Fix CUDA watcher thread segfault at exit due to shared memory unmapped before thread stops

[Copilot]

The CUDA watcher thread (start_cuda_watcher_thread) polls ctx->cuda_shared_mem->flag1 in shared memory. On exit, __destruct_shm() (.fini_array, priority 65535) calls bpftime_destroy_global_shm() which unmaps the Boost IPC segment — including CommSharedMem — before the thread stops, causing a SIGSEGV. ~bpf_attach_ctx() never fires because bpf_attach_ctx lives in a union with an empty destructor, and the existing atexit handler's ordering relative to .fini_array is not guaranteed.

Changes

• bpf_attach_ctx_cuda.cpp: Expose bpftime::bpftime_stop_cuda_watcher_thread() — a non-anonymous-namespace wrapper around the existing stop_cuda_watcher_thread_at_exit(). Idempotent; safe to call from both atexit and __destruct_shm.

• bpftime_shm_internal.cpp: Call bpftime_stop_cuda_watcher_thread() in __destruct_shm() before bpftime_destroy_global_shm(), guaranteeing the thread is joined before the segment is unmapped regardless of atexit/fini_array ordering:

  static __attribute__((destructor(65535))) void __destruct_shm()
  {
      // ...
  #ifdef BPFTIME_ENABLE_CUDA_ATTACH
      bpftime::bpftime_stop_cuda_watcher_thread(); // NEW: join before unmap
  #endif
      bpftime_destroy_global_shm();
  }

• bpf_attach_ctx.cpp: Guard cuda_ctx->... in ~bpf_attach_ctx() with a null check to prevent dereference on a partially-constructed object.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>[BUG] CUDA agent segfault at exit caused by watcher thread</issue_title>
<issue_description>Issue description:

When running GPU workloads (e.g. benchmark/gpu/workload/vec_add.cu) with the bpftime CUDA agent attached, the process sometimes segfaults on exit.
The crash happens in the CUDA watcher thread, which polls ctx->cuda_shared_mem->flag1 in bpf_attach_ctx::start_cuda_watcher_thread().

Root Cause:

The agent process opens the global shared memory with bpftime_initialize_global_shm(shm_open_type::SHM_OPEN_ONLY).
In runtime/src/bpftime_shm_internal.cpp, the global destructor __destruct_shm() unconditionally calls bpftime_destroy_global_shm().

bpftime_shm::~bpftime_shm() unmaps the Boost shared memory segment and calls cudaHostUnregister(base_addr) for the whole segment, which includes the cuda::CommSharedMem holding flag1/flag2.

The CUDA watcher thread is started in bpf_attach_ctx::start_cuda_watcher_thread() and is detached; because bpf_attach_ctx is stored in a union (bpf_attach_ctx_holder) whose destructor is empty, ~bpf_attach_ctx() never runs at process exit, so cuda_watcher_should_stop is never set to true.</issue_description>

Comments on the Issue (you are @copilot in this section)

• Fixes [BUG] CUDA agent segfault at exit caused by watcher thread #521

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.