.epoch_base_msgid = new_base_msgid;
}


/**
 * @brief Update/change the Producer ID for this toppar.
 *
 * Must only be called when pid is different from the current toppar pid.
 *
 * The epoch base sequence will be set to \p base_msgid, which must be the
 * first message in the partition
 * queue. However, if there are outstanding messages in-flight to the broker
 * we will need to wait for these ProduceRequests to finish (most likely
 * with failure) and have their messages re-enqueued to maintain original order.
 * In this case the pid will not be updated and this function should be
 * called again when there are no outstanding messages.
 *
 * @remark This function must only be called when rktp_xmitq is non-empty.
 *
 * @returns 1 if a new pid was set, else 0.
 *
 * @locality toppar handler thread
 * @locks none
 */
int rd_kafka_toppar_pid_change(rd_kafka_toppar_t *rktp,
                               rd_kafka_pid_t pid,
                               uint64_t base_msgid) {
        int inflight = rd_atomic32_get(&rktp->rktp_msgs_inflight);

        if (unlikely(inflight > 0)) {
                rd_kafka_dbg(
                    rktp->rktp_rkt->rkt_rk, TOPIC | RD_KAFKA_DBG_EOS, "NEWPID",
                    "%.*s [%" PRId32
                    "] will not change %s -> %s yet: "
                    "%d message(s) still in-flight from current "
                    "epoch",
                    RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                    rktp->rktp_partition, rd_kafka_pid2str(rktp->rktp_eos.pid),
                    rd_kafka_pid2str(pid), inflight);
                return 0;
        }

        rd_assert(base_msgid != 0 &&
                  *"BUG: pid_change() must only be called with "
                  "non-empty xmitq");

        rd_kafka_toppar_lock(rktp);
        rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, TOPIC | RD_KAFKA_DBG_EOS, "NEWPID",
                     "%.*s [%" PRId32
                     "] changed %s -> %s "
                     "with base MsgId %" PRIu64,
                     RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                     rktp->rktp_partition, rd_kafka_pid2str(rktp->rktp_eos.pid),
                     rd_kafka_pid2str(pid), base_msgid);

        rktp->rktp_eos.pid = pid;
        rd_kafka_toppar_reset_base_msgid(rktp, base_msgid);

        rd_kafka_toppar_unlock(rktp);

        return 1;
}


/**
 * @brief Purge messages in partition queues.
 *        Delivery reports will be enqueued for all purged messages, the error
 *        code is set to RD_KAFKA_RESP_ERR__PURGE_QUEUE.
 *
 * @param include_xmit_msgq If executing from the rktp's current broker handler
 *                          thread, also include the xmit message queue.
 *
 * @warning Only to be used with the producer.
 *
 * @returns the number of messages purged
 *
 * @locality any thread.
 * @locks_acquired rd_kafka_toppar_lock()
 * @locks_required none
 */
int rd_kafka_toppar_purge_queues(rd_kafka_toppar_t *rktp,
                                 int purge_flags,
                                 rd_bool_t include_xmit_msgq) {
        rd_kafka_t *rk       = rktp->rktp_rkt->rkt_rk;
        rd_kafka_msgq_t rkmq = RD_KAFKA_MSGQ_INITIALIZER(rkmq);
        int cnt;

        rd_assert(rk->rk_type == RD_KAFKA_PRODUCER);

        rd_kafka_dbg(rk, TOPIC, "PURGE",
                     "%s [%" PRId32
                     "]: purging queues "
                     "(purge_flags 0x%x, %s xmit_msgq)",
                     rktp->rktp_rkt->rkt_topic->str, rktp->rktp_partition,
                     purge_flags, include_xmit_msgq ? "include" : "exclude");

        if (!(purge_flags & RD_KAFKA_PURGE_F_QUEUE))
                return 0;

        if (include_xmit_msgq) {
                /* xmit_msgq is owned by the toppar handler thread
                 * (broker thread) and requires no locking. */
                rd_assert(rktp->rktp_broker);
                rd_assert(thrd_is_current(rktp->rktp_broker->rkb_thread));
                rd_kafka_msgq_concat(&rkmq, &rktp->rktp_xmit_msgq);
        }

        rd_kafka_toppar_lock(rktp);
        rd_kafka_msgq_concat(&rkmq, &rktp->rktp_msgq);
        cnt = rd_kafka_msgq_len(&rkmq);

        if (cnt > 0 && purge_flags & RD_KAFKA_PURGE_F_ABORT_TXN) {
                /* All messages in-queue are purged
                 * on abort_transaction(). Since these messages
                 * will not be produced (retried) we need to adjust the
                 * idempotence epoch's base msgid to skip the messages. */
                rktp->rktp_eos.epoch_base_msgid += cnt;
                rd_kafka_dbg(rk, TOPIC | RD_KAFKA_DBG_EOS, "ADVBASE",
                             "%.*s [%" PRId32
                             "] "
                             "advancing epoch base msgid to %" PRIu64
                             " due to %d message(s) in aborted transaction",
                             RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
                             rktp->rktp_partition,
                             rktp->rktp_eos.epoch_base_msgid, cnt);
        }
        rd_kafka_toppar_unlock(rktp);

        rd_kafka_dr_msgq(rktp->rktp_rkt, &rkmq, RD_KAFKA_RESP_ERR__PURGE_QUEUE);

        return cnt;
}


/**
 * @brief Purge queues for the unassigned toppars of all known topics.
 *
 * @locality application thread
 * @locks none
 */
void rd_kafka_purge_ua_toppar_queues(rd_kafka_t *rk) {
        rd_kafka_topic_t *rkt;
        int msg_cnt = 0, part_cnt = 0;

        rd_kafka_rdlock(rk);
        TAILQ_FOREACH(rkt, &rk->rk_topics, rkt_link) {
                rd_kafka_toppar_t *rktp;
                int r;

                rd_kafka_topic_rdlock(rkt);
                rktp = rkt->rkt_ua;
                if (rktp)
                        rd_kafka_toppar_keep(rktp);
                rd_kafka_topic_rdunlock(rkt);

                if (unlikely(!rktp))
                        continue;


                rd_kafka_toppar_lock(rktp);

                r = rd_kafka_msgq_len(&rktp->rktp_msgq);
                rd_kafka_dr_msgq(rkt, &rktp->rktp_msgq,
                                 RD_KAFKA_RESP_ERR__PURGE_QUEUE);
                rd_kafka_toppar_unlock(rktp);
                rd_kafka_toppar_destroy(rktp);

                if (r > 0) {
                        msg_cnt += r;
                        part_cnt++;
                }
        }
        rd_kafka_rdunlock(rk);

        rd_kafka_dbg(rk, QUEUE | RD_KAFKA_DBG_TOPIC, "PURGEQ",
                     "Purged %i message(s) from %d UA-partition(s)", msg_cnt,
                     part_cnt);
}


void rd_kafka_partition_leader_destroy_free(void *ptr) {
        struct rd_kafka_partition_leader *leader = ptr;
        rd_kafka_partition_leader_destroy(leader);
}


const char *rd_kafka_fetch_pos2str(const rd_kafka_fetch_pos_t fetchpos) {
        static RD_TLS char ret[2][64];
        static int idx;

        idx = (idx + 1) % 2;

        rd_snprintf(
            ret[idx], sizeof(ret[idx]), "offset %s (leader epoch %" PRId32 ")",
            rd_kafka_offset2str(fetchpos.offset), fetchpos.leader_epoch);

        return ret[idx];
}