SIGSEGV on call to cuvidMapVideoFrame. How to debug?

Graphics / Linux Linux
1

I have a program that uses GStreamer with an NVDEC plugin. Intermittently, this program receives a segmentation fault from within libcuda.so, when it calls cuvidMapVideoFrame. There are zero diagnostic messages or anything, libcuda.so has all symbols stripped, the assembler code looks inconspicuous

7FFFF2A58898: E8 43 00 F3 FF             callq  0x7ffff29888e0  ; symbol stub for: pthread_mutex_unlock
7FFFF2A5889D: 48 8B 5C 24 38             movq   0x38(%rsp), %rbx
7FFFF2A588A2: 48 89 DF                   movq   %rbx, %rdi
7FFFF2A588A5: FF 93 A0 00 00 00          callq  *0xa0(%rbx)
7FFFF2A588AB: 48 8B 6C 24 40             movq   0x40(%rsp), %rbp
7FFFF2A588B0: 48 8B 43 18                movq   0x18(%rbx), %rax
7FFFF2A588B4: 4C 8B 00                   movq   (%rax), %r8
7FFFF2A588B7: 41 8B 08                   movl   (%r8), %ecx

The address in r8 looks legit.

Dear NVIDIA, how I can find where the problem is?

2

Found some more information. cuvidMapVideoFrame eventually calls cuLaunchKernel, the seg fault happens inside one of the subroutines of the latter.