2016-12-19 10:25, Bruce Richardson: > On Sat, Dec 17, 2016 at 11:43:25AM +0100, tom.barbe...@ulg.ac.be wrote: > > Hi, > > > > Your comments made me saw the line "PMD: i40e_set_rx_function(): Vector rx > > enabled, please make sure RX burst size no less than 4 (port=0)." > > > > The problem was probably that I was under this limit... Is there a way to > > get that limit through a function or something? > > > > With 16.04 I received sometimes 5 or 7 packets with a burst_size of 4 which > > respects this limit. I see that "[dpdk-dev] net/i40e: fix out-of-bounds > > writes during vector Rx" fixed that, as the limit was in fact 32 no matter > > the message. > > > > At the end, what should be the minimal rx burst size? How to find it at > > runtime for any NIC? I imagine that vector rx will create a problem if I > > give a burst size of 1 even with a recent DPDK version, right? > > > > Sadly, there doesn't appear to be any way to discover this, and the i40e > driver requires at least a burst size of 4 even with the latest DPDK. > From i40e_rxtx_vec_sse.c: > > 243 /* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP */ > 244 nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP); > 245 > > I think in this case the gap is not so much having a discovery mechanism > to determine min burst size, but rather a driver gap so as to allow some > form of slower-path fallback when we get below min-size bursts for the > vector driver.
Yes this is a severe bug. Please Tom, could you keep asking/monitoring this bug until it is fixed? Thanks for reporting.
