Wrong rpath after repair

[ozon2]

I am using auditwheel 3.2.0 on manylinux2014_x86_64 docker and after running auditwheel repair on my wheel the rpath of a shared library is set to a wrong value.

When i install the wheel created by auditwheel i get this tree:

/usr/local/lib/python3.6/site-packages/
├── example_adder
│   ├── __init__.py
│   └── libexample_adder.so
├── gepetto_example_adder.libs
│   ├── libboost_python3-mt-1ace0776.so.1.53.0
│   └── libexample-adder-f1ca4966.so.3.0.2-1-g0463-dirty

So the rpath of libexample_adder.so should be set to $ORIGIN/../gepetto_example_adder.libs, however auditwheel set a wrong rpath:

readelf -d /usr/local/lib/python3.6/site-packages/example_adder/libexample_adder.so| grep 'R.*PATH'
 0x000000000000000f (RPATH)              Library rpath: [$ORIGIN/../../../../../../gepetto_example_adder.libs]

[lkollar]

This might be related to #245. Could you check if it happens with 3.1.1 as well? Is the code available somewhere to try to reproduce the issue?

[ozon2]

It happens on 3.1.1 as well.

Here is how to create the same wheel:

git clone --recursive https://github.com/Ozon2/example-adder.git
cd example-adder
git checkout dad2314
docker build . -t example-adder-manylinux
docker run --rm -it -e PLAT=manylinux2014_x86_64 -v `pwd`:/io example-adder-manylinux bash
./build_wheels.sh

I can't test it directly on the manylinux docker because for some reason it installs the shared libraries in /usr/local/lib64/python3.6/site-packages/gepetto_example_adder.libs instead of /usr/local/lib/python3.6/site-packages/gepetto_example_adder.libs.

You can reproduce the issue on another docker with:

docker run --rm -it -v `pwd`:/io python:3.6.11-buster bash
/io/test_wheels.sh
readelf -d /usr/local/lib/python3.6/site-packages/example_adder/libexample_adder.so| grep 'R.*PATH'
find / -name gepetto_example_adder.libs

[mayeut]

Auditwheel is putting grafted libraries at the root of the wheel:

auditwheel/auditwheel/repair.py

Lines 47 to 58 in 42a1e56

	with InWheelCtx(wheel_path) as ctx:
	ctx.out_wheel = pjoin(out_dir, wheel_fname)
	match = WHEEL_INFO_RE(wheel_fname)
	if not match:
	raise ValueError("Failed to parse wheel file name: %s",
	wheel_fname)
	dest_dir = match.group('name') + lib_sdir
	if not exists(dest_dir):
	os.mkdir(dest_dir)

It assumes 2 things:

• 'Root-Is-Purelib' == False: this seems OK as it changes this value to False when deemed necessary
• All modules are placed in root: this is wrong. According to PEP427, there are multiple places where things can go and this is where things go badly here.

Files can go in:

• root with destination being defined by Root-Is-Purelib (always False) so platlib
• {distribution}-{version}.data/(purelib|platlib|headers|scripts|data)

Here, skbuild has chosen to put all files in {distribution}-{version}.data/data e.g. ./gepetto_example_adder-3.0.2.data/data/lib/python3.7/site-packages/example_adder/libexample_adder.so.
This means RPATH is computed between ./gepetto_example_adder-3.0.2.data/data/lib/python3.7/site-packages/example_adder/libexample_adder.so and ./gepetto_example_adder.libs/libexample-adder-21dabdd7.so.3.0.2-6-gd319 instead of ./example_adder/libexample_adder.so and ./gepetto_example_adder.libs/libexample-adder-21dabdd7.so.3.0.2-6-gd319

That being said, I don't know the exact implications between the different destination directories allowed and how auditwheel should manage them. If every grafted library is grafted for only one destination directory, should be pretty easy to manage however, if it crosses multiple directories, that's another story. (I'm specifically thinking of the scripts vs platlib which I guess could be pretty common).

[brenthuisman]

Wrote a small script that patches the rpaths in our wheels post auditwheel. In case you're interested, it's easily adapted for your situation: https://github.com/arbor-sim/arbor/blob/3dc2f266ecb09e535ab35d878fd3a8375e7d2fa8/scripts/patchwheel.py