https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=240944

--- Comment #24 from Adrian Chadd <[email protected]> ---
Ok, this is where it stops being a driver thing and starts being a network
stack thing.
At least for e1000/igb/ixgbe LRO isn't a driver/hardware supported thing, it's
"just" leaning on packet RSS/hashid mapping flows into queues and then
net/netinet code (net/iflib.c, netinet/tcp_lro.c) is doing the batching and
such.

I never liked how it kind of was shoehorned on the side of the packet
processing path to improve cache/hotpath behaviour for the TCP stack. I get it,
but it's very specific to that.

Can we verify that the initial issue in the task (the locking up, crash, etc)
is fixed?
I have plenty of e1000/igb/ixgbe hardware here - my downstairs gateway is using
igb/em NICs in a bridge group! - and I'd like to create a separate task to
reproduce and properly deep dive into where and why LRO is failing here and
then run it by the network/transport stack folks so they can eyeball it.

(I don't mind suggesting them hacks like "ip forwarding is enabled, no LRO" and
"you can't add an interface to a bridge group if it has LRO enabled" as
enforced workarounds, but I first want to just get everyone on the same page
here.)

Thanks!

-- 
You are receiving this mail because:
You are the assignee for the bug.

Reply via email to