[BUG] threadhist example fails

[zhipenggong]

Try the example according to example/gpu/threadhist/README.md,
Run example/gpu/threadhist/vec_add standalone can succeed, but run it with bpftime fails with the following error.
Did I miss something here ?

I use one ubuntu 22.04 docker image to build bpftime

$ BPFTIME_LOG_OUTPUT=console LD_PRELOAD=build/runtime/agent/libbpftime-agent.so example/gpu/threadhist/vec_add
[2025-10-21 16:54:22.541] [info] [bpftime_shm_internal.cpp:927] Registered shared memory with CUDA: addr=0x7fbb3592c000 size=20971520
[2025-10-21 16:54:22.541] [info] [bpftime_shm_internal.cpp:755] Global shm constructed. shm_open_type 1 for bpftime_maps_shm
[2025-10-21 16:54:22.541] [info] [bpftime_shm_internal.cpp:42] Global shm initialized
[2025-10-21 16:54:22][info][3109801] Initializing CUDA shared memory
[2025-10-21 16:54:23][info][3109801] CUDA context created
[2025-10-21 16:54:23][info][3109801] bpf_attach_ctx constructed
[2025-10-21 16:54:23][info][3109818] CUDA watcher thread started
[2025-10-21 16:54:23][info][3109801] Register attach-impl defined helper bpf_get_func_arg, index 183
[2025-10-21 16:54:23][info][3109801] Register attach-impl defined helper bpf_get_func_ret_id, index 184
[2025-10-21 16:54:23][info][3109801] Register attach-impl defined helper bpf_get_retval, index 186
[2025-10-21 16:54:23][info][3109801] Starting nv_attach_impl
[2025-10-21 16:54:23][info][3109801] Initializing agent..
[2025-10-21 16:54:23][info][3109801] Instantiating bpf link 81 and the corresponding program cuda__retprobe is cuda program
[2025-10-21 16:54:23][info][3109801] Skipping nv attach handler 81 since we are not handling nv handles
[2025-10-21 16:54:23][info][3109801] Main initializing for handlers done, try to initialize cuda link handles....
[2025-10-21 16:54:23][info][3109801] Instantiating bpf link 81 and the corresponding program cuda__retprobe is cuda program
[2025-10-21 16:54:23][info][3109801] Handling link to CUDA program: 81, recording it..
[2025-10-21 16:54:23][info][3109801] Loaded 13 instructions (original) for cuda ebpf program
[2025-10-21 16:54:23][info][3109801] Initializing nv_gpu_array_map_impl at pid 3109801
[2025-10-21 16:54:23][info][3109801] Mapped GPU memory for gpu array map: 140432935747584
[2025-10-21 16:54:23][info][3109801] Copying map fd 38 to device, key size=4, value size=8, max ent=1, map_type = 1502, gpu_buffer = 7fb917400000
[2025-10-21 16:54:23][info][3109801] Recording kretprobe for _Z9vectorAddPKfS0_Pf
[2025-10-21 16:54:23][info][3109801] Attach successfully
[2025-10-21 16:54:23][info][3109801] Got CUBIN section header size = 16, size = 3312
[2025-10-21 16:54:23][info][3109801] Finally size = 3328
[2025-10-21 16:54:23][info][3109801] Temporary fatbin written to /tmp/bpftime-fatbin-store.sQZniF/temp.fatbin
[2025-10-21 16:54:23][info][3109801] Listing functions in the patched ptx
terminate called after throwing an instance of 'boost::process::process_error'
what(): execve failed: No such file or directory
Aborted (core dumped)

[yunwei37]

Hi @zhipenggong, thanks for reporting!

Have you installed nvcc in your computer?
This exec here is for nvcc. You also need to add nvcc to your path.

We do need to make the error notice more friendly and do not abort the process, that's what we need to fix.

[zhipenggong]

@yunwei37 Thank you very much for the help, nvcc was installed and added to the path

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Wed_Jan_15_19:20:09_PST_2025
Cuda compilation tools, release 12.8, V12.8.61
Build cuda_12.8.r12.8/compiler.35404655_0

$ LD_PRELOAD=build/runtime/agent/libbpftime-agent.so example/gpu/threadhist/vec_add
[2025-10-22 08:53:41.497] [info] [bpftime_shm_internal.cpp:927] Registered shared memory with CUDA: addr=0x7f5f2b75e000 size=20971520
[2025-10-22 08:53:41.497] [info] [bpftime_shm_internal.cpp:755] Global shm constructed. shm_open_type 1 for bpftime_maps_shm
[2025-10-22 08:53:41.497] [info] [bpftime_shm_internal.cpp:42] Global shm initialized
[2025-10-22 08:53:41][info][3856950] Initializing CUDA shared memory
[2025-10-22 08:53:42][info][3856950] CUDA context created
[2025-10-22 08:53:42][info][3856950] bpf_attach_ctx constructed
[2025-10-22 08:53:42][info][3856977] CUDA watcher thread started
[2025-10-22 08:53:42][info][3856950] Register attach-impl defined helper bpf_get_func_arg, index 183
[2025-10-22 08:53:42][info][3856950] Register attach-impl defined helper bpf_get_func_ret_id, index 184
[2025-10-22 08:53:42][info][3856950] Register attach-impl defined helper bpf_get_retval, index 186
[2025-10-22 08:53:42][info][3856950] Starting nv_attach_impl
[2025-10-22 08:53:42][info][3856950] Initializing agent..
[2025-10-22 08:53:42][info][3856950] Instantiating bpf link 81 and the corresponding program cuda__retprobe is cuda program
[2025-10-22 08:53:42][info][3856950] Skipping nv attach handler 81 since we are not handling nv handles
[2025-10-22 08:53:42][info][3856950] Main initializing for handlers done, try to initialize cuda link handles....
[2025-10-22 08:53:42][info][3856950] Instantiating bpf link 81 and the corresponding program cuda__retprobe is cuda program
[2025-10-22 08:53:42][info][3856950] Handling link to CUDA program: 81, recording it..
[2025-10-22 08:53:42][info][3856950] Loaded 13 instructions (original) for cuda ebpf program
[2025-10-22 08:53:42][info][3856950] Initializing nv_gpu_array_map_impl at pid 3856950
[2025-10-22 08:53:42][info][3856950] Mapped GPU memory for gpu array map: 140037664538624
[2025-10-22 08:53:42][info][3856950] Copying map fd 38 to device, key size=4, value size=8, max ent=1, map_type = 1502, gpu_buffer = 7f5d0f400000
[2025-10-22 08:53:42][info][3856950] Recording kretprobe for _Z9vectorAddPKfS0_Pf
[2025-10-22 08:53:42][info][3856950] Attach successfully
[2025-10-22 08:53:42][info][3856950] Got CUBIN section header size = 16, size = 3312
[2025-10-22 08:53:42][info][3856950] Finally size = 3328
[2025-10-22 08:53:42][info][3856950] Temporary fatbin written to /tmp/bpftime-fatbin-store.4q884q/temp.fatbin
[2025-10-22 08:53:42][info][3856950] Listing functions in the patched ptx
terminate called after throwing an instance of 'boost::process::process_error'
what(): execve failed: No such file or directory
Aborted (core dumped)

[zhipenggong]

The issue is resovled after export /usr/local/cuda-12.8/bin to PATH, export /usr/loca/cuda/bin to PATH not work.

[yunwei37]

Thanks for confirming!

Let's keep it open so we can make the error message more friendly in the future 🙂