Bug #9760: mkmf does not allow for linking against custom libraries when a system library is present

• Status: Open
• Assignee: -
• Target version: -
• ruby -v: ruby 2.1.1p76 (2014-02-24 revision 45161) [x86_64-linux]
• Backport: 2.0.0: UNKNOWN, 2.1: REQUIRED, 2.2: REQUIRED
• Tags: build

[zanegray (Andrew DeMaria)]

Hi,

Hopefully the title is not confusing, but the short story is that mkmf outputs a makefile that first searches the default lib path before searching any user provided lib paths. This is not an issue until one tries to link against an included library whose version is different than a preexisting system library.

The issue cropped up while trying to install the rugged gem (libgit2 wrapper) and a full dialog on the issue can be found on github https://github.com/libgit2/rugged/issues/351.

I was able to fix the issue with the attached patch (https://github.com/muff1nman/ruby/commit/a0c8bc32cfc11e61c5b9703bff243934c6509210)

[nobu (Nobuyoshi Nakada)]

• Status changed from Open to Closed
• % Done changed from 0 to 100

Applied in changeset r45640.

---

mkmf.rb: prefer $LIBPATH than $DEFLIBPATH

• lib/mkmf.rb (link_command, libpathflag, create_makefile): prefer
  user specified $LIBPATH than $DEFLIBPATH. [ruby-core:62100]
  [ruby-trunk - Bug #9760]

[knu (Akinori MUSHA)]

In the meantime, third parties can work around this problem by a monkey patch like this:

https://github.com/sparklemotion/nokogiri/commit/c98745d9098487c51685a732f1c5a21d8d07cdad

[knu (Akinori MUSHA)]

The monkey patch I mentioned above was backed out due to a serious problem found, that is, if an instance of libruby.so (of the same soname as the running ruby) is found in a user-given path that is different from the one for the running ruby used for build, it will be picked by the linker and the resulted extension will cause a SEGV in run time.

So, it turned out DEFLIBPATH took precedence over LIBPATH for a good reason and I'm afraid it should not have been simply changed this way.

[nagachika (Tomoyuki Chikanaga)]

• Backport changed from 2.0.0: UNKNOWN, 2.1: UNKNOWN to 2.0.0: UNKNOWN, 2.1: REQUIRED

Does this issue exist in 2.0.0 too?

[knu (Akinori MUSHA)]

Please do not backport this yet; I'm suggesting backing out the commit made in trunk for the reason I gave above.

[knu (Akinori MUSHA)]

• Status changed from Closed to Open

[sorah (Sorah Fukumori)]

Hit a problem as @knu (Akinori MUSHA) said: https://github.com/brianmario/mysql2/issues/657

[sorah (Sorah Fukumori)]

• File early-libdir.patch early-libdir.patch added

attached diff (early-libdir.patch) works well for my case... but I can't determine this is correct fix. If it's okay I'll commit into trunk.

[usa (Usaku NAKAMURA)]

• Backport changed from 2.0.0: UNKNOWN, 2.1: REQUIRED to 2.0.0: UNKNOWN, 2.1: REQUIRED, 2.2: REQUIRED

[sorah (Sorah Fukumori)]

Assume /usr/lib has librubyA and libfooB, and /home/someone/local/lib has libfooA. I think early-libdir.patch makes mkmf.rb can't link to libfooA never..

[sorah (Sorah Fukumori)]

I'm positive to revert the change at r45640.

[sorah (Sorah Fukumori)]

• Status changed from Open to Closed

Applied in changeset r52267.

---

• lib/mkmf.rb: Revert r45640 because it may lead to link
  with different libruby. [Bug #9760]

[sorah (Sorah Fukumori)]

• Status changed from Closed to Open

Reverted r45640 at r52267. Hmm...

[onur (Onur Küçük)]

We have a system where

• Ruby 1.8.7-p374 is installed system wide from distribution packages
• There is a user setup with RVM to run Ruby 2.x running Redmine with Mysql

During Redmine bundle install, mysql2 gem is linking against the system Ruby, leading to issues like missing symbols etc. for the generated SO. The issue is happening with 2.2.3, 2.2.4 and 2.3.0 but it seems it is fixed in 2.3.1. I did not bisect but r52267 is probably what fixed it, thank you !

[rhenium (Kazuki Yamaguchi)]

This is still an issue in the current master (e.g., https://github.com/ruby/openssl/issues/545).

If I understand correctly, mkmf.rb prepends libruby's path solely to prevent picking up a wrong libruby. Since the exact path of the correct libruby is known, I wonder if it's possible for mkmf.rb to just use it instead of the -L[...] -lruby combo to achieve the same thing.

BTW, Python 3.8 made an interesting change in https://bugs.python.org/issue21536: if the platform allows, extensions don't link to libpython (don't use -lpythonX.Y) and leave symbols unresolved until at runtime.