Fix a segmentation fault due to unloaded libcufile

[gigony]

Do not unload the cufile (GDS) library because libcufile registers a cleanup function with atexit() and unloading the library will cause a segfault (calling the cleanup function that doesn't exist anymore).

It turns out that CUDA 11.5 bundled GDS library (libcufile) and it is available in GPUCI's build/test container image (such as gpuci/rapidsai:21.12-cuda11.5-devel-ubuntu18.04-py3.7).
cuCIM would dynamically load libcufile.so shared library and unload it when a global static variable in cuCIM is destroyed.
Since libcufile's cleanup function(through atexit_thread handler) is registered after the libcufile is loaded, it causes a segmentation fault at exit if the libcufile is explicitly unloaded through dlclose().
(See #153)

You can see discussions with atexit in dynamically loaded shared library keywords (the actual root cause is the use of thread_local variable in libcufile.so).

• https://news.ycombinator.com/item?id=20228082
• https://forums.raspberrypi.com/viewtopic.php?t=237006

Maybe using destructor attribute could fix the issue from GDS(cufile) side.

This patch leaves the libcufile library loaded, without calling ::dlclose(library_handle) method to unload libcufile.so.

Update (2021-11-20): I couldn't find atexit() used in libcufile (though I can see atexit() call for an executable file[fio]) but it seems that a method is registered and called at the exit time so we cannot help but leave the dynamically loaded library without unloading.
Update (2021-11-23): libcufile.so started using thread_local variable since v1.1 which makes the shared library unloadable.
For this reason, this patch is a correct patch until libcufile is updated to make it possible.

Related information:

• https://sourceware.org/glibc/wiki/Destructor%20support%20for%20thread_local%20variables
• Thread local variable (thread_ec_ptr) prevents library to be unloaded emilk/loguru#146
• https://maskray.me/blog/2021-02-14-all-about-thread-local-storage
• https://www.spinics.net/lists/ceph-devel/msg42678.html

[gigony]

Need to take some time to dig the libcufile in detail if it is actually due to atexit() call, or cuCIM's destructor (by a static variable) is calling an unloaded method in libcufile, because I can see one atexit() method call but only in an executable code.

[gigony]

@jakirkham More update on the root-cause:

libcufile.so started using thread_local variable since v1.1 which makes the shared library unloadable.
For this reason, this patch is a correct patch until libcufile is updated to make it possible.

Related information:

• https://sourceware.org/glibc/wiki/Destructor%20support%20for%20thread_local%20variables
• Thread local variable (thread_ec_ptr) prevents library to be unloaded emilk/loguru#146
• https://maskray.me/blog/2021-02-14-all-about-thread-local-storage
• https://www.spinics.net/lists/ceph-devel/msg42678.html

[gigony]

GDS team suggested using RTLD_NODELETE when calling dlopen() so that the library is not unloaded.
They will fix the issue by adding -z nodelete in the link flags when building libcufile.so

dlopen("/usr/local/cuda/targets/x86_64-linux/lib/libcufile.so", RTLD_NOW | RTLD_LOCAL| RTLD_NODELETE);

see #177