[dpdk-dev] [PATCH v3 4/5] testpmd: handle all rxqs in rss setup

[Nélio Laranjeiro]

On Tue, Jun 14, 2016 at 07:08:05PM -0400, Zhihong Wang wrote:
> This patch removes constraints in rxq handling when multiqueue is enabled
> to handle all the rxqs.
> 
> Current testpmd forces a dedicated core for each rxq, some rxqs may be
> ignored when core number is less than rxq number, and that causes confusion
> and inconvenience.
> 
> One example: One Red Hat engineer was doing multiqueue test, there're 2
> ports in guest each with 4 queues, and testpmd was used as the forwarding
> engine in guest, as usual he used 1 core for forwarding, as a results he
> only saw traffic from port 0 queue 0 to port 1 queue 0, then a lot of
> emails and quite some time are spent to root cause it, and of course it's
> caused by this unreasonable testpmd behavior.  
> 
> Moreover, even if we understand this behavior, if we want to test the
> above case, we still need 8 cores for a single guest to poll all the
> rxqs, obviously this is too expensive.
> 
> We met quite a lot cases like this, one recent example:
> http://openvswitch.org/pipermail/dev/2016-June/072110.html
> 
> 
> Signed-off-by: Zhihong Wang <zhihong.wang at intel.com>
> ---
>  app/test-pmd/config.c | 8 +-------
>  1 file changed, 1 insertion(+), 7 deletions(-)
> 
> diff --git a/app/test-pmd/config.c b/app/test-pmd/config.c
> index ede7c78..4719a08 100644
> --- a/app/test-pmd/config.c
> +++ b/app/test-pmd/config.c
> @@ -1199,19 +1199,13 @@ rss_fwd_config_setup(void)
>  	cur_fwd_config.nb_fwd_ports = nb_fwd_ports;
>  	cur_fwd_config.nb_fwd_streams =
>  		(streamid_t) (nb_q * cur_fwd_config.nb_fwd_ports);
> -	if (cur_fwd_config.nb_fwd_streams > cur_fwd_config.nb_fwd_lcores)
> -		cur_fwd_config.nb_fwd_streams =
> -			(streamid_t)cur_fwd_config.nb_fwd_lcores;
> -	else
> -		cur_fwd_config.nb_fwd_lcores =
> -			(lcoreid_t)cur_fwd_config.nb_fwd_streams;
>  
>  	/* reinitialize forwarding streams */
>  	init_fwd_streams();
>  
>  	setup_fwd_config_of_each_lcore(&cur_fwd_config);
>  	rxp = 0; rxq = 0;
> -	for (lc_id = 0; lc_id < cur_fwd_config.nb_fwd_lcores; lc_id++) {
> +	for (lc_id = 0; lc_id < cur_fwd_config.nb_fwd_streams; lc_id++) {
>  		struct fwd_stream *fs;
>  
>  		fs = fwd_streams[lc_id];
> -- 
> 2.5.0

Hi Zhihong,

It seems this commits introduce a bug in pkt_burst_transmit(), this only
occurs when the number of cores present in the coremask is greater than
the number of queues i.e. coremask=0xffe --txq=4 --rxq=4.

  Port 0 Link Up - speed 40000 Mbps - full-duplex
  Port 1 Link Up - speed 40000 Mbps - full-duplex
  Done
  testpmd> start tx_first
    io packet forwarding - CRC stripping disabled - packets/burst=64
    nb forwarding cores=10 - nb forwarding ports=2
    RX queues=4 - RX desc=256 - RX free threshold=0
    RX threshold registers: pthresh=0 hthresh=0 wthresh=0
    TX queues=4 - TX desc=256 - TX free threshold=0
    TX threshold registers: pthresh=0 hthresh=0 wthresh=0
    TX RS bit threshold=0 - TXQ flags=0x0
  Segmentation fault (core dumped)


If I start testpmd with a coremask with at most as many cores as queues,
everything works well (i.e. coremask=0xff0, or 0xf00).

Are you able to reproduce the same issue?
Note: It only occurs on dpdk/master branch (commit f2bb7ae1d204).

Regards,

-- 
Nélio Laranjeiro
6WIND