SIGSEGV in Device::setupCpuAgent

[samcv]

When running clinfo (system or ROCm version) I get a segmentation fault. This is on tag rocm-3.7.0. My system is OpenSuse Tumbleweed. This happens if I install from the official ROCm SUSE repository, or if I compile rocclr and rocm-opencl-runtime myself the issue persists (with other packages as the official packages). Compiling those two with debug is how I got the nice backtrace.

This is with AMD Ryzen 7 PRO 2700U Vega using the APU (integrated graphics).

Thread 1 "clinfo" received signal SIGSEGV, Segmentation fault.
0x00007ffff7923266 in roc::Device::setupCpuAgent (this=0x4f3870) at /home/samantha/git/ROCm/ROCclr/device/rocm/rocdevice.cpp:170
170       cpu_agent_ = cpu_agents_[index].agent;
(gdb) bt
#0  0x00007ffff7923266 in roc::Device::setupCpuAgent (this=0x4f3870) at /home/samantha/git/ROCm/ROCclr/device/rocm/rocdevice.cpp:170
#1  0x00007ffff7926840 in roc::Device::populateOCLDeviceConstants (this=0x4f3870) at /home/samantha/git/ROCm/ROCclr/device/rocm/rocdevice.cpp:1047
#2  0x00007ffff7924d08 in roc::Device::create (this=0x4f3870) at /home/samantha/git/ROCm/ROCclr/device/rocm/rocdevice.cpp:593
#3  0x00007ffff79245a1 in roc::Device::init () at /home/samantha/git/ROCm/ROCclr/device/rocm/rocdevice.cpp:489
#4  0x00007ffff78b7390 in amd::Device::init () at /home/samantha/git/ROCm/ROCclr/device/device.cpp:194
#5  0x00007ffff78d2a53 in amd::Runtime::init () at /home/samantha/git/ROCm/ROCclr/platform/runtime.cpp:74
#6  0x00007ffff78a318b in ShouldLoadPlatform () at /home/samantha/git/ROCm/ROCm-OpenCL-Runtime/amdocl/cl_icd.cpp:222
#7  0x00007ffff78a3287 in operator() (__closure=0x7fffffffcdaf) at /home/samantha/git/ROCm/ROCm-OpenCL-Runtime/amdocl/cl_icd.cpp:272
#8  0x00007ffff78a366b in std::__invoke_impl<void, clIcdGetPlatformIDsKHR(cl_uint, _cl_platform_id**, cl_uint*)::<lambda()> >(std::__invoke_other, struct {...} &&) (__f=...) at /usr/include/c++/10/bits/invoke.h:60
#9  0x00007ffff78a363a in std::__invoke<clIcdGetPlatformIDsKHR(cl_uint, _cl_platform_id**, cl_uint*)::<lambda()> >(struct {...} &&) (__fn=...) at /usr/include/c++/10/bits/invoke.h:95
#10 0x00007ffff78a3531 in operator() (this=0x7fffffffcd50) at /usr/include/c++/10/mutex:717
#11 0x00007ffff78a355b in operator() (this=0x0) at /usr/include/c++/10/mutex:722
#12 0x00007ffff78a356c in _FUN () at /usr/include/c++/10/mutex:722
#13 0x00007ffff7c2e36f in __pthread_once_slow () from /lib64/libpthread.so.0
#14 0x00007ffff78a3084 in __gthread_once (__once=0x7ffff7a4b050 <clIcdGetPlatformIDsKHR::initOnce>, __func=0x7ffff7e754e0 <std::__once_proxy()>) at /usr/include/c++/10/x86_64-suse-linux/bits/gthr-default.h:700
#15 0x00007ffff78a35f2 in std::call_once<clIcdGetPlatformIDsKHR(cl_uint, _cl_platform_id**, cl_uint*)::<lambda()> >(std::once_flag &, struct {...} &&) (__once=..., __f=...) at /usr/include/c++/10/mutex:729
#16 0x00007ffff78a32e1 in clIcdGetPlatformIDsKHR (num_entries=0, platforms=0x0, num_platforms=0x7fffffffcdec) at /home/samantha/git/ROCm/ROCm-OpenCL-Runtime/amdocl/cl_icd.cpp:272
#17 0x00007ffff7fc236d in khrIcdVendorAdd (libraryName=0x442110 "libamdocl64.so") at /home/samantha/git/ROCm/ROCm-OpenCL-Runtime/khronos/icd/loader/icd.c:87
#18 0x00007ffff7fc5d3e in khrIcdOsVendorsEnumerate () at /home/samantha/git/ROCm/ROCm-OpenCL-Runtime/khronos/icd/loader/linux/icd_linux.c:125
#19 0x00007ffff7c2e36f in __pthread_once_slow () from /lib64/libpthread.so.0
#20 0x00007ffff7fc5dcc in khrIcdOsVendorsEnumerateOnce () at /home/samantha/git/ROCm/ROCm-OpenCL-Runtime/khronos/icd/loader/linux/icd_linux.c:149
#21 0x00007ffff7fc2272 in khrIcdInitialize () at /home/samantha/git/ROCm/ROCm-OpenCL-Runtime/khronos/icd/loader/icd.c:31
#22 0x00007ffff7fc27c3 in clGetPlatformIDs (num_entries=0, platforms=0x0, num_platforms=0x7fffffffcfa0) at /home/samantha/git/ROCm/ROCm-OpenCL-Runtime/khronos/icd/loader/icd_dispatch.c:34
#23 0x0000000000408216 in cl::Platform::get (platforms=0x7fffffffd370) at /home/samantha/git/ROCm/ROCm-OpenCL-Runtime/khronos/headers/opencl2.2/CL/cl2.hpp:2482
#24 0x0000000000403813 in main (argc=1, argv=0x7fffffffda98) at /home/samantha/git/ROCm/ROCm-OpenCL-Runtime/tools/clinfo/clinfo.cpp:75

[vsytch]

This has been fixed internally and the issue should be gone with the next ROCm release.

[samcv]

@vsytch good news. Is there a patch you can provide to fix this issue?

[faust3]

@samcv

I had the same problem with Ryzen 2400G and Ubuntu 20.04.

I don't know how the internal fix looks like but following change prevented the crash for me.

rocmPatch.patch.txt