/**
* Commit a list of offsets asynchronously. Response will be queued on 'replyq'.
* Optional 'op_cb' will be set on requesting op.
* 'opaque' will be set as 'rko_opaque'.
*/
rd_kafka_resp_err_t
rd_kafka_commit0 (rd_kafka_t *rk,
const rd_kafka_topic_partition_list_t *offsets,
rd_kafka_q_t *replyq, void (*op_cb) (rd_kafka_t *,
rd_kafka_op_t *)) {
rd_kafka_cgrp_t *rkcg;
rd_kafka_op_t *rko;
if (!(rkcg = rd_kafka_cgrp_get(rk)))
return RD_KAFKA_RESP_ERR__UNKNOWN_GROUP;
rko = rd_kafka_op_new(RD_KAFKA_OP_OFFSET_COMMIT);
rko->rko_op_cb = op_cb;
rko->rko_replyq = replyq;
if (replyq)
rd_kafka_q_keep(rko->rko_replyq);
if (offsets)
rd_kafka_op_payload_set(
rko, rd_kafka_topic_partition_list_copy(offsets),
(void *)rd_kafka_topic_partition_list_destroy);
rd_kafka_q_enq(&rkcg->rkcg_ops, rko);
return RD_KAFKA_RESP_ERR_NO_ERROR;
}
static rd_kafka_resp_err_t
rd_kafka_offset_broker_commit (rd_kafka_toppar_t *rktp) {
rd_kafka_topic_partition_list_t *offsets;
rd_kafka_topic_partition_t *rktpar;
rd_kafka_assert(rktp->rktp_rkt->rkt_rk, rktp->rktp_cgrp != NULL);
rd_kafka_assert(rktp->rktp_rkt->rkt_rk,
rktp->rktp_flags & RD_KAFKA_TOPPAR_F_OFFSET_STORE);
rktp->rktp_committing_offset = rktp->rktp_stored_offset;
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);
rktpar->offset = rktp->rktp_committing_offset;
rd_kafka_dbg(rktp->rktp_rkt->rkt_rk, CGRP, "OFFSETCMT",
"%.*s [%"PRId32"]: committing offset %"PRId64,
RD_KAFKAP_STR_PR(rktp->rktp_rkt->rkt_topic),
rktp->rktp_partition, rktp->rktp_committing_offset);
rd_kafka_q_keep(&rktp->rktp_ops);
rd_kafka_commit0(rktp->rktp_rkt->rkt_rk, offsets,
&rktp->rktp_ops, rd_kafka_offset_broker_commit_op_cb);
rd_kafka_topic_partition_list_destroy(offsets);
return RD_KAFKA_RESP_ERR__IN_PROGRESS;
}
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;
}
/**
* Pop an op from a queue.
*
* Locality: any thread.
*/
rd_kafka_op_t *rd_kafka_q_pop (rd_kafka_q_t *rkq, int timeout_ms,
int32_t version) {
rd_kafka_op_t *rko;
if (timeout_ms == RD_POLL_INFINITE)
timeout_ms = INT_MAX;
mtx_lock(&rkq->rkq_lock);
if (!rkq->rkq_fwdq) {
do {
/* Filter out outdated ops */
while ((rko = TAILQ_FIRST(&rkq->rkq_q)) &&
!(rko = rd_kafka_op_filter(rkq, rko)))
;
if (rko) {
/* Proper versioned op */
rd_kafka_q_deq0(rkq, rko);
break;
}
/* No op, wait for one */
rd_ts_t pre = rd_clock();
if (cnd_timedwait_ms(&rkq->rkq_cond,
&rkq->rkq_lock,
timeout_ms) ==
thrd_timedout) {
mtx_unlock(&rkq->rkq_lock);
return NULL;
}
/* Remove spent time */
timeout_ms -= (int) (rd_clock()-pre) / 1000;
if (timeout_ms < 0)
timeout_ms = RD_POLL_NOWAIT;
} while (timeout_ms != RD_POLL_NOWAIT);
mtx_unlock(&rkq->rkq_lock);
} else {
rd_kafka_q_t *fwdq = rkq->rkq_fwdq;
rd_kafka_q_keep(fwdq);
/* 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(fwdq, timeout_ms, version);
rd_kafka_q_destroy(fwdq);
}
return rko;
}
/**
* Send request to queue, wait for response.
*/
rd_kafka_op_t *rd_kafka_op_req0 (rd_kafka_q_t *destq,
rd_kafka_q_t *recvq,
rd_kafka_op_t *rko,
int timeout_ms) {
rd_kafka_op_t *reply;
/* Indicate to destination where to send reply. */
rko->rko_replyq = recvq;
if (recvq)
rd_kafka_q_keep(rko->rko_replyq);
/* Enqueue op */
rd_kafka_q_enq(destq, rko);
/* Wait for reply */
reply = rd_kafka_q_pop(recvq, timeout_ms, 0);
/* May be NULL for timeout */
return reply;
}
/**
* 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) {
if (do_lock)
mtx_lock(&srcq->rkq_lock);
if (srcq->rkq_fwdq) {
rd_kafka_q_destroy(srcq->rkq_fwdq);
srcq->rkq_fwdq = NULL;
}
if (destq) {
rd_kafka_q_keep(destq);
srcq->rkq_fwdq = destq;
/* If rkq has ops in queue, append them to fwdq's queue.
* This is an irreversible operation. */
if (srcq->rkq_qlen > 0)
rd_kafka_q_concat(destq, srcq);
}
if (do_lock)
mtx_unlock(&srcq->rkq_lock);
}
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);
rd_kafka_op_t *rko, *next;
rd_kafka_t *rk = rkq->rkq_rk;
mtx_lock(&rkq->rkq_lock);
if (rkq->rkq_fwdq) {
rd_kafka_q_t *fwdq = rkq->rkq_fwdq;
rd_kafka_q_keep(fwdq);
/* 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);
while (cnt < rkmessages_size) {
mtx_lock(&rkq->rkq_lock);
while (!(rko = TAILQ_FIRST(&rkq->rkq_q))) {
if (cnd_timedwait_ms(&rkq->rkq_cond, &rkq->rkq_lock,
timeout_ms) == thrd_timedout)
break;
}
if (!rko) {
mtx_unlock(&rkq->rkq_lock);
break; /* Timed out */
}
rd_kafka_q_deq0(rkq, rko);
mtx_unlock(&rkq->rkq_lock);
if (rd_kafka_op_version_outdated(rko, 0)) {
/* Outdated op, put on discard queue */
TAILQ_INSERT_TAIL(&tmpq, rko, rko_link);
continue;
}
/* Serve callbacks */
if (rd_kafka_poll_cb(rk, rko, _Q_CB_CONSUMER, NULL)) {
/* Callback served, rko is done, put on discard queue */
TAILQ_INSERT_TAIL(&tmpq, rko, rko_link);
continue;
}
/* Auto-commit offset, if enabled. */
if (!rko->rko_err && rko->rko_type == RD_KAFKA_OP_FETCH) {
rd_kafka_toppar_t *rktp;
rktp = rd_kafka_toppar_s2i(rko->rko_rktp);
rd_kafka_toppar_lock(rktp);
rktp->rktp_app_offset = rko->rko_u.fetch.rkm.rkm_offset+1;
if (rktp->rktp_cgrp &&
rk->rk_conf.enable_auto_offset_store)
rd_kafka_offset_store0(rktp,
rktp->rktp_app_offset,
0/* no lock */);
rd_kafka_toppar_unlock(rktp);
}
/* Get rkmessage from rko and append to array. */
rkmessages[cnt++] = rd_kafka_message_get(rko);
}
/* 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);
}
return cnt;
}
/**
* 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, int cb_type,
int (*callback) (rd_kafka_t *rk, rd_kafka_op_t *rko,
int cb_type, void *opaque),
void *opaque) {
rd_kafka_t *rk = rkq->rkq_rk;
rd_kafka_op_t *rko;
rd_kafka_q_t localq;
int cnt = 0;
int handled = 0;
mtx_lock(&rkq->rkq_lock);
rd_dassert(TAILQ_EMPTY(&rkq->rkq_q) || rkq->rkq_qlen > 0);
if (rkq->rkq_fwdq) {
rd_kafka_q_t *fwdq = rkq->rkq_fwdq;
int ret;
rd_kafka_q_keep(fwdq);
/* 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;
}
if (timeout_ms == RD_POLL_INFINITE)
timeout_ms = INT_MAX;
/* Wait for op */
while (!(rko = TAILQ_FIRST(&rkq->rkq_q)) && timeout_ms != 0) {
if (cnd_timedwait_ms(&rkq->rkq_cond,
&rkq->rkq_lock,
timeout_ms) != thrd_success)
break;
timeout_ms = 0;
}
if (!rko) {
mtx_unlock(&rkq->rkq_lock);
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))) {
handled += callback(rk, rko, cb_type, opaque);
rd_kafka_q_deq0(&localq, rko);
rd_kafka_op_destroy(rko);
cnt++;
if (unlikely(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;
}
}
/* Make sure no op was left unhandled. i.e.,
* a consumer op ended up on the global queue. */
rd_kafka_assert(NULL, handled == cnt);
rd_kafka_q_destroy(&localq);
return cnt;
}
/**
* 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, int timeout_ms,
int32_t version, int cb_type,
int (*callback) (rd_kafka_t *rk,
rd_kafka_op_t *rko,
int cb_type,
void *opaque),
void *opaque) {
rd_kafka_op_t *rko;
if (timeout_ms == RD_POLL_INFINITE)
timeout_ms = INT_MAX;
mtx_lock(&rkq->rkq_lock);
if (!rkq->rkq_fwdq) {
do {
/* Filter out outdated ops */
while ((rko = TAILQ_FIRST(&rkq->rkq_q)) &&
!(rko = rd_kafka_op_filter(rkq, rko, version)))
;
if (rko) {
int handled;
/* Proper versioned op */
rd_kafka_q_deq0(rkq, rko);
/* Ops with callbacks are considered handled
* and we move on to the next op, if any.
* Ops w/o callbacks are returned immediately */
if (callback) {
handled = callback(rkq->rkq_rk, rko,
cb_type, opaque);
if (handled) {
rd_kafka_op_destroy(rko);
rko = NULL;
}
} else
handled = 0;
if (!handled)
break;
}
/* No op, wait for one */
rd_ts_t pre = rd_clock();
if (cnd_timedwait_ms(&rkq->rkq_cond,
&rkq->rkq_lock,
timeout_ms) ==
thrd_timedout) {
mtx_unlock(&rkq->rkq_lock);
return NULL;
}
/* Remove spent time */
timeout_ms -= (int) (rd_clock()-pre) / 1000;
if (timeout_ms < 0)
timeout_ms = RD_POLL_NOWAIT;
} while (timeout_ms != RD_POLL_NOWAIT);
mtx_unlock(&rkq->rkq_lock);
} else {
rd_kafka_q_t *fwdq = rkq->rkq_fwdq;
rd_kafka_q_keep(fwdq);
/* 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_ms, version,
cb_type, callback, opaque);
rd_kafka_q_destroy(fwdq);
}
return rko;
}
