Re: [PATCH net] i40e: Prevent unwanted interface name changes

On Thursday 08/21 at 22:39 +0200, Przemek Kitszel wrote:

On 8/21/25 16:23, Calvin Owens wrote:

quoted

On Thursday 08/21 at 10:00 +0200, Przemek Kitszel wrote:

quoted

On 8/20/25 19:41, Calvin Owens wrote:

quoted

On Wednesday 08/20 at 08:31 -0700, Calvin Owens wrote:

quoted

On Wednesday 08/20 at 08:42 +0200, Michal Schmidt wrote:

quoted

On Wed, Aug 20, 2025 at 6:30 AM Calvin Owens [off-list ref] wrote:

quoted

The same naming regression which was reported in ixgbe and fixed in
commit e67a0bc3ed4f ("ixgbe: prevent from unwanted interface name
changes") still exists in i40e.

Fix i40e by setting the same flag, added in commit c5ec7f49b480
("devlink: let driver opt out of automatic phys_port_name generation").

Fixes: 9e479d64dc58 ("i40e: Add initial devlink support")

But this one's almost two years old. By now, there may be more users
relying on the new name than on the old one.
Michal

Well, I was relying on the new ixgbe names, and I had to revert them
all in a bunch of configs yesterday after e67a0bc3ed4f :)

we have fixed (changed to old naming scheme) ixgbe right after the
kernel was used by real users (modulo usual delay needed to invent
a good solution)

No, the "fix" actually broke me for a *second time*, because I'd
already converted my infrastructure to use the *new* names, which match
i40e and the rest of the world.

We've seen *two* user ABI regressions in the last several months in
ixgbe now, both of which completely broke networking on the system.

I'm not here to whine about that: I just want to save as many people out
there in the real world as I can the trouble of having to do the same
work (which has absolutely no benefit) over the next five years in i40e.

If it's acceptable to break me for a second time to "fix" this, because
I'm the minority of users (a viewpoint I am in agreement with), it
should also be acceptable to break the minority of i40e users who are
running newer kernels to "fix" it there too.

Why isn't it?

I think we agree that it is ok-ish to sometime break setups for bleeding
edge users, then fix (aka undo). It's bad that this time it was with
effect equivalent to the first breakage (hope that it was easier to fix
locally when it occurred second time in a row).

I just want to re-emphasize, it was *not* my intent to gripe at you
about this. A big reason I test new kernels is in the hope I can hit
things like this myself and get them fixed before they impact the wide
userbase, I'm only frustrated I'm probably too late here to do that.

But we dispute over change from Oct 2023, for me it is carved in stone
at this point. Every user either adjusted or worked it around [1]

IMHO the date of the release (Jan 2024) is more relevant than the
commit date, but it's not really that different in this case.

I think there's merit to the idea that the lack of complaining is a sign
that most users have not had to adjust yet, because if they had, they'd
have complained about it. But I don't have any real data either way.

The objections raised over the new interface naming in ixgbe are in no
way specific to ixgbe. You can s/ixgbe/i40e/ any mail about it and
nothing really changes. They're generalized objections against the
renaming of interfaces, so from a certain POV people *are* actively
complaining.

quoted

quoted

quoted

And, even if it is e67a0bc3ed4f that introduced it, v6.7 was the first
release with it. I strongly suspect most servers with i40e NICs running
in the wild are running older kernels than that, and have not yet
encountered the naming regression. But you probably have much better
data about that than I do :)

RedHat patches their kernels with current code of the drivers that their
customers use (including i40e and ixgbe)
One could expect that changes made today to those will reach RHEL 10.3,
even if it would be named "kernel 6.12".

(*) the changes will likely be also in 10.2, but I don't want to make
any promises from Intel or Redhat here

But how many i40e users are actually on the most recent version of RHEL?
Not very many, is my guess. RHEL9 is 5.14, and has the old behavior.

RHEL 9 backported devlink for i40e in July 2024 [0], together with undo
of interface name change [1] (this likely tells why there were zero
complains from RH users).

[0]
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/commit/bcbc349375ecd977aa429c3eff4d182b74dcdd8a

[1]
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/commit/5ab8aa31dc2b44fbd6761bb19463f5427b9be245

Heh. Thank you very much for checking that, and for the links.

quoted

If you actually have data on that, obviously that's different. But it
sounds like you're guessing just like I am.

I could only guess about other OS Vendors, one could check it also
for Ubuntu in their public git, but I don't think we need more data, as
ultimate judge here are Stable Maintainers

Maybe I'm barking up the wrong tree, it's udev after all that decides to
read the thing in /sys and name the interfaces differently because it's
there...

In any case, Debian stable is on 6.12 and didn't patch it (just
checked), so I concede, it is simply too late :/

Thanks,
Calvin