cd7e2f8e36
Certain applications (pytorch) decide that it's a great idea to distribute whole CUDA driver and link to it with DT_RPATH. This igores LD_LIBRARY_PATH. This code defeats that evil mechanism through any means necessary
1.1 KiB
This crate is a last resort Linux-specific solution.
Most of the time we can inject ourselves into a process by having users
set LD_LIBRARY_PATH.
Unfortunately, there is software out there which dynamically links to CUDA and
CUDA performance libraries using RPATH. On Linux, dynamic linker operates
using approximately this algorithm:
- If path contains
/treat the name as a (possibly relative) path and just use it - Otherwise return the first that succeeds:
- Library with this name already loaded into the process
- Try paths in
DT_RPATH(ifDT_RUNPATHis not present) - Try paths in
LD_LIBRARY_PATH - Try paths in
DT_RUNPATH - Try system paths
In order to defeat DT_RPATH this library needs to be preloaded with LD_PRELOAD.
On initialization we also preload all the performance libraries. We also hijack
dlopen and on every call to dlopen that tries to open a CUDA library we
redirect it to our libraries
We also expose zluda_dlopen_noredirect for the purpose of tracing libraries
so they can load real underlying library and not just get redirected to themselves