← Back to team overview

touch-packages team mailing list archive

Re: [Bug 1347859] Re: Introduction of Predictable Network Interface Names (aka biosdevname) breaks working systems

 

On Mon, 13 Apr 2015 13:32:15 -0000, Robie Basak wrote:

>A release
>upgrade should not cause biosdevname to get installed. So can you (or
>someone) please confirm that this is definitely the behaviour in some
>case,

It has now been most of a year since I filed the bug report, so as you might
imagine, I no longer remember the details of the upgrade that prompted it.
Here's what I know:

On the system in question, eth0 disappeared and was replaced by
p<something>p<something>.

The main.log files under /var/log/dist-upgrade include lsb-release: quantal,
raring, and saucy. The currently-installed release is trusty.  This implies
that the steps to reproduce include one or more of those distribution
upgrades.  I guess precise must have been the first-installed release.

Although I don't know when biosdevname was first installed, it is listed in
the Upgrade: log entries of all the /var/log/dist-upgrade/<date>/main.log
files, including one from a year before my bug report.  I suppose that means
biosdevname was likely installed by some earlier release, rather than being
newly installed during the problematic dist-upgrade.

The motherboard's chipset is an Intel Z77 Express. The ethernet device uses
a Realtek RTL8111/8168/8411 series chip. The r8169 kernel driver is
currently in use.

>and if so provide steps to reproduce so that we can understand the
>mechanism involved here that is making this happen?

Sorry, but I already spent too much time on this issue when it bit me in the
first place.  Reproducing it to determine step-by-step instructions would
require taking the computer out of service, saving all of its data and
state, taking it through several install and upgrade cycles, and then
restoring everything.  That is far more disruption and time than I can
spare.

>Second, you say that some distributions require both systemd and
>biosdevname interface renaming to be disabled. My understanding is that
>on Ubuntu with systemd (so Vivid only), we do not rename interfaces via
>systemd by default - this must be done explicitly. Can we assume for the
>rest of the discussion that this is true, or otherwise can you provide
>steps to reproduce that demonstrate that it is not (again, another bug
>might be a good idea to avoid cluttering this one)?

That's quite possible.  It has been most of a year since I wrote that text,
but I probably meant to include non-ubuntu distros when I wrote "some
distributions".

-- 
You received this bug notification because you are a member of Ubuntu
Touch seeded packages, which is subscribed to ubuntu-meta in Ubuntu.
https://bugs.launchpad.net/bugs/1347859

Title:
  Introduction of Predictable Network Interface Names (aka biosdevname)
  breaks working systems

Status in biosdevname package in Ubuntu:
  Opinion
Status in ubuntu-meta package in Ubuntu:
  New
Status in udev package in Ubuntu:
  Opinion

Bug description:
  Relatively recent linux distribution upgrades have been causing
  computers' ethernet devices to be unexpectedly renamed. While I
  understand that consistent device naming solves problems on some
  systems (mostly multi-NIC servers and a few specialty embedded
  devices), unilaterally forcing these changes on everyone is causing a
  lot of frustration. Here are some of the problems I've encountered:

  Interface names that were easily recognized as abbreviations for their
  device type have been replaced by cryptic names that have no obvious
  meaning whatsoever. It's easy to guess that eth0 is short for ethernet
  #0. What the heck is p4p1 supposed to mean? How is a human supposed to
  guess that the first p stands for "PCI slot", that the second p stands
  for "port number", and that the whole mysterious string represents an
  ethernet interface? This new naming convention is inferior to the old
  one in at least one significant respect: it makes things more
  difficult to understand.

  One of the more useful examples of consistency that unix-like systems
  have enjoyed for decades has been thrown out: the extremely well-known
  ethernet device names. This creates yet another hurtle for users and
  admins when switching between different operating systems or trying to
  apply general-purpose unix knowledge.

  A lot of documentation has been broken. I have no idea how many
  manuals, forum posts, bug reports, printed instructions, email
  messages, personal notes, books, and other forms of documentation in
  the world refer to a unix ethernet device as eth0, but I'll bet the
  number is huge. All that valuable guidance has just been rendered
  misleading or even useless to anyone who doesn't keep up with the
  latest distribution-specific device naming experiments; in other
  words: the people who need it most.

  Well-established workflows have been broken. The change trips up users
  and admins who have for years been getting tasks done quickly with
  commands that they could recall and execute without a second thought.
  They are suddenly finding that their workflows no longer work. This
  interrupts tasks that should have been quick and easy, forcing people
  figure out why known-good procedures are broken, think about how to
  modify their memorized commands to work on the affected systems, and
  train their fingers to type those new commands as quickly as they did
  the old ones. Beyond being irritating, it can eat up a bunch of time
  that some of us don't have to spare.

  Working systems have been broken. Tools and automation scripts,
  especially those developed for site-specific use, often make the
  difference between a computer that does real work and a useless
  generic OS installation. Sometimes they even make the difference
  between a malfunctioning headless box that can be fixed over the
  network and an expensive brick. It is quite common for such software
  to make some minimal assumptions about its runtime environment, like
  assuming that the name of the only network device that has ever been
  or will ever be present will not suddenly change after being stable
  for months or years. There are also applications (e.g. Matlab) and
  configuration files (e.g. smb.conf, dhclient.conf, isc-dhcp-server)
  that might depend on references to eth0. Renaming a critical and
  ubiquitous device like this is so very likely to cause problems that
  it should never, ever be done in an upgrade without the admin's
  explicit consent.

  Sufficient warning of the change was not given. On one of my machines,
  eth0 was renamed to p4p1 when I upgraded to Ubuntu 14.04 (trusty), yet
  I don't see any mention of it in the Trusty release notes, nor in any
  of the notes for releases of the previous several years. Is it buried
  in fine print someplace that I missed? Having to figure out for myself
  what changed, why, and how to revert it (in multiple ways on each
  machine) was a significant waste of my time. Multiply that by all the
  other people who were affected similarly, and I'll bet we'd get an
  embarrassing number of needlessly wasted person-months that could have
  been saved with a simple announcement and link to documentation.

  In short, the way this feature was forced on the world was an
  irresponsible blunder. It doesn't matter that the change was meant to
  address some other problem. Breaking working systems is far worse than
  allowing an existing problem to remain until someone opts in to a fix.
  This concept is so important that anyone who doesn't get it really has
  no business committing code to an operating system used by so many
  other people.

  I am filing this bug report against multiple projects because more
  than one is now overriding the kernel's device names, because each
  project must be reconfigured in a different way in order to disable
  this behavior, and because I believe the overall failure here lies not
  only in careless feature implementation but also in careless
  deployment. If the people involved in developing, integrating,
  testing, distributing, and using this new behavior had all been
  talking with each other, perhaps they would have coordinated better
  when tinkering with something upon which others have relied for longer
  than many coders have been alive.

  Here are some suggestions that might have mitigated the mess caused by
  eth0 renaming:

  It should be disabled by default for upgrades. (Don't break working
  systems.)

  It could be disabled by default for systems that have only one network
  device. (Fixing workstations when their eth0 was renamed to p4p1 was
  more than enough hassle. I do not look forward to them changing to
  p1p1 when their motherboards are upgraded.)

  It should not override custom udev rules. (People who added or edited
  /etc/udev/rules.d/70-persistent-net.rules did so for a reason.)

  It should be implemented in only one place. (On some distributions, it
  seems systemd and biosdevname are both doing the job, and both must be
  reconfigured in order to disable it.)

  It could generate names that are as understandable as the ones being
  replaced. (I can easily guess that "eth" is ethernet. What the heck is
  "p"?)

  It should have come with sufficient warning and documentation on how
  to disable it. (Where are the release notes?)

  I hope that those who read this report will genuinely try to
  understand the significance of the trouble that has been (and
  continues to be) caused here, rather than responding with the same
  stubborn arrogance that sometimes shows up when a developer's changes
  are criticized by his users. I think this can be done right, but it
  hasn't been so far. Solving one problem by creating another set of
  problems is not exactly a win.  I'm sure that the open source
  community can do better than this.  Thank you for your attention.

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/biosdevname/+bug/1347859/+subscriptions


References