KvikIO: Need to canonicalize dlopen'd library names

[jakirkham]

Documenting an offline discussion with @jakirkham in preparation of CUDA 12 bring-up on conda-forge.

Currently, there're some places where libXXXXX.so without any SONAME being dlopen'd, for example,

kvikio/cpp/include/kvikio/shim/cufile.hpp

Line 53 in db0a3e7

void* lib = load_library({"libcufile.so.1", "libcufile.so.0", "libcufile.so"});

There will become problematic because the libXXXXX.so symlink is supposed to exist only in the libXXXXX-dev package by the stand practice of Linux distros, not in libXXXXX which only contains libXXXXX.so.{major}. The dlopen'd names need to be canonicalized.

Note: Text borrowed from Leo's issue ( rapidsai/cudf#12708 ) and tweaked

[jakirkham]

Some comments:

• It appears we load libcufile.so.1. Where have we seen that in use?
• AFAICT libcufile.so.0 is the typical name. So this looks good.
• It looks like we fallback to libcufile.so. Does that show up in places where libcufile.so.0 isn't present?

[madsbk]

I have seen all of them in the wild :)
On my current workstation, I have libcufile.so, libcufile.so.0, libcufile.so.1.4.0, all pointing to the same library

[jakirkham]

Have seen those 3:

• Unversioned symlink ( libcufile.so ) - used in third-party development (not typically in runtime)
• SOVERSION library ( libcufile.so.0 ) - used by third-party tools/libraries at runtime
• Release versioned library ( libcufile.so.1.4.0 ) - (guessing?) for management of cuFile installed on system

Though haven't seen the SOVERSION bumped ( libcufile.so.1 ). Is that something that is showing up? If so, are there any more details about when it shows up?

[madsbk]

Sorry, I can't remember when I saw libcufile.so.1 :/

[wence-]

We need to dlopen libcufile.so.0 and no other. libcufile.so is installed in "dev" mode always pointing at the most recent "ABI-versioned" .so. That in turn symlinks to the actual versioned library. If both libcufile.so.1 and libcufile.so.0 were to exist (I do not believe that they do) they are (by definition) ABI-incompatible. This would mean that it is not safe to use a library built against an old version of libcufile with a new version (because, say, struct layouts will have changed).