Fix GPU-CPU memory synchronization and illegal memory access on NVIDI…

[dr-xjx]

…A Orin X

This commit addresses critical memory consistency and access issues when running on NVIDIA Tegra platforms (Orin/Thor).

Changes:

1. Add volatile qualifier to CommSharedMem members

   • Mark flag1, flag2, request_id, and map_id as volatile
   • Prevents compiler optimizations that could cause data inconsistency
   • Ensures each access reads from/writes to actual memory

2. Use ld.volatile.global.u32 for flag2 in make_helper_call

   • Forces memory load on each iteration in spin-wait loop
   • Bypasses L1 cache to ensure visibility of CPU writes
   • Critical for cross-device synchronization

3. Migrate to cudaHostAlloc for CommSharedMem allocation

   • Replaces standard library memory allocation (malloc/new)
   • Uses pinned memory to resolve CUDA_ERROR_ILLEGAL_ADDRESS
   • Follows NVIDIA Memory Management guidelines for Tegra platforms
   • Reference: https://docs.nvidia.com/cuda/cuda-for-tegra-appnote/index.html#memory-management

4. Align CommSharedMem lifetime with bpftime_attach_ctx

   • Implement RAII pattern using std::unique_ptr with custom deleter
   • Ensures proper resource cleanup on context destruction

Root Cause:
On NVIDIA Orin/Thor, GPU access to constData allocated via standard library functions triggers illegal memory access errors. The error occurs in _bpf_helper_ext_0006 when executing:
...
ld.const.u64 %rd1, [constData];
...
add.s64 %rd27, %rd1, %rd43;
st.u8 [%rd27+24], %rs1;

Error: CUDA_ERROR_ILLEGAL_ADDRESS (error code: 700)

Solution:
Using cudaHostAlloc with pinned memory ensures proper memory mapping between CPU and GPU address spaces on integrated GPU systems.

Testing:

• Platform: NVIDIA Orin X
• CPU: ARM Cortex-A78AE (12-core)
• CUDA Driver/Runtime: 12.1 / 11.4
• Compute Capability: 8.7
• Memory: 25767 MBytes
• Test case: example/gpu/cuda-counter
• Status: Verified working

Please try to use the copilot to summary your PR. You don't need to fill all info below, just it can help giving your a checklist.

Description

Fixes # (issue)

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

How Has This Been Tested?

• Test A
  example/gpu/cuda-counter
• Test B

Test Configuration:

• Platform: NVIDIA Orin X
• CPU: ARM Cortex-A78AE (12-core)
• CUDA Driver/Runtime: 12.1 / 11.4
• Compute Capability: 8.7
• Memory: 25767 MBytes
• Test case: example/gpu/cuda-counter
• Status: Verified working
• Firmware version:
• Hardware:
• Toolchain:
• SDK:

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• Any dependent changes have been merged and published in downstream modules
• I have checked my code and corrected any misspellings

[dr-xjx]

@yunwei37 please review this，tks~

[dr-xjx]

@copilot

[Copilot]

Pull Request Overview

This PR addresses critical GPU-CPU memory synchronization issues on NVIDIA Tegra platforms (Orin/Thor) by implementing proper memory management and synchronization primitives to fix illegal memory access errors and data consistency problems.

Key Changes:

• Migrated from standard memory allocation to CUDA pinned memory (cudaHostAlloc) to resolve illegal memory access on integrated GPU systems
• Added volatile qualifiers to shared memory synchronization flags to prevent compiler optimizations
• Used ld.volatile.global.u32 PTX instruction to ensure cache-bypassing memory reads in GPU spin-wait loops

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 12 comments.

Show a summary per file

File	Description
runtime/src/attach/bpf_attach_ctx_cuda.cpp	Implements CUDAContext with cudaHostAlloc-based memory allocation and worker thread management
runtime/src/attach/bpf_attach_ctx.cpp	Removes explicit thread lifecycle management (now handled by CUDAContext)
runtime/include/bpf_attach_ctx.hpp	Adds volatile qualifiers to CommSharedMem and refactors CUDAContext class structure
attach/nv_attach_impl/trampoline/default_trampoline.cu	Adds volatile qualifiers and uses ld.volatile.global.u32 for cache-bypassing reads
attach/nv_attach_impl/nv_attach_impl_frida_setup.cpp	Adds additional debug logging for kernel launch parameters
attach/nv_attach_impl/nv_attach_fatbin_record.cpp	Corrects error message text for clarity

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[yunwei37]

Thanks a lot for the PR!

It seems like it fix many bugs in one PR, would you mind seperate them into mmultiple?

Use ld.volatile.global.u32 for flag2 in make_helper_call

Also, does it has performance impact? If yes, how can we improve it while keep it correct?

@Officeyutong @Sy0307 What do you think?

[dr-xjx]

Thanks a lot for the PR!

It seems like it fix many bugs in one PR, would you mind seperate them into mmultiple?

Use ld.volatile.global.u32 for flag2 in make_helper_call

Also, does it has performance impact? If yes, how can we improve it while keep it correct?

@Officeyutong @Sy0307 What do you think?
These changes should be kept as a cohesive whole, and splitting them isn't really necessary.

I believe this PR should have no major impact, but there are minor considerations:

1. Memory Allocation: cudaHostAlloc requires contiguous physical memory allocation, which may fail when system memory is fragmented or under pressure

2. Instruction Overhead: ld.volatile is unnecessary for uncached memory access scenarios, but the performance impact is negligible

[Officeyutong]

Please fix the CI

[dr-xjx]

Please fix the CI
@Officeyutong @yunwei37
How should I debug CI? In my environment, can trampoline_ptx.h be compiled normally.
I have tried many times, but still haven't been able to fix CI

[dr-xjx]

@yunwei37 please review again,tks~

[yunwei37]

Thanks! We are testing it locally these days