[Bug 280074] em(4) temporarily hangs and re-enables TX CSUM if a bridge it's a member of is modified

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=280074

--- Comment #29 from Zhenlei Huang <[email protected]> ---
(In reply to Joshua Kinard from comment #28)

> So this is actually expected behavior for 14.x and up, when em(4) and other Intel
> chipsets were migrated to using iflib, or is this just a thing that got missed
> initially and hasn't really had attention?

From the current implementation of iflib(4), I think that is expected behavior.
But I think there're still room to improve. I checked the source of cxgbe(4)
and tested with t520cr and can confirm t520cr does not require HW reset or full
re-init on changing TXCSUM capability, thus no link flipping and packet loss
observed.

@Kevin

For the TX offload features, I think when they are disabled HW reset is
unnecessary, since the net stack will talk to if_hwassist and do e.g. soft
checksum when the HW offload is disabled. It appears that enabling the HW's TX
offload features has no side effect. Well, reseting HW and disabling the
features may save some resources but that depends on the HW design IMO.

For the RX offload features, the HW must be programmed correctly, or a reset is
required, to instruct the HW to do the desired work.

> > Can you also share the ifconfig before and after shutdown jail, and the dmesg ?
> I want to check whether the link of em0 is still flipping.

> That is already posted in comment#22, per Kevin's request earlier.

I think you misread my request. Kevin asked you to `Remove "EM_OPTS" in
entirety` and test and report. I asked to re-test with the following EM_OPTS,
```
EM_OPTS="-lro -tso -tso6 -mextpg"
```
and report if the issue persists.

Be aware with the disabled capability `-mextpg`. epair(4) does not support
that. I'd expect that you will no longer suffer em(4) link flipping once you
disable that capability along with "-lro -tso -tso6" .

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