arm64 runner stuck in git clean

[lazka]

https://github.com/msys2-arm/msys2-autobuild/actions/runs/14338197836/job/40190666280

  git clean/reset failed, sleeping for 0 seconds
  warning: failed to remove B/src/build-CLANGARM64/NUL: Invalid argument
  git clean/reset failed, sleeping for 1 seconds
  warning: failed to remove B/src/build-CLANGARM64/NUL: Invalid argument
  git clean/reset failed, sleeping for 2 seconds
  warning: failed to remove B/src/build-CLANGARM64/NUL: Invalid argument
  git clean/reset failed, sleeping for 3 seconds
  warning: failed to remove B/src/build-CLANGARM64/NUL: Invalid argument
  git clean/reset failed, sleeping for 4 seconds
  warning: failed to remove B/src/build-CLANGARM64/NUL: Invalid argument
  git clean/reset failed, sleeping for 5 seconds
  warning: failed to remove B/src/build-CLANGARM64/NUL: Invalid argument

fyi @jeremyd2019

[jeremyd2019]

How the heck did somebody create a file or directory called NUL? anyway, @dscho, more git clean brokenness

[jeremyd2019]

There was a 0-byte file called NUL in there. rmdir /s from cmd couldn't remove it, but rm -rf from MSYS2 could.

[dscho]

I guess you could remove those using the (unintuitive) core.longPaths=true setting (if you use Git for Windows)... but I guess these files were created using /usr/bin/git.exe?

[jeremyd2019]

No, I think they were created by some build process (these are untracked files trying to be removed in git clean). Do you recommend I enable core.longPaths=true in my Git for Windows config for the self-hosted runners?

[lazka]

Do you recommend I enable core.longPaths=true in my Git for Windows config for the self-hosted runners?

In theory we should already do that:

msys2-autobuild/msys2_autobuild/build.py

Line 195 in 23845c5

check_call(["git", "config", "core.longpaths", "true"], cwd=path)

The error occurred after building https://github.com/msys2/MINGW-packages/blob/master/mingw-w64-xavs/PKGBUILD - so no cygwin git involved.

[jeremyd2019]

Sure enough, it's in the .git/config of that repository.

[dscho]

Hmm. I guess there is code to skip the \\?\ dance when the path is short enough..

[lazka]

Easy enough to reproduce:

$ touch nul
$ git clean -f
fatal: Cannot lstat 'nul': No such file or directory

Not sure if that warrants creating a git4win issue, or if that is considered too much of an edge case (how to create such a file without cygwin?).

[dscho]

It might "work" with core.longPaths=true, which imitates how Cygwin can create that file IIUC.

[lazka]

It doesn't help sadly (see above)

[dscho]

It doesn't help sadly (see above)

Oops, sorry, I had missed that, only looked at your reply, my bad. My guess is that core.longPaths avoids the expensive \\?\<absolute-path> construction for paths that are short enough not to need it (or so it thinks).

I can think of a couple of work-arounds we could consider:

• introduce a mode core.longPaths=force that always uses that \\?\ prefix (which incurs worse performance, though, I had tested this about 10 years ago and the impact was in the double-digit percentages IIRC).
• use the is_valid_win32_path() function to see whether the \\?\ trick must be used (potentially guarded by special-casing the exact GetLastError() value after trying without, to lessen the performance impact, but then, I think I got that function quite fast and its run time will most likely be dwarfed by the I/O anyway).
• anything else I did not yet think of ;-)