Make rpath usage optional

[fabiand]

Currently (something after 0.5) rustc (IIUIC) is using rpaths to point libs and binaries to the appropriate locations. This is problematic when rustc and the other libs/bins are package e.g. for Fedora.
The rpath usage should be made optional - /etc/ld.so.conf.d can be used instead - and LD_LIBRARY_PATH.

[brson]

Why is this problematic for packaging? Is it because the rpath values end up with the wrong values?

Each crate's rpath has several values, one of which is supposed to be the final installation location (controlled with --prefix). Can we just provide more control over which locations are used during rpathing?

[fabiand]

In general Fedora recommends to not use rpaths [0], because the paths are hardcoded in the binary. Fedora preferres to point the dynamic linker to the appropriate paths [1].
But yes, we are more flexible about internal libraries - which the case is here. So yes, it would also help us if we could controll the rpath path (which goes a bit into the direction if issue #5223)

[0] https://fedoraproject.org/wiki/Packaging:Guidelines#Beware_of_Rpath
[1] https://fedoraproject.org/wiki/Packaging:Guidelines#Alternatives_to_Rpath

[fabiand]

The following rpath is used in e.g. the cargo binary:

ERROR   0002: file '/usr/bin/cargo' contains an invalid rpath '/home/fdeutsch/rpmbuild/BUILD/rust-0.6/x86_64-unknown-linux-gnu/stage1/lib/rustc/x86_64-unknown-linux-gnu/lib' in [$ORIGIN/../lib:/home/fdeutsch/rpmbuild/BUILD/rust-0.6/x86_64-unknown-linux-gnu/stage1/lib/rustc/x86_64-unknown-linux-gnu/lib:/usr/lib/rustc/x86_64-unknown-linux-gnu/lib]

The rpath referres to rpm's buildroot, where the lib won't stay (it will be instaled to --prefix)
Maybe the problem could adressed if the rpath would respect --prefix - but this could lead to problems at build time.
Maybe LD_LIBRARY_PATH can be used to point to the correct dir's at build time, and the rpath can be used to point to the correct dir's at runtime. Just my 2ct - but you ar ethe expert :)

[brson]

Thanks for those links. We do generate rpaths with the expectation that some of them will be invalidated by moving the binaries around.

It's possible we could consider removing rpath, but it sure is nice to not think about dynamic loading when running binaries out of my build directory.

We could easily add a --disable-rpath flag to the configure script, to turn off the rpath for all the rust-provided libraries, with the assumption that package managers would modify ld.so.conf.d during install. If we did that I feel like we would still want the installed rustc to generate rpaths with user-built libraries - otherwise the experience would be different depending on the rust installation.

We could also just remove rpath completely - it doesn't work on windows anyway.

Another thing to consider is that, if we just used rustpkg for all builds, and let rustpkg install all libraries at a known location, then we might be able to conveniently not use rpath while still doing the right thing at load time. Can user directories like ~/.rustpkg be put into ld.so.conf.d?

[fabiand]

Yes - I can fully understand how easy running rust in the build directory is when using rpaths :)
For the build directory use case it should be possible to create wrappers for rustc and friends which use LD_LIBRARY_PATH to point to the correct dirs.

You mention thet some libs might reside in ~/.rustpkg, could you explain the overall concept you are following with this?
In general I'd expect libraries to appear in /lib, /usr/lib some and some private /usr/lib/myapp. There are also some xdg specs that specify ~/.local/bin and ~/.local/lib - but let usfirst see what you say about ~/.rustpkg. Oh, I don't think that ~/.rustpkg would be expanded correctly if put into ld.so.conf.d

[fabiand]

Just to summarize my thoughts: If rpath is used, the rpath path beeing used should only use directories below --prefix (to point to non-standard dirs containing libs relevant for a rust component).
The rpath path should not point to any build-time dir.

To include libraries at buildtime or when using the binaries in the build directory I'd suggest using wrappers (isn't autotools doing it like this?) which set LD_LIBRARY_PATH appropriately.

[fabiand]

Any follow up thoughts?

[brson]

Our package manager will by default not be installing libraries to a system location and will instead put them in the user's directory, ~/.rustpkg. This doesn't seem out of the ordinary to me; consider cabal, etc.

The behavior you recommend requires some deep changes to how Rust works, so need to be considered carefully. Does the current behavior prevent packaging in Fedora, and is there a minimal change we can make to unblock packaging?

@graydon You'll probably be interested in this subject.

[fabiand]

Thanks for the rustpkg explanation.
Do I understand that rustpkg will create binaries and libs with rpaths pointing to the user's specific rustpkg dir ~/.rustpkg?
My feeling is that this user specific rpaths should be configurable at build time. The use case is obviously when the binary or lib is going to be installed systemwide. The only rpath exception is the rpath pointing to the system wide package specific libdir (e.g. /usr/lib64/rust).

I think that I'll spend a bit more time with looking at link.rs and rpath.rs and might get a better feeling for the problem.

Btw.: I noted that the libdir isn't configurabel. Fedora for example keeps x86_64 libs in (/usr)?/lib64, and i686 in (/usr)/lib. IIUIC this is currently not possible with rust as the libs will be installed to /usr/lib/arch-specific-path

[brson]

@fabiand The point about libdir is well taken and I updated #5223.

I am not entirely sure how rustpkg does it's build now or how @catamorphism and @graydon intend for it to work, but based on the current language design I would expect that, yes, binaries installed to ~/.rustpkg would have rpaths pointing there.

Currently there are several rpaths configured for each linked library so that they hopefully fall back to something useful when the libraries are moved around during install or during the bootstrapping process, something like:

• The relative path to the library
• The absolute path to where the library was at build time
• The system installation path

Under this scheme I would expect that rustpkg would, when linking against libraries in ~/.rustpkg, create rpaths to ~/.rustpkg.

[fabiand]

Yes, there are several rpaths used (seen in link.rs and rpath.rs) and I'd say - even with looking at rust itself, without the packaging glasses on - that it should be configurable what rpaths are used. Maybe it should be an option of rustc.

As said, system wide binaries are the best example for binaries where you don't want to many rpaths - including no rpaths pointing to some user dir.

[TJC]

Please don't tag TJC :(

[brson]

Sorry, TJC! Maybe it's time to give in and start working on Rust :) Seriously though, sorry. I'm sure it's a real PITA.