[C++][CI] arrow-json-test segfaults occasionally on AMD64 Windows MinGW MINGW64 C++ job

[raulcd]

Describe the bug, including details regarding any error messages, version, and platform.

I've seen on several PRs the AMD64 Windows MinGW MINGW64 C++ segfault when running the following test:

The following tests FAILED:
	 84 - arrow-json-test (SEGFAULT) arrow-tests unittest

The error doesn't show much (obvious with a segfault):

 Start 84: arrow-json-test
81/98 Test #84: arrow-json-test ..............................***Exception: SegFault 0.03 sec
Running main() from gmock_main.cc
[==========] Running 235 tests from 60 test suites.
[----------] Global test environment set-up.
[----------] 6 tests from ReaderTest
[ RUN ] ReaderTest.MultipleChunksParallel

 Start 84: arrow-json-test
 Test #84: arrow-json-test ..............................***Exception: SegFault 0.03 sec
Running main() from gmock_main.cc
[==========] Running 235 tests from 60 test suites.
[----------] Global test environment set-up.
[----------] 6 tests from ReaderTest
[ RUN ] ReaderTest.MultipleChunksParallel

 Start 84: arrow-json-test
 Test #84: arrow-json-test ..............................***Exception: SegFault 0.03 sec
Running main() from gmock_main.cc
[==========] Running 235 tests from 60 test suites.
[----------] Global test environment set-up.
[----------] 6 tests from ReaderTest
[ RUN ] ReaderTest.MultipleChunksParallel

Example of runs that failed on different PRs:
https://github.com/apache/arrow/actions/runs/21956381694/job/63422678894?pr=49262
https://github.com/apache/arrow/actions/runs/21981450495/job/63504813441?pr=49259

Component(s)

C++, Continuous Integration

[vanshaj2023]

Hi @raulcd
I'd like to work on this issue.

After reviewing the repository, I found that the segfault in ReaderTest.MultipleChunksParallel is likely related to:

Implementation Approach:

1. The parallel chunk processing in cpp/src/arrow/json/reader.cc might have race conditions when multiple threads access shared resources
2. Check memory management in ThreadedTaskGroup usage within the JSON reader - potential use-after-free or improper synchronization
3. MinGW-specific thread handling differences compared to MSVC - review thread pool initialization and cleanup
4. Add mutex guards around shared state modifications in parallel parsing code
5. Verify proper lifetime management of chunked JSON buffers during concurrent reads

The intermittent nature and the fact it only fails on MinGW suggests platform-specific threading or memory alignment issues in the parallel reader implementation.

Could you please assign this issue to me? Does this approach sound reasonable, or am I missing something? Any specific areas in the JSON reader code I should prioritize?

Thanks!

[vanshaj2023]

I've submitted a PR to fix this: #49297

The root cause is MinGW's __emutls TLS emulation racing during thread creation in the thread pool. The fix replaces thread_local with native Win32 TLS API (TlsAlloc/TlsGetValue/TlsSetValue) on MinGW.

@raulcd since you reported this, could you take a look when you get a chance?

[vanshaj2023]

Hi, could someone review this when you get a chance?
This fixes an intermittent MinGW CI segfault (#49272) that causes false failures across unrelated PRs. The fix is minimal, just replaces thread_local with Win32 TLS API on MinGW in thread_pool.cc to bypass a known __emutls race condition.
Thanks!