gpu: mount by-path directory

[tkatila]

oneCCL requires the /dev/dri/by-path folder to be available to create a mapping between GPUs.

[tkatila]

Apparently, I hadn't added the tests for bypath.go when I created the initial commit.

[codecov-commenter]

Codecov Report

Merging #1371 (3add9e6) into main (eeef8aa) will increase coverage by 0.32%.
The diff coverage is 100.00%.

❗ Current head 3add9e6 differs from pull request most recent head 5ac43f8. Consider uploading reports for the commit 5ac43f8 to get more accurate results

@@            Coverage Diff             @@
##             main    #1371      +/-   ##
==========================================
+ Coverage   50.56%   50.89%   +0.32%     
==========================================
  Files          44       44              
  Lines        4952     4985      +33     
==========================================
+ Hits         2504     2537      +33     
  Misses       2302     2302              
  Partials      146      146              

Impacted Files	Coverage Δ
cmd/gpu_plugin/gpu_plugin.go	87.21% <100.00%> (+2.26%)	⬆️

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[eero-t]

IMHO code would be more readable if all byPath, Bypath and bypath prefixes would be changed just to path:

$ sed -i \
  -e 's/by[Pp]ath\([a-zA-Z]\)/path\1/g' \
  -e 's/By[Pp]ath\([a-zA-Z]\)/Path\1/g' \
  $(git grep -il 'bypath[a-z]')

And copyright year in the new Go files should be 2023, not 2022.

Otherwise seems fine to me.

[eero-t]

New version with nicer local variable names looks IMHO much more readable! I.e. my comments are addressed.

[mythi]

looks great to me, thanks!

[tkatila]

Any further notes/comments?

[eero-t]

cmd/gpu_fakedev/fakedev.go should provide fake versions of all the files that GPU plugin checks (especially if their omission would cause constant log spam / errors).

Just make sure your changes to it do not conflict with ones in Harri's PR #1369.

[tkatila]

cmd/gpu_fakedev/fakedev.go should provide fake versions of all the files that GPU plugin checks (especially if their omission would cause constant log spam / errors).

I modified it to check in plugin creation for existence of the by-path dir. If it doesn't exist, it won't try to find the mounts. But if it does exist, it will warn about failures constantly.

[eero-t]

cmd/gpu_fakedev/fakedev.go should provide fake versions of all the files that GPU plugin checks (especially if their omission would cause constant log spam / errors).

I modified it to check in plugin creation for existence of the by-path dir. If it doesn't exist, it won't try to find the mounts. But if it does exist, it will warn about failures constantly.

Ok, looks good!

Fakedev can indeed skip things that are used only by (real) workloads, and do not impact GPU plugin behavior (e.g. what resources it provides).

[tkatila]

Hopefully all comments now addressed. Please review @mythi and @uniemimu .