On Wed, Oct 25, 2017 at 11:11:08AM -0700, Ferruh Yigit wrote:
> On 10/23/2017 10:42 AM, Roger B. Melton wrote:
> > On 10/20/17 3:04 PM, Ferruh Yigit wrote:
> >> On 10/12/2017 10:24 AM, Roger B Melton wrote:
> >>> When copying VLAN tags from the RX descriptor to the vlan_tci field
> >>> in the mbuf header, igb_rxtx.c:eth_igb_recv_pkts() and
> >>> eth_igb_recv_scattered_pkts() both assume that the VLAN tag is always
> >>> little endian. While i350, i354 and /i350vf VLAN non-loopback
> >>> packets are stored little endian, VLAN tags in loopback packets for
> >>> those devices are big endian.
> >>>
> >>> For i350, i354 and i350vf VLAN loopback packets, swap the tag when
> >>> copying from the RX descriptor to the mbuf header. This will ensure
> >>> that the mbuf vlan_tci is always little endian.
> >>>
> >>> Signed-off-by: Roger B Melton <rmel...@cisco.com>
> >> <...>
> >>
> >>> @@ -946,9 +954,16 @@ eth_igb_recv_pkts(void *rx_queue, struct rte_mbuf
> >>> **rx_pkts,
> >>>
> >>> rxm->hash.rss = rxd.wb.lower.hi_dword.rss;
> >>> hlen_type_rss =
> >>> rte_le_to_cpu_32(rxd.wb.lower.lo_dword.data);
> >>> - /* Only valid if PKT_RX_VLAN_PKT set in pkt_flags */
> >>> - rxm->vlan_tci = rte_le_to_cpu_16(rxd.wb.upper.vlan);
> >>> -
> >>> + /*
> >>> + * The vlan_tci field is only valid when PKT_RX_VLAN_PKT is
> >>> + * set in the pkt_flags field and must be in CPU byte order.
> >>> + */
> >>> + if ((staterr & rte_cpu_to_le_32(E1000_RXDEXT_STATERR_LB)) &&
> >>> + (rxq->flags & IGB_RXQ_FLAG_LB_BSWAP_VLAN)) {
> >> This is adding more condition checks into Rx path.
> >> What is the performance cost of this addition?
> >
> > I have not measured the performance cost, but I can collect data. What
> > specifically are you looking for?
> >
> > To be clear the current implementation incorrect as it does not
> > normalize the vlan tag to CPU byte order before copying it into mbuf and
> > applications have no visibility to determine if the tag in the mbuf is
> > big or little endian.
> >
> > Do you have any suggestions for an alternative approach to avoid rx
> > patch checks?
>
> No suggestion indeed. And correctness matters.
>
> But this add a cost and I wonder how much it is, based on that result it may
> be
> possible to do more investigation for alternate solutions or trade-offs.
>
> Konstantin, Bruce, Wenzhuo,
>
> What do you think, do you have any comment?
>
For a 1G driver, is performance really that big an issue? Unless you
have a *lot* of 1G ports, I would expect most platforms not to notice an
extra couple of cycles when dealing with 1G line rates.
/Bruce
