external-cannbot-ops-ascendc-crash-debug

Ascend C Operator Hang/Crash Debugging

Systematically debug issues where Ascend C operators cannot finish running, including program hang/suspension, Kernel timeout, and program crash.

Quick Diagnosis

程序无法运行完
    │
    ├─ 程序崩溃？
    │   ├─ 启用 coredump → ulimit -c unlimited
    │   ├─ 生成 core 文件 → 运行程序
    │   └─ GDB 分析 → gdb <exe> <core> → bt / bt full / info locals
    │
    └─ 程序卡死？
        ├─ 查看 plog 日志 → 定位卡死位置
        ├─ 检查 Buffer 配对 → AllocTensor/FreeTensor
        ├─ 检查同步 → EnQue/DeQue 配对
        └─ Kernel 调试 → AscendC::PRINTF / DumpTensor / msDebug

Symptom-Cause Quick Reference Table

Symptom	Possible Cause	Diagnosis Direction
Program hang/timeout	Buffer not released / deadlock	Check AllocTensor/FreeTensor pairing, EnQue/DeQue pairing
Core timeout/hang	Buffer conflict/deadlock	Check Alloc / Free pairing
Segmentation Fault	Null pointer dereference / memory out-of-bounds	GDB analyze coredump -> bt to view call stack
Abort	Assertion failure / abnormal termination	GDB analyze coredump -> check assertion condition
Stack overflow	Recursion too deep / large array	Check array size within Kernel
aic error	Memory access out-of-bounds / unaligned access	Check DataCopy length, 32B alignment
AIV hang	Missing cross-core synchronization	Check CrossCoreWaitFlag corresponding SetFlag
drain hang	AIV->AIC flag not sent	Check flag sending chain

Enter detailed process by scenario

Scenario	Description	Detailed Steps
Program crash	Segmentation Fault / Abort	Coredump debugging
Program hang/timeout	Kernel unresponsive, hang	Kernel hang debugging
Buffer conflict/deadlock	Alloc/Free mismatch, missing synchronization	Buffer issue troubleshooting

Detailed Resources

• crash_workflow.md: Complete debugging process + debugging tool quick reference
• parse_plog.py: plog log parsing script (automatically identifies hang/crash/hardware exception signals)
• ascendc-env-check skill: Environment variable configuration guide