segfault or free(): invalid pointer when importing dgl with other libraries due to RTLD_GLOBAL

[skrsna]

🐛 Bug

importing dgl after importing C++ based library with pybind interface leads to segfault or free(): invalid pointer. The C++ library in question is an internal library that is not available publicly. I found some relevant issues on pytorch repo pytorch/pytorch#3059 and RobotLocomotion/drake#12073. I was able to find a workaround by deleting ctypes.RTLD_GLOBAL here in the dgl source code. Pytorch and tensorflow seemed to move away from RTLD_GLOBAL. ref (pytorch/pytorch#28536). Just wondering if something similar can be done in dgl.

To Reproduce

Sorry the library I'm using that causes this error is not available publicly and uses TBB allocator.
Steps to reproduce the behavior:

Expected behavior

import dgl without segfault or free(): invalid pointer Aborted

Environment

• DGL Version (e.g., 1.0): 0.5.2 cpu
• Backend Library & Version (e.g., PyTorch 0.4.1, MXNet/Gluon 1.3): pytorch 1.7 nightly
• OS (e.g., Linux): linux
• How you installed DGL (conda, pip, source): conda
• Build command you used (if compiling from source):
• Python version: 3.8
• CUDA/cuDNN version (if applicable): 10.2, 7.6
• GPU models and configuration (e.g. V100):
• Any other relevant information:

Additional context

[VoVAllen]

bump this

[BarclayII]

Does it immediately crash after importing DGL after importing the said library?

[skrsna]

Hi @BarclayII,

If I import dgl then import the private library, it doesn't crash right away but crashes when there's a call to any dgl functions or the library's functions. On the other hand if I import the library first and then dgl it crashes right away.

[dgasmith]

This issue was references in #2328, but then the line was crossed out. I didn't see an immediate reason of why in the issue.

If this is a longer term item, could we introduce an env variable to dynamically change the CDLL load in specific circumstance-- perhaps DGL_RTLD_SETTING?

[BarclayII]

As I mentioned in the crossed-out text, directly changing RTLD_GLOBAL will make some examples (namely examples/pytorch/graphsage/train_sampling.py with num_workers=0) freeze.

I wasn't able to figure out the reason yet, so I had to work around it by ensuring PyTorch/MXNet/Tensorflow C library to be loaded before libdgl.so. Obviously not a fix to this issue per se.

[dgasmith]

Ah got it. In the meantime would take a PR that allows us to alter this setting via env variable?

[BarclayII]

@dgasmith @skrsna I removed the flag in a recent PR. Please give the nightly builds a try. I tested on the GraphSAGE examples and it currently run without any problems.

[dgasmith]

@BarclayII Thanks! I really appreciate that, we will evaluate the nightly builds ASAP.

[BarclayII]

So far no issues as per our experience. Please reopen the issue if the problem still exists in your case.