Fix broken brew formula due to loading operating_system.rb customizations too late

[deivid-rodriguez]

What was the end-user or developer problem that led to this PR?

Hombrew ruby formula no longer works since #5044 .

What is your fix for the problem, implemented in this PR?

The problem is that setting up default gems needs Gem.default_dir, which is something commonly customized in this file. So if we setup default gems before loading operating_system.rb customizations, we will setup the wrong default gems.

Unfortunately fixing this requires reverting #5044, but I wasn't even sure that enhancement was the right thing to do, and the culprit ended up being something else, so we should fix the real culprit there instead.

Fixes #5151.
Fixes #5043.

Make sure the following tasks are checked

• Describe the problem / feature
• Write tests for features and bug fixes
• Write code to solve the problem
• Make sure you follow the current code style and write meaningful commit messages without tags

[voxik]

  1) Failure:
GemTest#test_operating_system_customizing_default_dir [/builddir/build/BUILD/ruby-3.0.3/test/rubygems/test_rubygems.rb:39]:
Expected "/builddir/build/BUILD/ruby-3.0.3/ruby: No such file or directory -- gem (LoadError)" to be empty.

[voxik]

Just FTR, the test failure happened when I applied the patch to Ruby 3.0.3 and executed the test suite via make check. I am not sure that the gem command could be found on PATH in this configuration.

[voxik]

And again, I am not happy about the symlink expansion. This is not right. This is not intuitive.

Just for example, if you execute Ruby via /bin/ruby, then it is not acceptable Gems to be reported from /urs/lib. This causes just confusion.

If the symlink expansion would be so good idea, then this would be the shell behavior:

$ $ ll /
celkem 20
dr-xr-xr-x.   1 root root    0 22. zář 11.54 afs
lrwxrwxrwx.   1 root root    7 22. zář 11.54 bin -> usr/bin
dr-xr-xr-x.   7 root root 4096  4. lis 18.40 boot
drwxr-xr-x.  20 root root 4660 10. pro 09.31 dev
drwxr-xr-x.   1 root root 4716  8. pro 16.22 etc
drwxr-xr-x.   1 root root   16 22. zář 11.54 home
lrwxrwxrwx.   1 root root    7 22. zář 11.54 lib -> usr/lib
lrwxrwxrwx.   1 root root    9 22. zář 11.54 lib64 -> usr/lib64
drwx------.   1 root root    0 29. říj  2020 lost+found
drwxr-xr-x.   1 root root    0 22. zář 11.54 media
drwxr-xr-x.   1 root root   30 22. zář 11.54 mnt
drwxr-xr-x.   1 root root    0 22. zář 11.54 opt
dr-xr-xr-x. 389 root root    0 12. lis 10.25 proc
dr-xr-x---.   1 root root  206  1. pro 15.59 root
drwxr-xr-x.  51 root root 1340 10. pro 06.28 run
lrwxrwxrwx.   1 root root    8 22. zář 11.54 sbin -> usr/sbin
drwxr-xr-x.   1 root root    0 22. zář 11.54 srv
dr-xr-xr-x.  13 root root    0 12. lis 10.25 sys
drwxrwxrwt.  30 root root  780 10. pro 13.28 tmp
drwxr-xr-x.   1 root root  100  1. říj 10.29 usr
drwxr-xr-x.   1 root root  200  1. říj 10.36 var

$ cd /bin

$ pwd
/usr/bin

But in reality, the result is /bin.

Just as an example this is place, where previously used ln -s needed to be replaced by cp -a when require_relative was introduced into RDoc. This is the stuff we need to do in Fedora to be able to execute the test suite. Having multiple copies of test suite around is not better then symlink.

[voxik]

Forgot to mention that I have reported this ticket yesterday to solve the issue:

https://bugs.ruby-lang.org/issues/18401

The idea is to really make require_relative to behave the same as require without expanding the symlinks

[deivid-rodriguez]

We have been expanding $LOAD_PATH entries in other places for a long time when we need to compare them with other paths, in order to fix issues with symlinked rubies. For example, here:

rubygems/bundler/lib/bundler/shared_helpers.rb

Lines 315 to 331 in a969cf0

	def clean_load_path
	loaded_gem_paths = Bundler.rubygems.loaded_gem_paths
	$LOAD_PATH.reject! do |p|
	resolved_path = resolve_path(p)
	next if $LOADED_FEATURES.any? {|lf| lf.start_with?(resolved_path) }
	loaded_gem_paths.delete(p)
	end
	$LOAD_PATH.uniq!
	end
	def resolve_path(path)
	expanded = File.expand_path(path)
	return expanded unless File.respond_to?(:realpath) && File.exist?(expanded)
	File.realpath(expanded)
	end

. I think it was added because otherwise bundler was crashing when run on some rvm gemsets that use symlinks.

In general, I think avoiding runtime crashes or redefinition warnings is a higher priority than potentially causing confusion when symlinked rubies are used.

I updated the test so that it hopefully works in your environment now.

[voxik]

I updated the test so that it hopefully works in your environment now.

Yes, the test passes and it seems to fix also the issues #5156 👍 Going to apply this fix to Ruby in Fedora in a bit.