net/mlx5: Correctly set gso_size when LRO is used

gso_size is expected by the networking stack to be the size of the
payload (thus, not including ethernet/IP/TCP-headers). However, cqe_bcnt
is the full sized frame (including the headers). Dividing cqe_bcnt by
lro_num_seg will then give incorrect results.

For example, running a bpftrace higher up in the TCP-stack
(tcp_event_data_recv), we commonly have gso_size set to 1450 or 1451 even
though in reality the payload was only 1448 bytes.

This can have unintended consequences:
- In tcp_measure_rcv_mss() len will be for example 1450, but. rcv_mss
will be 1448 (because tp->advmss is 1448). Thus, we will always
recompute scaling_ratio each time an LRO-packet is received.
- In tcp_gro_receive(), it will interfere with the decision whether or
not to flush and thus potentially result in less gro'ed packets.

So, we need to discount the protocol headers from cqe_bcnt so we can
actually divide the payload by lro_num_seg to get the real gso_size.

v2:
 - Use "(unsigned char *)tcp + tcp->doff * 4 - skb->data)" to compute header-len
   (Tariq Toukan <tariqt@nvidia.com>)
 - Improve commit-message (Gal Pressman <gal@nvidia.com>)

Fixes: e586b3b0 ("net/mlx5: Ethernet Datapath files")
Signed-off-by: Christoph Paasch <cpaasch@openai.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Reviewed-by: Gal Pressman <gal@nvidia.com>
Link: https://patch.msgid.link/20250715-cpaasch-pf-925-investigate-incorrect-gso_size-on-cx-7-nic-v2-1-e06c3475f3ac@openai.com


Signed-off-by: Jakub Kicinski <kuba@kernel.org>