On 1/27/21 3:10 PM, Loftus, Ciara wrote:
[...]

Thanks for your feedback Daniel.

The stats tell us that there is *a* problem whereas the traces will shed
light on what that problem is. eg. The XSK_TRACE_DROP_PKT_TOO_BIG
trace tells us we dropped a packet on RX due to it being too big vs. ss
would just tell us the packet was dropped.

But wouldn't that just be a matter of extending struct xdp_diag_stats +
xsk_diag_put_stats() with more fine-grained counters? Just wondering given
you add the trace_xsk_packet_drop() tracepoints at locations where we
bump most of these counters already for ss tool. I guess this was my confusion -
as far as I can see only the two XSK_TRACE_DROP_{PKT_TOO_BIG,DRV_ERR_TX} are
not covered in xdp_diag_stats. Is there any other reason that the diag is
not sufficient for your use case?

Thanks,
Daniel