[dpdk-dev] [PATCH v6 08/10] examples/l2fwd-event: add eventdev main loop
Varghese, Vipin
vipin.varghese at intel.com
Tue Oct 22 05:13:03 CEST 2019
Hi Pavan,
snipped
> >> Add event dev main loop based on enabled l2fwd options and
> >eventdev
> >> capabilities.
> >>
> >> Signed-off-by: Pavan Nikhilesh <pbhagavatula at marvell.com>
> >> ---
> >> + if (rsrc->event_mode) {
> >> + port_conf.rxmode.mq_mode = ETH_MQ_RX_RSS;
> >> + port_conf.rx_adv_conf.rss_conf.rss_key = NULL;
> >> + port_conf.rx_adv_conf.rss_conf.rss_hf = ETH_RSS_IP;
> >> + }
> >Question, is RSS hash configured for generating flow id for Eventdev?
> >As my understanding. RSS for single RX port-queue pair does not require
> >the same.
Thanks for the confirmation.
> >snipped
>
> In case of SW event device i.e. vdev=event_sw0 the software Rx adapter requires
> mbuf::rss:hash to distribute packets else it has to calculate rss.
>
> @see lib/librte_eventdev/rte_event_eth_rx_adapter.c +817 '
> rss = do_rss ?
> rxa_do_softrss(m, rx_adapter->rss_key_be) :
> m->hash.rss;
> '
>
> >> + if (is_master && timer_period > 0) {
> >> + cur_tsc = rte_rdtsc();
> >> + diff_tsc = cur_tsc - prev_tsc;
> >> +
> >> + /* advance the timer */
> >> + timer_tsc += diff_tsc;
> >> +
> >> + /* if timer has reached its timeout */
> >> + if (unlikely(timer_tsc >= timer_period)) {
> >> + print_stats(rsrc);
> >> + /* reset the timer */
> >> + timer_tsc = 0;
> >> + }
> >> + prev_tsc = cur_tsc;
> >> + }
> >Is it possible to move the print_stats to service core, as 'CALL_MASTER'
> >is enabled in remote_launch making this a potential worker?
> >
>
> Since not every eventdevice requires Service core user might not pass service
> core mask.
> Instead we could do SKIP_MASTER and prints stats here.
Thanks
>
> >> +
> >> + /* Read packet from eventdev */
> >> + if (!rte_event_dequeue_burst(event_d_id, port_id,
> >&ev, 1, 0))
> >> + continue;
> >Is not this unlikely `nb_burst == 0`
> >
>
> Not necessarily refer
> https://mails.dpdk.org/archives/dev/2018-July/108610.html
Thanks for the article and links it is helpful. Following the article suggestion, should not the code be enhanced for execute code rather than continue?
snipped
> >> + /* Read packet from eventdev */
> >> + nb_rx = rte_event_dequeue_burst(event_d_id,
> >port_id, ev,
> >> + deq_len, 0);
> >> + if (nb_rx == 0)
> >Can we use `unlikely`?
>
> Not necessarily refer
> https://mails.dpdk.org/archives/dev/2018-July/108610.html
Same as above.
>
> >> + continue;
> >> +
> >> + for (i = 0; i < nb_rx; i++) {
> >> + l2fwd_event_fwd(rsrc, &ev[i], tx_q_id,
> >timer_period,
> >> + flags);
> >> + }
> >> +
> >> + if (flags & L2FWD_EVENT_TX_ENQ) {
> >> + nb_tx =
> >rte_event_enqueue_burst(event_d_id, port_id,
> >> + ev, nb_rx);
> >> + while (nb_tx < nb_rx && !rsrc->force_quit)
> >> + nb_tx +=
> >> rte_event_enqueue_burst(event_d_id,
> >> + port_id, ev + nb_tx,
> >> + nb_rx - nb_tx);
> >Can we use `continue` as we do not transmit from the same worker int
> >his case?
>
> I'm not sure I follow what you meant here. We are trying to transmit the work
> present on the worker till we succeed, if we do a continue then we would loose
> the untransmitted packets.
Maybe I mistook the L2FWD against ETHDEV_EVENT flags. As per my current understanding L2FWD_EVENT_TX_ENQ uses extra queue stage for single-event-enqueue for TX adapter, while L2FWD_EVENT_TX_DIRECT allows the worker to do transmit via port. May be this is wrong expectation.
>
> @see examples/eventdev_pipeline/pipeline_worker_generic.c +109
I will check the same and get back as required.
>
> >> + }
> >> +
> >snipped
More information about the dev
mailing list