This is for discussion for now. I'm not sure where is the best place to put something like this.

@camierjs This adds a debug tool specifically to be used in conjunction with the debug memory backend. The idea is that if you get an MmuError while running your program with the debug backend, you can preload this library to find why that happens. The debug backend uses mprotect to enforce validity of memory between host and device which this library intercepts and tracks. The backend has some false positives currently, but as they are fixed I think a tool like this could make the debug backend even more useful for debugging GPU code. As far as I am aware, there aren't existing tools for this. Tools like nvidia's compute-sanitizer probably aren't useful in this case because it's just going to detect access to unallocated or uninitialized memory, not whether or not values got out of sync between host and device.

To use, set LD_PRELOAD=/path/to/mprotect_trace.so before executing your program.
If it is in the same directory as your present directory, you may need to use
LD_PRELOAD=./mprotect_trace.so

It may be possible to rewrite this using libunwind, but I found it easier to get informative traces with libbacktrace.

Limitations:

• Right now I only have this working with the cmake build. This requires linux, libbacktrace, and I am using c++17. I tested with gcc, but it may also work with clang.
• I have to probably refine the logic around the lock guards for thread safety, but that wasn't my focus.
• This finds the offending protection call based on an address, memory range, and most recent timestamp meeting those parameters. If the timestamp isn't resolved enough, it's possible that multiple calls could meet the same criteria. This seems like an unlikely case, but it could be eliminated with more work.