@platformatic/rdkafka 4.0.0

package/deps/librdkafka/src/rdkafka_queue.c

@@ -0,0 +1,1153 @@
+/*
+ * librdkafka - The Apache Kafka C/C++ library
+ *
+ * Copyright (c) 2016-2022, Magnus Edenhill,
+ *               2023, Confluent Inc.
+ * 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_offset.h"
+#include "rdkafka_topic.h"
+#include "rdkafka_interceptor.h"
+
+int RD_TLS rd_kafka_yield_thread = 0;
+
+void rd_kafka_yield(rd_kafka_t *rk) {
+        rd_kafka_yield_thread = 1;
+}
+
+
+/**
+ * @brief Check and reset yield flag.
+ * @returns rd_true if caller should yield, otherwise rd_false.
+ * @remarks rkq_lock MUST be held
+ */
+static RD_INLINE rd_bool_t rd_kafka_q_check_yield(rd_kafka_q_t *rkq) {
+        if (!(rkq->rkq_flags & RD_KAFKA_Q_F_YIELD))
+                return rd_false;
+
+        rkq->rkq_flags &= ~RD_KAFKA_Q_F_YIELD;
+        return rd_true;
+}
+/**
+ * Destroy a queue. refcnt must be at zero.
+ */
+void rd_kafka_q_destroy_final(rd_kafka_q_t *rkq) {
+
+        mtx_lock(&rkq->rkq_lock);
+        if (unlikely(rkq->rkq_qio != NULL)) {
+                rd_free(rkq->rkq_qio);
+                rkq->rkq_qio = NULL;
+        }
+        /* Queue must have been disabled prior to final destruction,
+         * this is to catch the case where the queue owner/poll does not
+         * use rd_kafka_q_destroy_owner(). */
+        rd_dassert(!(rkq->rkq_flags & RD_KAFKA_Q_F_READY));
+        rd_kafka_q_disable0(rkq, 0 /*no-lock*/); /* for the non-devel case */
+        rd_kafka_q_fwd_set0(rkq, NULL, 0 /*no-lock*/, 0 /*no-fwd-app*/);
+        rd_kafka_q_purge0(rkq, 0 /*no-lock*/);
+        assert(!rkq->rkq_fwdq);
+        mtx_unlock(&rkq->rkq_lock);
+        mtx_destroy(&rkq->rkq_lock);
+        cnd_destroy(&rkq->rkq_cond);
+
+        if (rkq->rkq_flags & RD_KAFKA_Q_F_ALLOCATED)
+                rd_free(rkq);
+}
+
+
+
+/**
+ * Initialize a queue.
+ */
+void rd_kafka_q_init0(rd_kafka_q_t *rkq,
+                      rd_kafka_t *rk,
+                      rd_bool_t for_consume,
+                      const char *func,
+                      int line) {
+        rd_kafka_q_reset(rkq);
+        rkq->rkq_fwdq   = NULL;
+        rkq->rkq_refcnt = 1;
+        rkq->rkq_flags  = RD_KAFKA_Q_F_READY;
+        if (for_consume)
+                rkq->rkq_flags |= RD_KAFKA_Q_F_CONSUMER;
+        rkq->rkq_rk     = rk;
+        rkq->rkq_qio    = NULL;
+        rkq->rkq_serve  = NULL;
+        rkq->rkq_opaque = NULL;
+        mtx_init(&rkq->rkq_lock, mtx_plain);
+        cnd_init(&rkq->rkq_cond);
+#if ENABLE_DEVEL
+        rd_snprintf(rkq->rkq_name, sizeof(rkq->rkq_name), "%s:%d", func, line);
+#else
+        rkq->rkq_name = func;
+#endif
+}
+
+
+/**
+ * Allocate a new queue and initialize it.
+ */
+rd_kafka_q_t *rd_kafka_q_new0(rd_kafka_t *rk,
+                              rd_bool_t for_consume,
+                              const char *func,
+                              int line) {
+        rd_kafka_q_t *rkq = rd_malloc(sizeof(*rkq));
+        if (!for_consume)
+                rd_kafka_q_init(rkq, rk);
+        else
+                rd_kafka_consume_q_init(rkq, rk);
+        rkq->rkq_flags |= RD_KAFKA_Q_F_ALLOCATED;
+#if ENABLE_DEVEL
+        rd_snprintf(rkq->rkq_name, sizeof(rkq->rkq_name), "%s:%d", func, line);
+#else
+        rkq->rkq_name = func;
+#endif
+        return rkq;
+}
+
+/*
+ * Sets the flag RD_KAFKA_Q_F_CONSUMER for rkq, any queues it's being forwarded
+ * to, recursively.
+ * Setting this flag indicates that polling this queue is equivalent to calling
+ * consumer poll, and will reset the max.poll.interval.ms timer. Only used
+ * internally when forwarding queues.
+ * @locks rd_kafka_q_lock(rkq)
+ */
+static void rd_kafka_q_consumer_propagate(rd_kafka_q_t *rkq) {
+        mtx_lock(&rkq->rkq_lock);
+        rkq->rkq_flags |= RD_KAFKA_Q_F_CONSUMER;
+
+        if (!rkq->rkq_fwdq) {
+                mtx_unlock(&rkq->rkq_lock);
+                return;
+        }
+
+        /* Recursively propagate the flag to any queues rkq is already
+         * forwarding to. There will be a deadlock here if the queues are being
+         * forwarded circularly, but that is a user error. We can't resolve this
+         * deadlock by unlocking before the recursive call, because that leads
+         * to incorrectness if the rkq_fwdq is forwarded elsewhere and the old
+         * one destroyed between recursive calls. */
+        rd_kafka_q_consumer_propagate(rkq->rkq_fwdq);
+        mtx_unlock(&rkq->rkq_lock);
+}
+
+/**
+ * Set/clear forward queue.
+ * Queue forwarding enables message routing inside rdkafka.
+ * Typical use is to re-route all fetched messages for all partitions
+ * to one single queue.
+ *
+ * All access to rkq_fwdq are protected by rkq_lock.
+ */
+void rd_kafka_q_fwd_set0(rd_kafka_q_t *srcq,
+                         rd_kafka_q_t *destq,
+                         int do_lock,
+                         int fwd_app) {
+        if (unlikely(srcq == destq))
+                return;
+
+        if (do_lock)
+                mtx_lock(&srcq->rkq_lock);
+        if (fwd_app)
+                srcq->rkq_flags |= RD_KAFKA_Q_F_FWD_APP;
+        if (srcq->rkq_fwdq) {
+                rd_kafka_q_destroy(srcq->rkq_fwdq);
+                srcq->rkq_fwdq = NULL;
+        }
+        if (destq) {
+                rd_kafka_q_keep(destq);
+
+                /* If rkq has ops in queue, append them to fwdq's queue.
+                 * This is an irreversible operation. */
+                if (srcq->rkq_qlen > 0) {
+                        rd_dassert(destq->rkq_flags & RD_KAFKA_Q_F_READY);
+                        rd_kafka_q_concat(destq, srcq);
+                }
+
+                srcq->rkq_fwdq = destq;
+
+                if (srcq->rkq_flags & RD_KAFKA_Q_F_CONSUMER)
+                        rd_kafka_q_consumer_propagate(destq);
+        }
+        if (do_lock)
+                mtx_unlock(&srcq->rkq_lock);
+}
+
+/**
+ * Purge all entries from a queue.
+ */
+int rd_kafka_q_purge0(rd_kafka_q_t *rkq, int do_lock) {
+        rd_kafka_op_t *rko, *next;
+        TAILQ_HEAD(, rd_kafka_op_s) tmpq = TAILQ_HEAD_INITIALIZER(tmpq);
+        rd_kafka_q_t *fwdq;
+        int cnt = 0;
+
+        if (do_lock)
+                mtx_lock(&rkq->rkq_lock);
+
+        if ((fwdq = rd_kafka_q_fwd_get(rkq, 0))) {
+                if (do_lock)
+                        mtx_unlock(&rkq->rkq_lock);
+                cnt = rd_kafka_q_purge(fwdq);
+                rd_kafka_q_destroy(fwdq);
+                return cnt;
+        }
+
+        /* Move ops queue to tmpq to avoid lock-order issue
+         * by locks taken from rd_kafka_op_destroy(). */
+        TAILQ_MOVE(&tmpq, &rkq->rkq_q, rko_link);
+
+        rd_kafka_q_mark_served(rkq);
+
+        /* Zero out queue */
+        rd_kafka_q_reset(rkq);
+
+        if (do_lock)
+                mtx_unlock(&rkq->rkq_lock);
+
+        /* Destroy the ops */
+        next = TAILQ_FIRST(&tmpq);
+        while ((rko = next)) {
+                next = TAILQ_NEXT(next, rko_link);
+                rd_kafka_op_destroy(rko);
+                cnt++;
+        }
+
+        return cnt;
+}
+
+
+/**
+ * Purge all entries from a queue with a rktp version smaller than `version`
+ * This shaves off the head of the queue, up until the first rko with
+ * a non-matching rktp or version.
+ */
+void rd_kafka_q_purge_toppar_version(rd_kafka_q_t *rkq,
+                                     rd_kafka_toppar_t *rktp,
+                                     int version) {
+        rd_kafka_op_t *rko, *next;
+        TAILQ_HEAD(, rd_kafka_op_s) tmpq = TAILQ_HEAD_INITIALIZER(tmpq);
+        int32_t cnt  = 0;
+        int64_t size = 0;
+        rd_kafka_q_t *fwdq;
+
+        mtx_lock(&rkq->rkq_lock);
+
+        if ((fwdq = rd_kafka_q_fwd_get(rkq, 0))) {
+                mtx_unlock(&rkq->rkq_lock);
+                rd_kafka_q_purge_toppar_version(fwdq, rktp, version);
+                rd_kafka_q_destroy(fwdq);
+                return;
+        }
+
+        /* Move ops to temporary queue and then destroy them from there
+         * without locks to avoid lock-ordering problems in op_destroy() */
+        while ((rko = TAILQ_FIRST(&rkq->rkq_q)) && rko->rko_rktp &&
+               rko->rko_rktp == rktp && rko->rko_version < version) {
+                TAILQ_REMOVE(&rkq->rkq_q, rko, rko_link);
+                TAILQ_INSERT_TAIL(&tmpq, rko, rko_link);
+                cnt++;
+                size += rko->rko_len;
+        }
+
+        rd_kafka_q_mark_served(rkq);
+
+        rkq->rkq_qlen -= cnt;
+        rkq->rkq_qsize -= size;
+        mtx_unlock(&rkq->rkq_lock);
+
+        next = TAILQ_FIRST(&tmpq);
+        while ((rko = next)) {
+                next = TAILQ_NEXT(next, rko_link);
+                rd_kafka_op_destroy(rko);
+        }
+}
+
+
+/**
+ * Move 'cnt' entries from 'srcq' to 'dstq'.
+ * If 'cnt' == -1 all entries will be moved.
+ * Returns the number of entries moved.
+ */
+int rd_kafka_q_move_cnt(rd_kafka_q_t *dstq,
+                        rd_kafka_q_t *srcq,
+                        int cnt,
+                        int do_locks) {
+        rd_kafka_op_t *rko;
+        int mcnt = 0;
+
+        if (do_locks) {
+                mtx_lock(&srcq->rkq_lock);
+                mtx_lock(&dstq->rkq_lock);
+        }
+
+        if (!dstq->rkq_fwdq && !srcq->rkq_fwdq) {
+                if (cnt > 0 && dstq->rkq_qlen == 0)
+                        rd_kafka_q_io_event(dstq);
+
+                /* Optimization, if 'cnt' is equal/larger than all
+                 * items of 'srcq' we can move the entire queue. */
+                if (cnt == -1 || cnt >= (int)srcq->rkq_qlen) {
+                        mcnt = srcq->rkq_qlen;
+                        rd_kafka_q_concat0(dstq, srcq, 0 /*no-lock*/);
+                } else {
+                        while (mcnt < cnt &&
+                               (rko = TAILQ_FIRST(&srcq->rkq_q))) {
+                                TAILQ_REMOVE(&srcq->rkq_q, rko, rko_link);
+                                if (likely(!rko->rko_prio))
+                                        TAILQ_INSERT_TAIL(&dstq->rkq_q, rko,
+                                                          rko_link);
+                                else
+                                        TAILQ_INSERT_SORTED(
+                                            &dstq->rkq_q, rko, rd_kafka_op_t *,
+                                            rko_link, rd_kafka_op_cmp_prio);
+
+                                srcq->rkq_qlen--;
+                                dstq->rkq_qlen++;
+                                srcq->rkq_qsize -= rko->rko_len;
+                                dstq->rkq_qsize += rko->rko_len;
+                                mcnt++;
+                        }
+                }
+
+                rd_kafka_q_mark_served(srcq);
+
+        } else
+                mcnt = rd_kafka_q_move_cnt(
+                    dstq->rkq_fwdq ? dstq->rkq_fwdq : dstq,
+                    srcq->rkq_fwdq ? srcq->rkq_fwdq : srcq, cnt, do_locks);
+
+        if (do_locks) {
+                mtx_unlock(&dstq->rkq_lock);
+                mtx_unlock(&srcq->rkq_lock);
+        }
+
+        return mcnt;
+}
+
+
+/**
+ * Filters out outdated ops.
+ */
+static RD_INLINE rd_kafka_op_t *
+rd_kafka_op_filter(rd_kafka_q_t *rkq, rd_kafka_op_t *rko, int version) {
+        if (unlikely(!rko))
+                return NULL;
+
+        if (unlikely(rd_kafka_op_version_outdated(rko, version))) {
+                rd_kafka_q_deq0(rkq, rko);
+                rd_kafka_op_destroy(rko);
+                return NULL;
+        }
+
+        return rko;
+}
+
+
+
+/**
+ * Pop an op from a queue.
+ *
+ * Locality: any thread.
+ */
+
+
+/**
+ * Serve q like rd_kafka_q_serve() until an op is found that can be returned
+ * as an event to the application.
+ *
+ * @returns the first event:able op, or NULL on timeout.
+ *
+ * Locality: any thread
+ */
+rd_kafka_op_t *rd_kafka_q_pop_serve(rd_kafka_q_t *rkq,
+                                    rd_ts_t timeout_us,
+                                    int32_t version,
+                                    rd_kafka_q_cb_type_t cb_type,
+                                    rd_kafka_q_serve_cb_t *callback,
+                                    void *opaque) {
+        rd_kafka_op_t *rko;
+        rd_kafka_q_t *fwdq;
+
+        rd_dassert(cb_type);
+
+        mtx_lock(&rkq->rkq_lock);
+
+        rd_kafka_yield_thread = 0;
+        if (!(fwdq = rd_kafka_q_fwd_get(rkq, 0))) {
+                const rd_bool_t can_q_contain_fetched_msgs =
+                    rd_kafka_q_can_contain_fetched_msgs(rkq, RD_DONT_LOCK);
+
+                struct timespec timeout_tspec;
+
+                rd_timeout_init_timespec_us(&timeout_tspec, timeout_us);
+
+                if (can_q_contain_fetched_msgs)
+                        rd_kafka_app_poll_start(rkq->rkq_rk, 0, timeout_us);
+
+                while (1) {
+                        rd_kafka_op_res_t res;
+                        /* Keep track of current lock status to avoid
+                         * unnecessary lock flapping in all the cases below. */
+                        rd_bool_t is_locked = rd_true;
+
+                        /* Filter out outdated ops */
+                retry:
+                        while ((rko = TAILQ_FIRST(&rkq->rkq_q)) &&
+                               !(rko = rd_kafka_op_filter(rkq, rko, version)))
+                                ;
+
+                        rd_kafka_q_mark_served(rkq);
+
+                        if (rko) {
+                                /* Proper versioned op */
+                                rd_kafka_q_deq0(rkq, rko);
+
+                                /* Let op_handle() operate without lock
+                                 * held to allow re-enqueuing, etc. */
+                                mtx_unlock(&rkq->rkq_lock);
+                                is_locked = rd_false;
+
+                                /* Ops with callbacks are considered handled
+                                 * and we move on to the next op, if any.
+                                 * Ops w/o callbacks are returned immediately */
+                                res = rd_kafka_op_handle(rkq->rkq_rk, rkq, rko,
+                                                         cb_type, opaque,
+                                                         callback);
+
+                                if (res == RD_KAFKA_OP_RES_HANDLED ||
+                                    res == RD_KAFKA_OP_RES_KEEP) {
+                                        mtx_lock(&rkq->rkq_lock);
+                                        is_locked = rd_true;
+                                        goto retry; /* Next op */
+                                } else if (unlikely(res ==
+                                                    RD_KAFKA_OP_RES_YIELD)) {
+                                        if (can_q_contain_fetched_msgs)
+                                                rd_kafka_app_polled(
+                                                    rkq->rkq_rk);
+                                        /* Callback yielded, unroll */
+                                        return NULL;
+                                } else {
+                                        if (can_q_contain_fetched_msgs)
+                                                rd_kafka_app_polled(
+                                                    rkq->rkq_rk);
+                                        break; /* Proper op, handle below. */
+                                }
+                        }
+
+                        if (unlikely(rd_kafka_q_check_yield(rkq))) {
+                                if (is_locked)
+                                        mtx_unlock(&rkq->rkq_lock);
+                                if (can_q_contain_fetched_msgs)
+                                        rd_kafka_app_polled(rkq->rkq_rk);
+                                return NULL;
+                        }
+
+                        if (!is_locked)
+                                mtx_lock(&rkq->rkq_lock);
+
+                        if (cnd_timedwait_abs(&rkq->rkq_cond, &rkq->rkq_lock,
+                                              &timeout_tspec) != thrd_success) {
+                                mtx_unlock(&rkq->rkq_lock);
+                                if (can_q_contain_fetched_msgs)
+                                        rd_kafka_app_polled(rkq->rkq_rk);
+                                return NULL;
+                        }
+                }
+
+        } else {
+                /* Since the q_pop may block we need to release the parent
+                 * queue's lock. */
+                mtx_unlock(&rkq->rkq_lock);
+                rko = rd_kafka_q_pop_serve(fwdq, timeout_us, version, cb_type,
+                                           callback, opaque);
+                rd_kafka_q_destroy(fwdq);
+        }
+
+
+        return rko;
+}
+
+rd_kafka_op_t *
+rd_kafka_q_pop(rd_kafka_q_t *rkq, rd_ts_t timeout_us, int32_t version) {
+        return rd_kafka_q_pop_serve(rkq, timeout_us, version,
+                                    RD_KAFKA_Q_CB_RETURN, NULL, NULL);
+}
+
+
+/**
+ * Pop all available ops from a queue and call the provided
+ * callback for each op.
+ * `max_cnt` limits the number of ops served, 0 = no limit.
+ *
+ * Returns the number of ops served.
+ *
+ * Locality: any thread.
+ */
+int rd_kafka_q_serve(rd_kafka_q_t *rkq,
+                     int timeout_ms,
+                     int max_cnt,
+                     rd_kafka_q_cb_type_t cb_type,
+                     rd_kafka_q_serve_cb_t *callback,
+                     void *opaque) {
+        rd_kafka_t *rk = rkq->rkq_rk;
+        rd_kafka_op_t *rko;
+        rd_kafka_q_t localq;
+        rd_kafka_q_t *fwdq;
+        int cnt = 0;
+        struct timespec timeout_tspec;
+        const rd_bool_t can_q_contain_fetched_msgs =
+            rd_kafka_q_can_contain_fetched_msgs(rkq, RD_DONT_LOCK);
+
+        rd_dassert(cb_type);
+
+        mtx_lock(&rkq->rkq_lock);
+
+        rd_dassert(TAILQ_EMPTY(&rkq->rkq_q) || rkq->rkq_qlen > 0);
+        if ((fwdq = rd_kafka_q_fwd_get(rkq, 0))) {
+                int ret;
+                /* Since the q_pop may block we need to release the parent
+                 * queue's lock. */
+                mtx_unlock(&rkq->rkq_lock);
+                ret = rd_kafka_q_serve(fwdq, timeout_ms, max_cnt, cb_type,
+                                       callback, opaque);
+                rd_kafka_q_destroy(fwdq);
+                return ret;
+        }
+
+
+        rd_timeout_init_timespec(&timeout_tspec, timeout_ms);
+
+        if (can_q_contain_fetched_msgs)
+                rd_kafka_app_poll_start(rk, 0, timeout_ms);
+
+        /* Wait for op */
+        while (!(rko = TAILQ_FIRST(&rkq->rkq_q)) &&
+               !rd_kafka_q_check_yield(rkq) &&
+               cnd_timedwait_abs(&rkq->rkq_cond, &rkq->rkq_lock,
+                                 &timeout_tspec) == thrd_success)
+                ;
+
+        rd_kafka_q_mark_served(rkq);
+
+        if (!rko) {
+                mtx_unlock(&rkq->rkq_lock);
+                if (can_q_contain_fetched_msgs)
+                        rd_kafka_app_polled(rk);
+                return 0;
+        }
+
+        /* Move the first `max_cnt` ops. */
+        rd_kafka_q_init(&localq, rkq->rkq_rk);
+        rd_kafka_q_move_cnt(&localq, rkq, max_cnt == 0 ? -1 /*all*/ : max_cnt,
+                            0 /*no-locks*/);
+
+        mtx_unlock(&rkq->rkq_lock);
+
+        rd_kafka_yield_thread = 0;
+
+        /* Call callback for each op */
+        while ((rko = TAILQ_FIRST(&localq.rkq_q))) {
+                rd_kafka_op_res_t res;
+
+                rd_kafka_q_deq0(&localq, rko);
+                res = rd_kafka_op_handle(rk, &localq, rko, cb_type, opaque,
+                                         callback);
+                /* op must have been handled */
+                rd_kafka_assert(NULL, res != RD_KAFKA_OP_RES_PASS);
+                cnt++;
+
+                if (unlikely(res == RD_KAFKA_OP_RES_YIELD ||
+                             rd_kafka_yield_thread)) {
+                        /* Callback called rd_kafka_yield(), we must
+                         * stop our callback dispatching and put the
+                         * ops in localq back on the original queue head. */
+                        if (!TAILQ_EMPTY(&localq.rkq_q))
+                                rd_kafka_q_prepend(rkq, &localq);
+                        break;
+                }
+        }
+
+        if (can_q_contain_fetched_msgs)
+                rd_kafka_app_polled(rk);
+
+        rd_kafka_q_destroy_owner(&localq);
+
+        return cnt;
+}
+
+/**
+ * @brief Filter out and destroy outdated messages.
+ *
+ * @returns Returns the number of valid messages.
+ *
+ * @locality Any thread.
+ */
+static size_t
+rd_kafka_purge_outdated_messages(rd_kafka_toppar_t *rktp,
+                                 int32_t version,
+                                 rd_kafka_message_t **rkmessages,
+                                 size_t cnt,
+                                 struct rd_kafka_op_tailq *ctrl_msg_q) {
+        size_t valid_count = 0;
+        size_t i;
+        rd_kafka_op_t *rko, *next;
+
+        for (i = 0; i < cnt; i++) {
+                rko = rkmessages[i]->_private;
+                if (rko->rko_rktp == rktp &&
+                    rd_kafka_op_version_outdated(rko, version)) {
+                        /* This also destroys the corresponding rkmessage. */
+                        rd_kafka_op_destroy(rko);
+                } else if (i > valid_count) {
+                        rkmessages[valid_count++] = rkmessages[i];
+                } else {
+                        valid_count++;
+                }
+        }
+
+        /* Discard outdated control msgs ops */
+        next = TAILQ_FIRST(ctrl_msg_q);
+        while (next) {
+                rko  = next;
+                next = TAILQ_NEXT(rko, rko_link);
+                if (rko->rko_rktp == rktp &&
+                    rd_kafka_op_version_outdated(rko, version)) {
+                        TAILQ_REMOVE(ctrl_msg_q, rko, rko_link);
+                        rd_kafka_op_destroy(rko);
+                }
+        }
+
+        return valid_count;
+}
+
+
+/**
+ * Populate 'rkmessages' array with messages from 'rkq'.
+ * If 'auto_commit' is set, each message's offset will be committed
+ * to the offset store for that toppar.
+ *
+ * Returns the number of messages added.
+ */
+
+int rd_kafka_q_serve_rkmessages(rd_kafka_q_t *rkq,
+                                int timeout_ms,
+                                rd_kafka_message_t **rkmessages,
+                                size_t rkmessages_size) {
+        unsigned int cnt = 0;
+        TAILQ_HEAD(, rd_kafka_op_s) tmpq = TAILQ_HEAD_INITIALIZER(tmpq);
+        struct rd_kafka_op_tailq ctrl_msg_q =
+            TAILQ_HEAD_INITIALIZER(ctrl_msg_q);
+        rd_kafka_op_t *rko, *next;
+        rd_kafka_t *rk = rkq->rkq_rk;
+        rd_kafka_q_t *fwdq;
+        struct timespec timeout_tspec;
+        int i;
+
+        mtx_lock(&rkq->rkq_lock);
+        if ((fwdq = rd_kafka_q_fwd_get(rkq, 0))) {
+                /* Since the q_pop may block we need to release the parent
+                 * queue's lock. */
+                mtx_unlock(&rkq->rkq_lock);
+                cnt = rd_kafka_q_serve_rkmessages(fwdq, timeout_ms, rkmessages,
+                                                  rkmessages_size);
+                rd_kafka_q_destroy(fwdq);
+                return cnt;
+        }
+
+        mtx_unlock(&rkq->rkq_lock);
+
+        rd_timeout_init_timespec(&timeout_tspec, timeout_ms);
+
+        rd_kafka_app_poll_start(rk, 0, timeout_ms);
+
+        rd_kafka_yield_thread = 0;
+        while (cnt < rkmessages_size) {
+                rd_kafka_op_res_t res;
+
+                mtx_lock(&rkq->rkq_lock);
+
+                while (!(rko = TAILQ_FIRST(&rkq->rkq_q)) &&
+                       !rd_kafka_q_check_yield(rkq) &&
+                       cnd_timedwait_abs(&rkq->rkq_cond, &rkq->rkq_lock,
+                                         &timeout_tspec) == thrd_success)
+                        ;
+
+                rd_kafka_q_mark_served(rkq);
+
+                if (!rko) {
+                        mtx_unlock(&rkq->rkq_lock);
+                        break; /* Timed out */
+                }
+
+                rd_kafka_q_deq0(rkq, rko);
+
+                mtx_unlock(&rkq->rkq_lock);
+
+                if (unlikely(rko->rko_type == RD_KAFKA_OP_BARRIER)) {
+                        cnt = (unsigned int)rd_kafka_purge_outdated_messages(
+                            rko->rko_rktp, rko->rko_version, rkmessages, cnt,
+                            &ctrl_msg_q);
+                        rd_kafka_op_destroy(rko);
+                        continue;
+                }
+
+                if (rd_kafka_op_version_outdated(rko, 0)) {
+                        /* Outdated op, put on discard queue */
+                        TAILQ_INSERT_TAIL(&tmpq, rko, rko_link);
+                        continue;
+                }
+
+                /* Serve non-FETCH callbacks */
+                res =
+                    rd_kafka_poll_cb(rk, rkq, rko, RD_KAFKA_Q_CB_RETURN, NULL);
+                if (res == RD_KAFKA_OP_RES_KEEP ||
+                    res == RD_KAFKA_OP_RES_HANDLED) {
+                        /* Callback served, rko is destroyed (if HANDLED). */
+                        continue;
+                } else if (unlikely(res == RD_KAFKA_OP_RES_YIELD ||
+                                    rd_kafka_yield_thread)) {
+                        /* Yield. */
+                        break;
+                }
+                rd_dassert(res == RD_KAFKA_OP_RES_PASS);
+
+                /* If this is a control messages, don't return message to
+                 * application. Add it to a tmp queue from where we can store
+                 * the offset and destroy the op */
+                if (unlikely(rd_kafka_op_is_ctrl_msg(rko))) {
+                        TAILQ_INSERT_TAIL(&ctrl_msg_q, rko, rko_link);
+                        continue;
+                }
+
+                /* Get rkmessage from rko and append to array. */
+                rkmessages[cnt++] = rd_kafka_message_get(rko);
+        }
+
+        for (i = cnt - 1; i >= 0; i--) {
+                rko = (rd_kafka_op_t *)rkmessages[i]->_private;
+                rd_kafka_toppar_t *rktp = rko->rko_rktp;
+                int64_t offset          = rkmessages[i]->offset + 1;
+                if (unlikely(rktp && (rktp->rktp_app_pos.offset < offset)))
+                        rd_kafka_update_app_pos(
+                            rk, rktp,
+                            RD_KAFKA_FETCH_POS(
+                                offset,
+                                rd_kafka_message_leader_epoch(rkmessages[i])),
+                            RD_DO_LOCK);
+        }
+
+        /* Discard non-desired and already handled ops */
+        next = TAILQ_FIRST(&tmpq);
+        while (next) {
+                rko  = next;
+                next = TAILQ_NEXT(next, rko_link);
+                rd_kafka_op_destroy(rko);
+        }
+
+        /* Discard ctrl msgs */
+        next = TAILQ_FIRST(&ctrl_msg_q);
+        while (next) {
+                rko                     = next;
+                next                    = TAILQ_NEXT(next, rko_link);
+                rd_kafka_toppar_t *rktp = rko->rko_rktp;
+                int64_t offset = rko->rko_u.fetch.rkm.rkm_rkmessage.offset + 1;
+                if (rktp && (rktp->rktp_app_pos.offset < offset))
+                        rd_kafka_update_app_pos(
+                            rk, rktp,
+                            RD_KAFKA_FETCH_POS(
+                                offset,
+                                rd_kafka_message_leader_epoch(
+                                    &rko->rko_u.fetch.rkm.rkm_rkmessage)),
+                            RD_DO_LOCK);
+                rd_kafka_op_destroy(rko);
+        }
+
+        rd_kafka_app_polled(rk);
+
+        return cnt;
+}
+
+
+
+void rd_kafka_queue_destroy(rd_kafka_queue_t *rkqu) {
+        if (rkqu->rkqu_is_owner)
+                rd_kafka_q_destroy_owner(rkqu->rkqu_q);
+        else
+                rd_kafka_q_destroy(rkqu->rkqu_q);
+        rd_free(rkqu);
+}
+
+rd_kafka_queue_t *rd_kafka_queue_new0(rd_kafka_t *rk, rd_kafka_q_t *rkq) {
+        rd_kafka_queue_t *rkqu;
+
+        rkqu = rd_calloc(1, sizeof(*rkqu));
+
+        rkqu->rkqu_q = rkq;
+        rd_kafka_q_keep(rkq);
+
+        rkqu->rkqu_rk = rk;
+
+        return rkqu;
+}
+
+
+rd_kafka_queue_t *rd_kafka_queue_new(rd_kafka_t *rk) {
+        rd_kafka_q_t *rkq;
+        rd_kafka_queue_t *rkqu;
+
+        rkq  = rd_kafka_q_new(rk);
+        rkqu = rd_kafka_queue_new0(rk, rkq);
+        rd_kafka_q_destroy(rkq); /* Loose refcount from q_new, one is held
+                                  * by queue_new0 */
+        rkqu->rkqu_is_owner = 1;
+        return rkqu;
+}
+
+
+rd_kafka_queue_t *rd_kafka_queue_get_main(rd_kafka_t *rk) {
+        return rd_kafka_queue_new0(rk, rk->rk_rep);
+}
+
+
+rd_kafka_queue_t *rd_kafka_queue_get_consumer(rd_kafka_t *rk) {
+        if (!rk->rk_cgrp)
+                return NULL;
+        return rd_kafka_queue_new0(rk, rk->rk_cgrp->rkcg_q);
+}
+
+rd_kafka_queue_t *rd_kafka_queue_get_partition(rd_kafka_t *rk,
+                                               const char *topic,
+                                               int32_t partition) {
+        rd_kafka_toppar_t *rktp;
+        rd_kafka_queue_t *result;
+
+        if (rk->rk_type == RD_KAFKA_PRODUCER)
+                return NULL;
+
+        rktp = rd_kafka_toppar_get2(rk, topic, partition, 0, /* no ua_on_miss */
+                                    1 /* create_on_miss */);
+
+        if (!rktp)
+                return NULL;
+
+        result = rd_kafka_queue_new0(rk, rktp->rktp_fetchq);
+        rd_kafka_toppar_destroy(rktp);
+
+        return result;
+}
+
+rd_kafka_queue_t *rd_kafka_queue_get_background(rd_kafka_t *rk) {
+        rd_kafka_queue_t *rkqu;
+
+        rd_kafka_wrlock(rk);
+        if (!rk->rk_background.q) {
+                char errstr[256];
+
+                if (rd_kafka_background_thread_create(rk, errstr,
+                                                      sizeof(errstr))) {
+                        rd_kafka_log(rk, LOG_ERR, "BACKGROUND",
+                                     "Failed to create background thread: %s",
+                                     errstr);
+                        rd_kafka_wrunlock(rk);
+                        return NULL;
+                }
+        }
+
+        rkqu = rd_kafka_queue_new0(rk, rk->rk_background.q);
+        rd_kafka_wrunlock(rk);
+        return rkqu;
+}
+
+
+rd_kafka_resp_err_t rd_kafka_set_log_queue(rd_kafka_t *rk,
+                                           rd_kafka_queue_t *rkqu) {
+        rd_kafka_q_t *rkq;
+
+        if (!rk->rk_logq)
+                return RD_KAFKA_RESP_ERR__NOT_CONFIGURED;
+
+        if (!rkqu)
+                rkq = rk->rk_rep;
+        else
+                rkq = rkqu->rkqu_q;
+        rd_kafka_q_fwd_set(rk->rk_logq, rkq);
+        return RD_KAFKA_RESP_ERR_NO_ERROR;
+}
+
+void rd_kafka_queue_forward(rd_kafka_queue_t *src, rd_kafka_queue_t *dst) {
+        rd_kafka_q_fwd_set0(src->rkqu_q, dst ? dst->rkqu_q : NULL,
+                            1, /* do_lock */
+                            1 /* fwd_app */);
+}
+
+
+size_t rd_kafka_queue_length(rd_kafka_queue_t *rkqu) {
+        return (size_t)rd_kafka_q_len(rkqu->rkqu_q);
+}
+
+/**
+ * @brief Enable or disable(fd==-1) fd-based wake-ups for queue
+ */
+void rd_kafka_q_io_event_enable(rd_kafka_q_t *rkq,
+                                rd_socket_t fd,
+                                const void *payload,
+                                size_t size) {
+        struct rd_kafka_q_io *qio = NULL;
+
+        if (fd != -1) {
+                qio                  = rd_malloc(sizeof(*qio) + size);
+                qio->fd              = fd;
+                qio->size            = size;
+                qio->payload         = (void *)(qio + 1);
+                qio->sent            = rd_false;
+                qio->event_cb        = NULL;
+                qio->event_cb_opaque = NULL;
+                memcpy(qio->payload, payload, size);
+        }
+
+        mtx_lock(&rkq->rkq_lock);
+        if (rkq->rkq_qio) {
+                rd_free(rkq->rkq_qio);
+                rkq->rkq_qio = NULL;
+        }
+
+        if (fd != -1) {
+                rkq->rkq_qio = qio;
+        }
+
+        mtx_unlock(&rkq->rkq_lock);
+}
+
+void rd_kafka_queue_io_event_enable(rd_kafka_queue_t *rkqu,
+                                    int fd,
+                                    const void *payload,
+                                    size_t size) {
+        rd_kafka_q_io_event_enable(rkqu->rkqu_q, fd, payload, size);
+}
+
+
+void rd_kafka_queue_yield(rd_kafka_queue_t *rkqu) {
+        rd_kafka_q_yield(rkqu->rkqu_q);
+}
+
+
+/**
+ * @brief Enable or disable(event_cb==NULL) callback-based wake-ups for queue
+ */
+void rd_kafka_q_cb_event_enable(rd_kafka_q_t *rkq,
+                                void (*event_cb)(rd_kafka_t *rk, void *opaque),
+                                void *opaque) {
+        struct rd_kafka_q_io *qio = NULL;
+
+        if (event_cb) {
+                qio                  = rd_malloc(sizeof(*qio));
+                qio->fd              = -1;
+                qio->size            = 0;
+                qio->payload         = NULL;
+                qio->event_cb        = event_cb;
+                qio->event_cb_opaque = opaque;
+        }
+
+        mtx_lock(&rkq->rkq_lock);
+        if (rkq->rkq_qio) {
+                rd_free(rkq->rkq_qio);
+                rkq->rkq_qio = NULL;
+        }
+
+        if (event_cb) {
+                rkq->rkq_qio = qio;
+        }
+
+        mtx_unlock(&rkq->rkq_lock);
+}
+
+void rd_kafka_queue_cb_event_enable(rd_kafka_queue_t *rkqu,
+                                    void (*event_cb)(rd_kafka_t *rk,
+                                                     void *opaque),
+                                    void *opaque) {
+        rd_kafka_q_cb_event_enable(rkqu->rkqu_q, event_cb, opaque);
+}
+
+
+/**
+ * Helper: wait for single op on 'rkq', and return its error,
+ * or .._TIMED_OUT on timeout.
+ */
+rd_kafka_resp_err_t rd_kafka_q_wait_result(rd_kafka_q_t *rkq, int timeout_ms) {
+        rd_kafka_op_t *rko;
+        rd_kafka_resp_err_t err;
+
+        rko = rd_kafka_q_pop(rkq, rd_timeout_us(timeout_ms), 0);
+        if (!rko)
+                err = RD_KAFKA_RESP_ERR__TIMED_OUT;
+        else {
+                err = rko->rko_err;
+                rd_kafka_op_destroy(rko);
+        }
+
+        return err;
+}
+
+
+/**
+ * Apply \p callback on each op in queue.
+ * If the callback wishes to remove the rko it must do so using
+ * using rd_kafka_op_deq0().
+ *
+ * @returns the sum of \p callback() return values.
+ * @remark rkq will be locked, callers should take care not to
+ *         interact with \p rkq through other means from the callback to avoid
+ *         deadlocks.
+ */
+int rd_kafka_q_apply(rd_kafka_q_t *rkq,
+                     int (*callback)(rd_kafka_q_t *rkq,
+                                     rd_kafka_op_t *rko,
+                                     void *opaque),
+                     void *opaque) {
+        rd_kafka_op_t *rko, *next;
+        rd_kafka_q_t *fwdq;
+        int cnt = 0;
+
+        mtx_lock(&rkq->rkq_lock);
+        if ((fwdq = rd_kafka_q_fwd_get(rkq, 0))) {
+                mtx_unlock(&rkq->rkq_lock);
+                cnt = rd_kafka_q_apply(fwdq, callback, opaque);
+                rd_kafka_q_destroy(fwdq);
+                return cnt;
+        }
+
+        next = TAILQ_FIRST(&rkq->rkq_q);
+        while ((rko = next)) {
+                next = TAILQ_NEXT(next, rko_link);
+                cnt += callback(rkq, rko, opaque);
+        }
+
+        rd_kafka_q_mark_served(rkq);
+
+        mtx_unlock(&rkq->rkq_lock);
+
+        return cnt;
+}
+
+/**
+ * @brief Convert relative to absolute offsets and also purge any messages
+ *        that are older than \p min_offset.
+ * @remark Error ops with ERR__NOT_IMPLEMENTED will not be purged since
+ *         they are used to indicate unknnown compression codecs and compressed
+ *         messagesets may have a starting offset lower than what we requested.
+ * @remark \p rkq locking is not performed (caller's responsibility)
+ * @remark Must NOT be used on fwdq.
+ */
+void rd_kafka_q_fix_offsets(rd_kafka_q_t *rkq,
+                            int64_t min_offset,
+                            int64_t base_offset) {
+        rd_kafka_op_t *rko, *next;
+        int adj_len      = 0;
+        int64_t adj_size = 0;
+
+        rd_kafka_assert(NULL, !rkq->rkq_fwdq);
+
+        next = TAILQ_FIRST(&rkq->rkq_q);
+        while ((rko = next)) {
+                next = TAILQ_NEXT(next, rko_link);
+
+                if (unlikely(rko->rko_type != RD_KAFKA_OP_FETCH))
+                        continue;
+
+                rko->rko_u.fetch.rkm.rkm_offset += base_offset;
+
+                if (rko->rko_u.fetch.rkm.rkm_offset < min_offset &&
+                    rko->rko_err != RD_KAFKA_RESP_ERR__NOT_IMPLEMENTED) {
+                        adj_len++;
+                        adj_size += rko->rko_len;
+                        TAILQ_REMOVE(&rkq->rkq_q, rko, rko_link);
+                        rd_kafka_op_destroy(rko);
+                        continue;
+                }
+        }
+
+
+        rkq->rkq_qlen -= adj_len;
+        rkq->rkq_qsize -= adj_size;
+}
+
+
+/**
+ * @brief Print information and contents of queue
+ */
+void rd_kafka_q_dump(FILE *fp, rd_kafka_q_t *rkq) {
+        mtx_lock(&rkq->rkq_lock);
+        fprintf(fp,
+                "Queue %p \"%s\" (refcnt %d, flags 0x%x, %d ops, "
+                "%" PRId64 " bytes)\n",
+                rkq, rkq->rkq_name, rkq->rkq_refcnt, rkq->rkq_flags,
+                rkq->rkq_qlen, rkq->rkq_qsize);
+
+        if (rkq->rkq_qio)
+                fprintf(fp, " QIO fd %d\n", (int)rkq->rkq_qio->fd);
+        if (rkq->rkq_serve)
+                fprintf(fp, " Serve callback %p, opaque %p\n", rkq->rkq_serve,
+                        rkq->rkq_opaque);
+
+        if (rkq->rkq_fwdq) {
+                fprintf(fp, " Forwarded ->\n");
+                rd_kafka_q_dump(fp, rkq->rkq_fwdq);
+        } else {
+                rd_kafka_op_t *rko;
+
+                if (!TAILQ_EMPTY(&rkq->rkq_q))
+                        fprintf(fp, " Queued ops:\n");
+                TAILQ_FOREACH(rko, &rkq->rkq_q, rko_link) {
+                        fprintf(fp,
+                                "  %p %s (v%" PRId32
+                                ", flags 0x%x, "
+                                "prio %d, len %" PRId32
+                                ", source %s, "
+                                "replyq %p)\n",
+                                rko, rd_kafka_op2str(rko->rko_type),
+                                rko->rko_version, rko->rko_flags, rko->rko_prio,
+                                rko->rko_len,
+#if ENABLE_DEVEL
+                                rko->rko_source
+#else
+                                "-"
+#endif
+                                ,
+                                rko->rko_replyq.q);
+                }
+        }
+
+        mtx_unlock(&rkq->rkq_lock);
+}
+
+
+void rd_kafka_enq_once_trigger_destroy(void *ptr) {
+        rd_kafka_enq_once_t *eonce = ptr;
+
+        rd_kafka_enq_once_trigger(eonce, RD_KAFKA_RESP_ERR__DESTROY, "destroy");
+}