[bug](be) Undefined _Unwind_* symbols when building BE UT in UBSAN mode

[heguanhui]

Search before asking

• I had searched in the issues and found no similar issues.

Version

master

What's Wrong?

When building BE unit tests in UBSAN mode, the linker reports undefined _Unwind_* symbols, causing the build to fail. The same issue also affects LSAN builds.

The root cause is that when using sanitizer link flags, the _Unwind_* family of symbols (from libgcc_eh / libgcc_s) are not being linked. These symbols are required by:

1. libclang_rt.ubsan_standalone (references _Unwind_Backtrace, _Unwind_GetIP)
2. libprofiler.a from gperftools (references _Unwind_Backtrace, _Unwind_GetIP)
3. libglog.a (references _Unwind_GetIP)
4. libstdc++.a (references _Unwind_Resume, _Unwind_RaiseException, _Unwind_Resume_or_Rethrow, _Unwind_GetTextRelBase, _Unwind_GetRegionStart, _Unwind_GetDataRelBase, _Unwind_SetGR, _Unwind_SetIP, _Unwind_GetLanguageSpecificData, _Unwind_GetIPInfo, _Unwind_DeleteException)

Full Error Log

ld.lld: error: undefined symbol: _Unwind_Backtrace
>>> referenced by sanitizer_unwind_linux_libcdep.cpp.o:(__sanitizer::BufferedStackTrace::UnwindSlow(unsigned long, unsigned int)) in archive /var/local/ldb-toolchain/tmp/gentoo/llvm/bin/../../../../lib/clang/20/lib/linux/libclang_rt.ubsan_standalone-x86_64.a
>>> referenced by sanitizer_unwind_linux_libcdep.cpp.o:(__sanitizer::BufferedStackTrace::UnwindSlow(unsigned long, void*, unsigned int)) in archive /var/local/ldb-toolchain/tmp/gentoo/llvm/bin/../../../../lib/clang/20/lib/linux/libclang_rt.ubsan_standalone-x86_64.a
>>> referenced by stacktrace_libgcc-inl.h:100 (src/stacktrace_libgcc-inl.h:100)
>>>               stacktrace.o:(GetStackTraceWithContext_libgcc(void**, int, int, void const*)) in archive /var/local/thirdparty/installed/gperftools/lib/libprofiler.a
>>> referenced 8 more times

ld.lld: error: undefined symbol: _Unwind_GetIP
>>> referenced by sanitizer_unwind_linux_libcdep.cpp.o:(__sanitizer::(anonymous namespace)::Unwind_Trace(_Unwind_Context*, void*)) in archive /var/local/ldb-toolchain/tmp/gentoo/llvm/bin/../../../../lib/clang/20/lib/linux/libclang_rt.ubsan_standalone-x86_64.a
>>> referenced by stacktrace_libgcc-inl.h:69 (src/stacktrace_libgcc-inl.h:69)
>>>               stacktrace.o:(libgcc_backtrace_helper(_Unwind_Context*, void*)) in archive /var/local/thirdparty/installed/gperftools/lib/libprofiler.a
>>> referenced by utilities.cc.o:(.text+0x5a) in archive /var/local/thirdparty/installed/lib/libglog.a
>>> referenced 2 more times

ld.lld: error: undefined symbol: _Unwind_Resume
>>> referenced by task_worker_pool_test.cpp:38 (./be/ut_build_UBSAN/./be/test/agent/task_worker_pool_test.cpp:38)
>>>               test/CMakeFiles/doris_be_test.dir/agent/task_worker_pool_test.cpp.o:(doris::TaskWorkerPoolTest_TaskWorkerPool_Test::TestBody())
>>> referenced by task_worker_pool_test.cpp:60 (./be/ut_build_UBSAN/./be/test/agent/task_worker_pool_test.cpp:60)
>>>               test/CMakeFiles/doris_be_test.dir/agent/task_worker_pool_test.cpp.o:(doris::TaskWorkerPoolTest_PreSubmitCallback_Test::TestBody())
>>> referenced by task_worker_pool_test.cpp:85 (./be/ut_build_UBSAN/./be/test/agent/task_worker_pool_test.cpp:85)
>>>               test/CMakeFiles/doris_be_test.dir/agent/task_worker_pool_test.cpp.o:(doris::TaskWorkerPoolTest_PreSubmitCallbackWithDedup_Test::TestBody())
>>> referenced 257716 more times

ld.lld: error: undefined symbol: _Unwind_RaiseException
>>> referenced by eh_throw.o:(__cxa_throw) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a
>>> referenced by eh_ptr.o:(std::rethrow_exception(std::__exception_ptr::exception_ptr)) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a

ld.lld: error: undefined symbol: _Unwind_Resume_or_Rethrow
>>> referenced by eh_throw.o:(__cxa_rethrow) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a

ld.lld: error: undefined symbol: _Unwind_GetTextRelBase
>>> referenced by eh_personality.o:(base_of_encoded_value(unsigned char, _Unwind_Context*)) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a

ld.lld: error: undefined symbol: _Unwind_GetRegionStart
>>> referenced by eh_personality.o:(base_of_encoded_value(unsigned char, _Unwind_Context*)) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a
>>> referenced by eh_personality.o:(parse_lsda_header(_Unwind_Context*, unsigned char const*, lsda_header_info*)) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a

ld.lld: error: undefined symbol: _Unwind_GetDataRelBase
>>> referenced by eh_personality.o:(base_of_encoded_value(unsigned char, _Unwind_Context*)) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a

ld.lld: error: undefined symbol: _Unwind_SetGR
>>> referenced by eh_personality.o:(__gxx_personality_v0) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a
>>> referenced by eh_personality.o:(__gxx_personality_v0) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a

ld.lld: error: undefined symbol: _Unwind_SetIP
>>> referenced by eh_personality.o:(__gxx_personality_v0) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a

ld.lld: error: undefined symbol: _Unwind_GetLanguageSpecificData
>>> referenced by eh_personality.o:(__gxx_personality_v0) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a

ld.lld: error: undefined symbol: _Unwind_GetIPInfo
>>> referenced by eh_personality.o:(__gxx_personality_v0) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a

ld.lld: error: undefined symbol: _Unwind_DeleteException
>>> referenced by eh_catch.o:(__cxa_end_catch) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a
>>> referenced by eh_catch.o:(__cxa_end_catch) in archive /var/local/ldb-toolchain/bin/../lib/gcc/x86_64-pc-linux-gnu/15/libstdc++.a
clang++: error: linker command failed with exit code 1 (use -v to see invocation)
ninja: build stopped: subcommand failed.

What You Expected to Work

BE UT should build successfully in UBSAN mode without linker errors.

How to Reproduce?

1. Build BE with UBSAN mode: BUILD_TYPE=UBSAN ./build.sh --be --ut
2. The linker will fail with undefined _Unwind_* symbol errors

Anything Else?

The fix is to add -lgcc_eh to the sanitizer link flags for both GCC and Clang compilers. For GCC, -lgcc_eh provides the _Unwind_* symbols. For Clang with -rtlib=compiler-rt, the compiler-rt runtime does not include unwind support by default, so -lgcc_eh must also be explicitly linked.

This fix applies to both be/CMakeLists.txt and cloud/CMakeLists.txt.

Are you willing to submit PR?

• Yes I am willing to submit a PR!

[zclllyybb]

Breakwater-GitHub-Analysis-Slot: slot_fbda170906e0

Initial analysis based on the issue log and current upstream/master code:

This looks like a real sanitizer link dependency issue, not a failure in task_worker_pool_test itself. In current master, be/CMakeLists.txt enables LSAN/UBSAN compile flags and then appends sanitizer-specific link flags through LSAN_LIBS / UBSAN_LIBS; cloud/CMakeLists.txt has the same pattern. Those link paths do not add an explicit EH unwind provider. For Clang UBSAN, the link flags currently keep -rtlib=compiler-rt plus the allocator, and for LSAN the Clang-side extra library variable is effectively not adding an unwind provider. The reported undefined symbols are exactly the Itanium unwind APIs consumed by the sanitizer runtime, gperftools stacktrace fallback, glog, and static libstdc++ exception handling.

The proposed fix direction is reasonable: add the EH unwind provider to the sanitizer link path in both be/CMakeLists.txt and cloud/CMakeLists.txt, at least for LSAN and UBSAN. Prefer adding it through the sanitizer library variables so it is appended after the static libraries that reference _Unwind_*. Please keep both GNU and Clang paths covered; for Clang/compiler-rt this should add the missing unwind provider, not replace -rtlib=compiler-rt.

One reproduction detail should be clarified in the issue or PR: current build.sh does not parse a --ut option. The failing path be/ut_build_UBSAN matches run-be-ut.sh, which uses BUILD_TYPE_UT and passes -DMAKE_TEST=ON. Please include the exact command used, for example the BUILD_TYPE_UT=UBSAN ./run-be-ut.sh ... invocation, plus the compiler/toolchain values.

Suggested PR validation:

• BE UT UBSAN link step succeeds, for example with the exact BUILD_TYPE_UT=UBSAN ./run-be-ut.sh ... command used by the reporter.
• BE UT LSAN link step also succeeds.
• If cloud/CMakeLists.txt is changed, the matching cloud UT CMake/link step for UBSAN and LSAN is also checked.
• Attach or summarize the generated link command / CMake "Link Flags" snippet showing the unwind provider after the sanitizer/static-library entries.

I did not find an existing public Doris issue or PR for the same _Unwind_* / libgcc_eh symptom during a narrow search.