Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

host-device: use temp network namespace for rename #1073

Merged
merged 1 commit into from
Oct 2, 2024

Conversation

champtar
Copy link
Contributor

Using a temporary name / doing a fast rename causes some race conditions with udev and NetworkManager: https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/issues/1599

Fixes #1072

@champtar champtar force-pushed the host-device-temp-ns branch from 12cb704 to e23a29e Compare August 14, 2024 04:40
@champtar champtar force-pushed the host-device-temp-ns branch 7 times, most recently from 9d0a6aa to 49e807e Compare August 16, 2024 15:01
@champtar champtar requested a review from aojea August 19, 2024 19:01
@champtar champtar force-pushed the host-device-temp-ns branch from 49e807e to a75d829 Compare August 22, 2024 22:39
@champtar
Copy link
Contributor Author

Rebased on main, removed go version bump

@champtar champtar force-pushed the host-device-temp-ns branch 5 times, most recently from 5c6ae75 to d97a084 Compare August 27, 2024 20:11
@champtar
Copy link
Contributor Author

Rebased on main
@aojea do you have time to do a full review ?

@champtar
Copy link
Contributor Author

Maybe @squeed you have some time to review this ?

@champtar champtar force-pushed the host-device-temp-ns branch from d97a084 to 0c69818 Compare September 3, 2024 02:00
@champtar
Copy link
Contributor Author

champtar commented Sep 3, 2024

Rebased

Copy link
Contributor

@s1061123 s1061123 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't have not so much time to review whole code, but I'm wondering that when temp network namespace is deleted, created by Unshare()?

tempNS.Close() just closes the netns handle(i.e. inode) in process, not remove netns. (currently pkg/ns/ns_linux.go is not designed to create/delete network ns, actually. Just manipulate 'already created' netns).

@EdDev
Copy link
Contributor

EdDev commented Sep 3, 2024

I would expect CNI plugins to execute on servers and not desktops. The auto-detection of new interfaces (i.e. detecting and managing them) is supposed to be disabled on servers that use NM.

Can that be an alternative solution?

@champtar
Copy link
Contributor Author

champtar commented Sep 3, 2024

I would expect CNI plugins to execute on servers and not desktops. The auto-detection of new interfaces (i.e. detecting and managing them) is supposed to be disabled on servers that use NM.

Can that be an alternative solution?

I'm already using no-auto-default=* in NM config, but I use cockpit and want some interfaces (identified by mac pattern) not to be configurable at all (NM_UNMANAGED), and this race condition makes it impossible

@EdDev
Copy link
Contributor

EdDev commented Sep 3, 2024

I'm already using no-auto-default=* in NM config, but I use cockpit and want some interfaces (identified by mac pattern) not to be configurable at all (NM_UNMANAGED), and this race condition makes it impossible

I think the basic is setting the no-auto-default and ignore-carrier as set with the NetworkManager-config-server package. But I think NM has more options in this regard.
If you already exhausted this direction with the NM experts, then I guess that is it.

The proposed solution you provided to overcome the problem is innovative, but at the same time feels a bit hacky. It attempts to overcome a problem of NM & udev, which a general CNI should not care about.
I guess NM had not faced this problem before because this is not a common action, but if we think about it, the same operation could have been done by a human or automation, regardless of a CNI, and one could expect NM to still behave correctly. Solving this at the core, would solve the CNI scenario as well.

If the workaround is accepted here, then at the minimum I would suggest to have an open bug on the NM project, asking to solve it there and ref in the code. I.e. making the current solution temporary.

@champtar
Copy link
Contributor Author

champtar commented Sep 3, 2024

I'm already using no-auto-default=* in NM config, but I use cockpit and want some interfaces (identified by mac pattern) not to be configurable at all (NM_UNMANAGED), and this race condition makes it impossible

I think the basic is setting the no-auto-default and ignore-carrier as set with the NetworkManager-config-server package. But I think NM has more options in this regard. If you already exhausted this direction with the NM experts, then I guess that is it.

The proposed solution you provided to overcome the problem is innovative, but at the same time feels a bit hacky. It attempts to overcome a problem of NM & udev, which a general CNI should not care about. I guess NM had not faced this problem before because this is not a common action, but if we think about it, the same operation could have been done by a human or automation, regardless of a CNI, and one could expect NM to still behave correctly. Solving this at the core, would solve the CNI scenario as well.

Agreed that NM should do better, but the bug is 3 weeks old and still waiting for triage.
Also having the NM fix backported will take a long time.

If the workaround is accepted here, then at the minimum I would suggest to have an open bug on the NM project, asking to solve it there and ref in the code. I.e. making the current solution temporary.

My fix also leads to less udev events / rules execution, so I really don't see it as a workaround
And maybe some other software handling network configuration are impacted, each distro uses its own ...

@EdDev
Copy link
Contributor

EdDev commented Sep 3, 2024

My fix also leads to less udev events / rules execution, so I really don't see it as a workaround
And maybe some other software handling network configuration are impacted, each distro uses its own ...

I see it as a workaround because it fixes a problem of a different component. I.e. the CNI needs to take into account the host OS configuration and components.
Fixing it at the "edge" is not scaling well and burdens this CNI logic.

The cost of changing this CNI and supporting the added logic is not easier (or faster) than fixing the root cause.
NM tasks could be prioritized through D/S if this is possible from your side.

@champtar champtar force-pushed the host-device-temp-ns branch from 0c69818 to 065d805 Compare September 3, 2024 13:39
@champtar
Copy link
Contributor Author

champtar commented Sep 3, 2024

Fixing it at the "edge" is not scaling well

Depending on how you write your udev rules they might be impacted as well, for example if you do negative matches ATTR{address}!="00:11:22:*" this will match because the device doesn't exists, so ATTR{address} is empty

Fast rename is also problematic for udev alone, and race condition are always fun to debug

and burdens this CNI logic.

it's the same number of steps

With temp name:

  • link down
  • rename 1
  • change namespace
  • rename 2

With temp namespace:

  • create a temp namespace
  • change namespace (this also down the link)
  • rename
  • change namespace

@champtar champtar force-pushed the host-device-temp-ns branch 2 times, most recently from 3761ba9 to e530620 Compare September 10, 2024 18:15
@squeed
Copy link
Member

squeed commented Sep 17, 2024

Honestly, this makes sense to me. I understand that NM changes can take far too long to roll out, especially when there is no appreciable workaround.

Personally, I'm in favor of this PR. It's not like namespaces are so expensive. IIRC there have been other issues regarding "churn" as the host-network device is reconfigured, so this seems like a reasonable change.

Using a temporary name / doing a fast rename causes
some race conditions with udev and NetworkManager:
https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/issues/1599

Signed-off-by: Etienne Champetier <e.champetier@ateme.com>
@champtar
Copy link
Contributor Author

@squeed as it make sense to you, any chance you can review this PR ?

@EdDev
Copy link
Contributor

EdDev commented Sep 19, 2024

and burdens this CNI logic.

it's the same number of steps

With temp name:
* link down
* rename 1
* change namespace
* rename 2

With temp namespace:
* create a temp namespace
* change namespace (this also down the link)
* rename
* change namespace

Ok, this argument is convincing.

@champtar
Copy link
Contributor Author

This as been approved by @squeed for a week, ready for review for a good month, how can we move forward ?

@squeed
Copy link
Member

squeed commented Sep 26, 2024

@champtar we’ll be cutting a release in a week or so, this will get in.

@squeed squeed dismissed s1061123’s stale review October 2, 2024 08:30

Discussed in person, PR is Ok to merge.

@squeed squeed merged commit a4b80cc into containernetworking:main Oct 2, 2024
6 checks passed
@champtar champtar deleted the host-device-temp-ns branch December 27, 2024 16:59
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

host-device temp interface name causes race conditions
5 participants