On Tue, May 20, 2025 at 02:40:05PM +0900, Alexander Bluhm wrote:
> I saw counters in FreeBSD that indicate strange
> hardware behaviour.  Does it make sense to add them to kstat to
> find potential bugs?

In ixl(4) kstat is mapped directly to traffic counters read from hardware
register space. ice(4) does not yet use kstat but if it was using kstat
then the use case should probably be the same for consistency.

FreeBSD ice(4) checksum offload counters are a pure software construct
and mostly count the number of checksum errors.

Which probably maps best to our netstat's "discarded for bad checksums"
counters, tcps_rcvbadsum and udps_badsum. Which should already be counted
by tcp_input() and udp_input() since the driver sets M_TCP_CSUM_IN_BAD
or M_UDP_CSUM_IN_BAD when the hardware signals a checksum error.

I would rather avoid introducing driver-specific software counters.