Go 1.24 breaks runc exec

[kolyshkin]

Description

I opened #4598 to check if runc works with Go 1.24 (currently using go1.24rc2) and found out that runc exec is badly broken.

The cause is allegedly a bug in the kernel, which is reported in https://lore.kernel.org/linux-fsdevel/20250122020850.2175427-1-kolyshkin@gmail.com/T/#u

I've also reported a Go bug to golang/go#71375

We can work around it by reimplementing io.Copy but first let's see what kernel and Go developers have to say.

Steps to reproduce the issue

See reproducers in the links above. You can also compile runc with Go 1.24 and run runc exec in the foreground.

Describe the results you received and expected

See above links.

What version of runc are you using?

I belive that any and all versions of runc are affected, when using Go 1.24rc2 (or rc1). The code introducing the use of io.Copy to copy from os.Stdin to a process' stdin is introduced in the very first runc commit 9fac183 aka "Initial commit of runc binary".

Host OS information

No response

Host kernel information

No response

[fuweid]

cc @Burning1020 @abel-von

If I remember it correctly, rust io copy uses splice by default and run into tty copy issue in shim side. Just tag you if it can help.

[abel-von]

@fuweid Yes, and we replaced the std::io::copy by a copy function imlemented ourselves.

[kolyshkin]

@fuweid Yes, and we replaced the std::io::copy by a copy function imlemented ourselves.

Was it the same issue (the pipe being locked in the kernel so its other end can't be closed)? Can you point out to the discussion you were having?

[abel-von]

It is a different issue but I think the root cause is the same. containerd/rust-extensions#79

[abel-von]

Does this bug only effect runc? as there are also many calls of io.Copy in containerd.

[akhilerm]

Does this bug only effect runc? as there are also many calls of io.Copy in containerd.

I had ran the containerd CI on 1.24rc1 and 1.24rc2 here containerd/containerd#11211. But couldnt see any failures or we may be missing tests for this case.

[kolyshkin]

It is a different issue but I think the root cause is the same. containerd/rust-extensions#79

Yes, looks like it's the same underlying issue. @fuweid did you end up reporting this to the kernel folks?

[fuweid]

@kolyshkin @abel-von sorry, I forgot to report it to kernel community. 😢

[kolyshkin]

This is now fixed in Go.