Problem:
When TheRock-based PyTorch package is installed along with amdsmi, importing
torch causes a double-free crash on exit (GitHub issue ROCm/TheRock#2269).
Root cause:
Both librocm_smi64.so and libamd_smi.so export the C++ static member
'amd::smi::Device::devInfoTypesStrings'. When libraries are loaded with
RTLD_GLOBAL, the dynamic linker resolves libamd_smi.so's reference to this
symbol to the one in librocm_smi64.so. This causes:
1. librocm_smi64.so registers its destructor for devInfoTypesStrings
2. libamd_smi.so also registers a destructor, but for the SAME address
3. On exit, both destructors run on the same object -> double-free
Fix:
Change devInfoTypesStrings from a class static member to a file-local static
variable. This ensures the symbol has internal linkage and is not exported,
preventing the symbol collision.
Changes:
- rocm_smi_device.h: Remove static member declaration
- rocm_smi_device.cc: Change from 'Device::devInfoTypesStrings' to file-local
'static const std::map<...> devInfoTypesStrings'
- rocm_smi.cc: Remove the global alias to the (now removed) class member
Tested on gfx1151. `import torch` crashed on exit before the fix, and doesn't crash after the fix.
Matthias Gehre
1883f736ad