[dpdk-stable] [dpdk-dev] [PATCH v2] net/failsafe: fix source port ID in Rx packets

[Thomas Monjalon]

18/04/2019 18:46, Adrien Mazarguil:
> On Thu, Apr 18, 2019 at 05:51:18PM +0200, Thomas Monjalon wrote:
> > 18/04/2019 17:39, Thomas Monjalon:
> > > 18/04/2019 17:32, Adrien Mazarguil:
> > > > When passed to the application, Rx packets retain the port ID value
> > > > originally set by slave devices. Unfortunately these IDs have no meaning to
> > > > applications, which are typically unaware of their existence.
> > > > 
> > > > This confuses those caring about the source port field in mbufs (m->port)
> > > > which experience issues ranging from traffic drop to crashes.
> > [...]
> > > > +/*
> > > > + * Override source port in Rx packets.
> > > > + *
> > > > + * Make Rx packets originate from this PMD instance instead of one of its
> > > > + * slaves. This is mandatory to avoid breaking applications.
> > > > + */
> <snip>
> > > "slave" is a wording from bonding.
> > > In failsafe, it is sub-device, isn't it?
> 
> I don't mind, although grep shows a couple of comments talking about slaves
> already. Either way I think it fits as those are failsafe's pets, as in
> failsafe does whatever it wants to them and they don't have a say :)
> 
> Does it warrant a v3?

Yes please, except if Ferruh is already doing the change on apply.

> > > > +static void
> > > > +failsafe_rx_set_port(struct rte_mbuf **rx_pkts, uint16_t nb_pkts, uint16_t port)
> > > > +{
> > > > +	unsigned int i;
> > > > +
> > > > +	for (i = 0; i != nb_pkts; ++i)
> > > > +		rx_pkts[i]->port = port;
> > > > +}
> > > > +
> > > >  uint16_t
> > > >  failsafe_rx_burst(void *queue,
> > > >  		  struct rte_mbuf **rx_pkts,
> > > > @@ -87,6 +102,9 @@ failsafe_rx_burst(void *queue,
> > > >  		sdev = sdev->next;
> > > >  	} while (nb_rx == 0 && sdev != rxq->sdev);
> > > >  	rxq->sdev = sdev;
> > > > +	if (nb_rx)
> > > > +		failsafe_rx_set_port(rx_pkts, nb_rx,
> > > > +				     rxq->priv->data->port_id);
> > > >  	return nb_rx;
> > > >  }
> > > 
> > > I'm afraid the performance drop to be hard.
> 
> Mbufs are still hot from the oven at this stage, so it's not *that*
> expensive. I don't see a more efficient approach.

Yes, Ali did some quick tests showing no perf drop.

> > > How the port id in mbuf is used exactly?
> 
> Applications that dissociate Rx itself from packet processing, or whenever a
> networking stack is involved. Basically every time some code wonders where a
> packet comes from due to lack of context and looks at m->port for the
> answer (e.g. checking that a packet arrives on the right port given its
> destination address).
> 
> > > What crash are you seeing?
> 
> None, thankfully. In my specific use case, 6WINDGate's stack simply drops
> traffic coming from unknown ports.
> 
> However nothing prevents applications from using m->port as an index of some
> array they allocated to quickly retrieve port context without looking it
> up. They wouldn't expect indices they do not know about in there; assuming
> it will result in a crash is not far fetched.
> 
> > Another way to fix it without performance drop would be to add
> > a new driver op to set the top-level port id.
> > This top-level id would be stored in the private structure of the port,
> > initialized with the port id of the port itself, and used to fill mbufs.
> > 
> > Thoughts?
> 
> Adding a new devop as a fix would be a problem for stable releases, so this
> patch is definitely needed, at least as a first step.
> 
> I'm not against a new API, however would it be worth the trouble? Especially
> considering it would only be used by failsafe-like drivers with something to
> hide from applications which is not the main use case.
> 
> For some PMDs, this operation could only be done at init time before port ID
> is stored in private Rx queue data for fast retrieval. Retrieving it through
> a pointer so it can be updated anytime would make it more expensive than
> necessary for them.

I don't understand this comment.
The port id is currently retrieved via some pointers already.
I suggest to look at private structure, it is not different.

> It's understood that having failsafe in the dataplane has a cost, but even
> with the proposed fix, that cost is dwarfed by the amount of work done by a
> true PMD (and the application) for Rx processing.
> 
> My suggestion is to wait for someone to complain about the performance
> compared to what they had before that fix, only then see what we can do.

OK