RE: [PATCH] net/tap: do not include l2 header in gso size when compared with mtu

[Ophir Munk <ophirmu@nvidia.com>]

The effective limit in Tx is the maximum packet size and not the MTU.
On my Ethernet NIC the MTU size is 1500 bytes. As long as I send packets
up to 1514 bytes size (including an Ethernet header of 14 bytes) - TX succeeds.
Conversely, if I add to the packet VLAN, Q-in-Q, VXLAX or any other L2 headers - TX fails
This is because the NIC's maximum packet size is 1514 bytes, regardless of the L2 header size.
(Remark: NIC max size may include 4 bytes CRC as well).
Having said that - I suggest setting up an MTU TAP that considers the maximum supported packet size.
Using mbuf_in-> l2_len in calculating the tso_segsz limit - does not seem right to me.

> -----Original Message-----
> From: Harold Huang <baymaxhuang@gmail.com>
> Sent: Tuesday, March 8, 2022 4:35 PM
> To: dev@dpdk.org
> Cc: Ophir Munk <ophirmu@nvidia.com>; stable@dpdk.org; Keith Wiles
> <keith.wiles@intel.com>; Raslan Darawsheh <rasland@nvidia.com>;
> jiayu.hu@intel.com
> Subject: Re: [PATCH] net/tap: do not include l2 header in gso size when
> compared with mtu
> 
> On Mon, Feb 28, 2022 at 4:27 PM Harold Huang <baymaxhuang@gmail.com>
> wrote:
> >
> > The gso size is calculated with all of the headers and payload. As a
> > result, the l2 header should not be included when comparing gso size
> > with mtu.
> >
> > Fixes: 050316a88313 ("net/tap: support TSO (TCP Segment Offload)")
> > Cc: stable@dpdk.org
> > Signed-off-by: Harold Huang <baymaxhuang@gmail.com>
> > ---
> >  drivers/net/tap/rte_eth_tap.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/drivers/net/tap/rte_eth_tap.c
> > b/drivers/net/tap/rte_eth_tap.c index f1b48cae82..2b561d232c 100644
> > --- a/drivers/net/tap/rte_eth_tap.c
> > +++ b/drivers/net/tap/rte_eth_tap.c
> > @@ -731,7 +731,7 @@ pmd_tx_burst(void *queue, struct rte_mbuf
> **bufs, uint16_t nb_pkts)
> >                                         mbuf_in->l4_len;
> >                         tso_segsz = mbuf_in->tso_segsz + hdrs_len;
> >                         if (unlikely(tso_segsz == hdrs_len) ||
> > -                               tso_segsz > *txq->mtu) {
> > +                               tso_segsz > *txq->mtu +
> > + mbuf_in->l2_len) {
> >                                 txq->stats.errs++;
> >                                 break;
> >                         }
> > --
> > 2.27.0
> >
> 
> Hi, Jiayu,
> 
> This is the only example in the driver to use GSO. I think it is important for us
> to calculate a correct GSO size. I see you are the GSO lib maintainer, could
> you please help review this patch?