Add: memfd-based SO loading for all runtimes

[puddingfjz]

Summary

Add memfd-based SO loading mechanism for AICPU orchestration SO files across all runtimes.

Changes

• Add memfd_loader.h: In-memory SO loading using memfd_create()
• Update AICPU executors to use memfd loading instead of file-based loading
• Eliminates temporary file pollution in /tmp directory
• Kernel-managed cleanup via MFD_CLOEXEC

Files Modified

• src/a2a3/runtime/tensormap_and_ringbuffer/aicpu/memfd_loader.h (new)
• src/a2a3/runtime/tensormap_and_ringbuffer/aicpu/aicpu_executor.cpp
• src/a5/runtime/tensormap_and_ringbuffer/aicpu/memfd_loader.h (new)
• src/a5/runtime/tensormap_and_ringbuffer/aicpu/aicpu_executor.cpp
• src/a2a3/runtime/aicpu_build_graph/aicpu/memfd_loader.h (new)
• src/a2a3/runtime/aicpu_build_graph/aicpu/aicpu_executor.cpp

Test plan

• Tested on a2a3 hardware with tensormap_and_ringbuffer and aicpu_build_graph runtime
• Performance profiled: ~5% faster core loading for tensormap_and_ringbuffer and 7.55% faster for aicpu_build_graph.

[gemini-code-assist]

Code Review

This pull request introduces a memory file descriptor (memfd) based mechanism for loading orchestration shared objects, aimed at bypassing dynamic linker issues in AICPU environments. The changes include the addition of a memfd_loader.h utility and updates to aicpu_executor.cpp across several runtime paths to implement a 'memfd-first' loading strategy with a file-based fallback. Review feedback highlights opportunities to improve debuggability by adding logging for failure cases in the new loader and ensuring logging consistency across the various versions of the executor modified in this PR.

[gemini-code-assist]

To improve debuggability, it would be beneficial to add logging for failure cases within this function. When memfd_create, write, symlink, or dlopen fail, logging the error (e.g., from errno or dlerror()) would provide valuable insight before falling back to the file-based method. This would also make the behavior consistent with the fallback path, which does have logging for its failures.

Suggested change

	if (handle == nullptr) {
	close(fd);
	return -1;
	}
	if (handle == nullptr) {
	const char *dl_err = dlerror();
	DEV_INFO("Thread %d: dlopen from memfd symlink failed: %s, falling back to file-based loading",
	orch_thread_num, dl_err ? dl_err : "unknown error");
	close(fd);
	return -1;
	}

References

1. Always check the return values of resource allocation functions (like pthread_key_create) and handle potential failures gracefully, for example, by rolling back partial allocations.

[gemini-code-assist]

The logging for file creation and writing has been removed from this loop, which is inconsistent with the other aicpu_executor.cpp files in this PR. Please restore the DEV_INFO logs to help with debugging the file-based fallback path.

                        snprintf(so_path, sizeof(so_path), "%s/libdevice_orch_%d.so", candidate_dirs[i], getpid());
                        int32_t fd = open(so_path, O_WRONLY | O_CREAT | O_TRUNC, 0755);
                        if (fd < 0) {
                            DEV_INFO(
                                "Thread %d: Cannot create SO at %s (errno=%d), trying next path", thread_idx, so_path, errno
                            );
                            continue;
                        }
                        ssize_t written = write(fd, so_data, so_size);
                        close(fd);

                        if (written != static_cast<ssize_t>(so_size)) {
                            DEV_INFO(
                                "Thread %d: Cannot write SO to %s (errno=%d), trying next path", thread_idx, so_path, errno
                            );
                            unlink(so_path);
                            continue;
                        }
                        file_created = true;
                        DEV_INFO("Thread %d: Created SO file at %s (%zu bytes)", thread_idx, so_path, so_size);

References

1. For files shared across different runtimes, maintain identical code for consistency, even if a more robust implementation is possible.

[gemini-code-assist]

A log message for a successful dlopen in the fallback path is missing here, while it's present in the other aicpu_executor.cpp files. For consistency and better debugging, please add it back.

                    if (handle == nullptr) {
                        DEV_ERROR("Thread %d: dlopen failed: %s", thread_idx, dlopen_err ? dlopen_err : "unknown");
                        unlink(so_path);
                        return -1;
                    }
                    DEV_INFO("Thread %d: dlopen succeeded, handle=%p", thread_idx, handle);

References

1. For files shared across different runtimes, maintain identical code for consistency, even if a more robust implementation is possible.