Re: [dpdk-dev] ixgbe vector rx does not conform torte_eth_rx_burst() API

[Bruce Richardson <bruce.richardson@intel.com>]

On Thu, Jul 16, 2020 at 11:08:49AM +0200, Morten Brørup wrote:
> > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Zhao1, Wei
> > Sent: Thursday, July 16, 2020 10:50 AM
> > 
> > Hi,
> > 
> > > -----Original Message-----
> > > From: Morten Brørup <mb@smartsharesystems.com>
> > > Sent: Thursday, July 16, 2020 12:03 AM
> > > To: Zhao1, Wei <wei.zhao1@intel.com>; Guo, Jia <jia.guo@intel.com>
> > > Cc: dev@dpdk.org; Thomas Monjalon <thomas@monjalon.net>; Yigit,
> > Ferruh
> > > <ferruh.yigit@intel.com>; Andrew Rybchenko
> > <arybchenko@solarflare.com>
> > > Subject: ixgbe vector rx does not conform to rte_eth_rx_burst() API
> > >
> > > Wei, Jeff,
> > >
> > > For the ixgbe driver using vector functions, i.e.
> > ixgbe_recv_pkts_vec(), calling
> > > rte_eth_rx_burst() with nb_pkts > RTE_IXGBE_MAX_RX_BURST only returns
> > > RTE_IXGBE_MAX_RX_BURST packets. E.g. calling rte_eth_rx_burst() with
> > > nb_pkts=64 only returns 32 packets.
> > >
> > >
> > > The API description of rte_eth_rx_burst() says:
> > >
> > > <quote>
> > > The rte_eth_rx_burst() function returns the number of packets
> > actually
> > > retrieved, which is the number of rte_mbuf data structures
> > effectively supplied
> > > into the rx_pkts array. A return value equal to nb_pkts indicates
> > that the RX
> > > queue contained at least rx_pkts packets, and this is likely to
> > signify that other
> > > received packets remain in the input queue. Applications implementing
> > a
> > > "retrieve as much received packets as possible" policy can check this
> > specific
> > > case and keep invoking the rte_eth_rx_burst() function until a value
> > less than
> > > nb_pkts is returned.
> > > </quote>
> > >
> > > The driver implementation does not conform to the documented behavior
> > for
> > > "retrieve as much received packets as possible" applications.
> > 
> > It seems not an issue, this function has comment bellow, it is design
> > work in that way.
> > 
> > 
> > /*
> >  * vPMD receive routine, only accept(nb_pkts >=
> > RTE_IXGBE_DESCS_PER_LOOP)
> >  *
> >  * Notice:
> >  * - nb_pkts < RTE_IXGBE_DESCS_PER_LOOP, just return no packet
> >  * - nb_pkts > RTE_IXGBE_MAX_RX_BURST, only scan RTE_IXGBE_MAX_RX_BURST
> >  *   numbers of DD bit
> >  * - floor align nb_pkts to a RTE_IXGBE_DESC_PER_LOOP power-of-two
> >  */
> > 
> 
> I noticed this already. And yes, ixgbe_recv_pkts_vec() does what its comments says.
> 
> However, when ixgbe_recv_pkts_vec() is used as the driver's implementation of the rte_eth_rx_burst() function call, the rte_eth_rx_burst() function does not do what is expected of the rte_eth_rx_burst() function.
> 
> The implementation must conform to the API that it implements.
> 
> If you don't want to update the ixgbe_recv_pkts_vec() function, I propose that you add a wrapper function that calls ixgbe_recv_pkts_vec() repeatedly, and use the wrapper function as the implementation of the rte_eth_rx_burst() function.
>
I agree it does not conform to the API as documented. Looking at the i40e
AVX2 vector code paths, the regular (no-fragmented packets) path does not
seem to put a limit on the size of burst received, while the scattered
receive path uses the latter scheme of a wrapper function. Therefore,
either approach should work for ixgbe driver, I think.

/Bruce