Page MenuHomeFreeBSD

cxgbe: Allow parse_pkt to internally queue a packet.
ClosedPublic

Authored by jhb on Feb 14 2023, 12:21 AM.
Tags
None
Referenced Files
F168674178: D38577.id.diff
Sat, Aug 29, 1:52 PM
Unknown Object (File)
Sat, Aug 29, 10:15 AM
Unknown Object (File)
Fri, Aug 28, 3:36 PM
Unknown Object (File)
Thu, Aug 27, 7:06 PM
Unknown Object (File)
Thu, Aug 27, 11:30 AM
Unknown Object (File)
Thu, Aug 27, 10:11 AM
Unknown Object (File)
Wed, Aug 26, 1:19 PM
Unknown Object (File)
Tue, Aug 25, 5:32 PM
Subscribers

Details

Summary

If parse_pkt returns EINPROGRESS, return from cxgbe_transmit
without queueing the packet in a txq. Use this to move the call
to ethofld_transmit for packet pacing into parse_pkt.

Diff Detail

Repository
rG FreeBSD src repository
Lint
Lint Not Applicable
Unit
Tests Not Applicable

Event Timeline

np added inline comments.
sys/dev/cxgbe/t4_main.c
3063–3070

Should we get rid of the __predict_false if EINPROGRESS can commonly occur on the fast path for tx?

This revision is now accepted and ready to land.Feb 16 2023, 5:33 PM
sys/dev/cxgbe/t4_main.c
3063–3070

Hmmm, it depends on what we want to optimize for. Plain NIC traffic will still take the rc == 0 case. Also, for the EINPROGRESS case, all the work has already been done in parse_pkt and we are just trying to return after possibly draining the mp_ring, so I think that's still less important to optimize for vs getting down to the fast path for plain TX.