Re: [dpdk-dev] [PATCH v6 2/2] testpmd: add mode 4 support v6

[Thomas Monjalon <thomas.monjalon@6wind.com>]

2014-11-26 13:00, Wodkowski, PawelX:
> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Thomas Monjalon
> > 2014-11-26 11:17, Michal Jastrzebski:
> > > From: Pawel Wodkowski <pawelx.wodkowski@intel.com>
> > > --- a/app/test-pmd/csumonly.c
> > > +++ b/app/test-pmd/csumonly.c
> > > @@ -254,8 +254,17 @@ pkt_burst_checksum_forward(struct fwd_stream *fs)
> > >  	 */
> > >  	nb_rx = rte_eth_rx_burst(fs->rx_port, fs->rx_queue, pkts_burst,
> > >  				 nb_pkt_per_burst);
> > > +#ifndef RTE_LIBRTE_PMD_BOND
> > >  	if (unlikely(nb_rx == 0))
> > >  		return;
> > > +#else
> > > +	if (unlikely(nb_rx == 0 && (fs->forward_timeout == 0 ||
> > > +			fs->next_forward_time > rte_rdtsc())))
> > > +		return;
> > > +
> > > +	if (fs->forward_timeout != 0)
> > > +		fs->next_forward_time = rte_rdtsc() + fs->forward_timeout;
> > > +#endif
> > 
> > I don't understand why you need to make such change for bonding,
> > and there is no comment to explain.
> > Bonding should be a PMD like any other and shouldn't require such change.
> > I don't know mode 4 but it seems there is a design problem here.
> > 
> 
> It is an implication of requirement that was formed on beginning of bonding 
> implementation - bonded interface should be transparent to user app. But this
> requirement in is in collision with mode 4. It need to periodically receive and 
> transmit frames (LACP and marker) that are not passed to user app but 
> processed/produced in background. If this will not happen in at least 10 times
> per second mode 4 will not work.
> 
> Most of (all?) user applications do RX/TX more often than 10 times per second, 
> so this will have neglectable impact to those apps (it will have to check this 
> 100ms maximum interval of rx/tx as I did in code you pointed).
> 
> We had discussed all options with Declan and Bruce, and this seems to be the
> most transparent way to implement mode 4 without using any kind of locking
> inside library.

So you agree there is a design problem and you were initially trying to push it
without raising the problem in the hope that nobody will see it?
It's really not the good way to work in an Open Source project.

Is there any comment in the API to explain this new constraint?
Do you think we can change how Rx/Tx works in DPDK to integrate this feature?

Actually, I think these bonding features should be implemented in a layer on
top of DPDK. It's not the DPDK responsibility to make some protocol processing.
Bonding was integrated with the promise that it's transparent and really close
to the hardware ports.

Today I see we clearly need a discussion to know what should be implemented
in DPDK. Which protocol layer is the limit?
I explained my point of view but the decision belongs to the whole community.

-- 
Thomas