2017-02-10 09:02, Liu, Yong:
> From: Thomas Monjalon [mailto:thomas.monja...@6wind.com]
> > 2017-02-10 07:53, Liu, Yong:
> > > From: Thomas Monjalon
> > > > 2017-02-09 22:25, Marvin Liu:
> > > > > Add new option parse-ptype in this sample in case of pmd driver
> > > > > not provide packet type info. If this option enabled, packet type
> > > > > will be analyzed in Rx callback function.
> > > > [...]
> > > > > + if (parse_ptype) {
> > > > > + if (add_cb_parse_ptype(portid, queueid) < 0)
> > > > > + rte_exit(EXIT_FAILURE,
> > > > > + "Fail to add ptype cb\n");
> > > > > + } else if (!check_ptype(portid))
> > > > > + rte_exit(EXIT_FAILURE,
> > > > > + "PMD can not provide needed ptypes\n");
> > > >
> > > > Instead of adding a new option, why not adding the callback
> > automatically
> > > > if the packet type is not supported by the hardware?
> > >
> > > Thomas,
> > > We want to let user choice which kind of method for packet type parsing.
> > > If start application with parse-type option, is meaning user want to use
> > software parsing otherwise will use hardware parsing.
> >
> > I do not understand why this user choice matters.
> > If it is available, hardware ptype is better, isn't it?
> > It it is not available, we need to be aware of this specific issue,
> > otherwise we have the error "PMD can not provide needed ptypes"
> > (without suggesting to use the option).
>
> Yes, hardware always has better performance than software. I think it matters
> in some performance measurement scenarios.
> Like l3fwd, we compared performance with software and hardware packet parsers
> and this option may not have much value in other samples.
> I will rework this patch and fallback to software if hardware not support.
The ip_fragmentation case has been reworked this way:
http://dpdk.org/patch/21769
