FIX a number of race conditions in threadpoolctl's OpenBLAS-pthreads ownership

.github/workflows/test.yml

@@ -125,7 +125,7 @@ jobs:
 
           # Linux environment with development versions of numpy and scipy
           - name: pylatest_pip_dev
-            os : ubuntu-latest
+            os : ubuntu-24.04
             PACKAGER: "pip-dev"
             PYTHON_VERSION: "*"
             CC_OUTER_LOOP: "gcc"
@@ -134,7 +134,7 @@ jobs:
 
           # Linux + Python 3.10 and homogeneous runtime nesting.
           - name: py310_conda_openblas_clang_clang
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PACKAGER: "conda"
             PYTHON_VERSION: "3.10"
             BLAS: "openblas"
@@ -145,7 +145,7 @@ jobs:
           # Linux environment with MKL and Clang (known to be unsafe for
           # threadpoolctl) to only test the warning from multiple OpenMP.
           - name: pylatest_conda_mkl_clang_gcc
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PYTHON_VERSION: "*"
             PACKAGER: "conda"
             BLAS: "mkl"
@@ -154,7 +154,7 @@ jobs:
             TESTS: "libomp_libiomp_warning"
           # Linux environment with MKL, safe for threadpoolctl.
           - name: pylatest_conda_mkl_gcc_gcc
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PYTHON_VERSION: "*"
             PACKAGER: "conda"
             BLAS: "mkl"
@@ -166,15 +166,15 @@ jobs:
           # Linux + Python 3.11 with numpy / scipy installed with pip from PyPI
           # and heterogeneous OpenMP runtimes.
           - name: py311_pip_openblas_gcc_clang
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PACKAGER: "pip"
             PYTHON_VERSION: "3.11"
             CC_OUTER_LOOP: "gcc"
             CC_INNER_LOOP: "clang-18"
           # Linux + conda-forge Python RC, extra deps from PyPI.
           # OpenCV is omitted: no wheels for the current RC ABI yet.
           - name: pyrc_pip
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PACKAGER: "pip"
             # Using the "*" marker does not select the latest RC release
             # from the conda-forge `python_rc` channel hence the need to pin
@@ -185,7 +185,7 @@ jobs:
             CC_INNER_LOOP: "gcc"
           # Linux + conda-forge free-threaded Python RC, extra deps from PyPI.
           - name: pyrc_ft_pip
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PACKAGER: "pip"
             # Using the "*" marker does not select the latest RC release
             # from the conda-forge `python_rc` channel hence the need to pin
@@ -197,7 +197,7 @@ jobs:
             CC_INNER_LOOP: "gcc"
           # Linux environment with numpy from conda-forge channel and openblas-openmp
           - name: pylatest_conda_forge
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PACKAGER: "conda-forge"
             PYTHON_VERSION: "*"
             BLAS: "openblas"
@@ -207,7 +207,7 @@ jobs:
             CC_INNER_LOOP: "gcc"
           # Linux environment with no numpy and heterogeneous OpenMP runtimes.
           - name: pylatest_conda_nonumpy_gcc_clang
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PACKAGER: "conda"
             PYTHON_VERSION: "*"
             NO_NUMPY: "true"
@@ -216,7 +216,7 @@ jobs:
 
           # Linux environments with numpy linked to BLIS
           - name: pylatest_blis_gcc_clang_openmp
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PYTHON_VERSION: "*"
             INSTALL_BLAS: "blis"
             BLIS_NUM_THREAEDS: "4"
@@ -225,7 +225,7 @@ jobs:
             BLIS_CC: "clang-18"
             BLIS_ENABLE_THREADING: "openmp"
           - name: pylatest_blis_clang_gcc_pthreads
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PYTHON_VERSION: "*"
             INSTALL_BLAS: "blis"
             BLIS_NUM_THREADS: "4"
@@ -234,7 +234,7 @@ jobs:
             BLIS_CC: "gcc-12"
             BLIS_ENABLE_THREADING: "pthreads"
           - name: pylatest_blis_no_threading
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PYTHON_VERSION: "*"
             INSTALL_BLAS: "blis"
             BLIS_NUM_THREADS: "1"
@@ -245,7 +245,7 @@ jobs:
 
           # Linux env with FlexiBLAS
           - name: pylatest_flexiblas
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PYTHON_VERSION: "*"
             INSTALL_BLAS: "flexiblas"
             PLATFORM_SPECIFIC_PACKAGES: "mkl"
@@ -254,7 +254,7 @@ jobs:
 
           # Python 3.14, free-threaded
           - name: py314_freethreaded_pip_openblas
-            os: ubuntu-latest
+            os: ubuntu-24.04
             PACKAGER: "conda-forge"
             BLAS: "openblas"
             PYTHON_VERSION: "3.14"

CHANGES.md

@@ -1,6 +1,11 @@
-3.9.0 (under development)
+3.8.1 (under development)
 =========================
 
+- Fix a number of race conditions in OpenBLAS-pthreads/win32 controlling thread
+  logic, that could result in unexpected number of threads being set.
+  https://github.com/joblib/threadpoolctl/pull/300
+
+
 3.8.0 (2026-11-05)
 ==================
 

tests/test_threadpoolctl.py

@@ -8,7 +8,7 @@
 import subprocess
 import sys
 import time
-from threading import Lock, Thread, current_thread
+from threading import Event, Lock, Thread, current_thread
 
 import pytest
 
@@ -1063,7 +1063,7 @@ def test_restore_original_limits_keeps_openblas_controlling_thread(request):
     libs, lib, orig = _require_resizable_non_openmp_openblas()
     request.addfinalizer(_restore_openblas_thread_counts(libs))
 
-    limiter = threadpool_limits(limits=1)
+    limiter = ThreadpoolController().select(internal_api="openblas").limit(limits=1)
     assert lib.get_num_threads() == 1
     assert lib._SHARED_STATE.controlling_thread is current_thread()
 
@@ -1075,6 +1075,144 @@ def test_restore_original_limits_keeps_openblas_controlling_thread(request):
     assert lib._SHARED_STATE.controlling_thread is None
 
 
+def _race_condition_skip_problem_platforms():
+    flexiblas_controller = ThreadpoolController().select(internal_api="flexiblas")
+
+    if sys.platform == "win32":
+        pytest.xfail("OpenMP on Windows has its own issues with overlapping setting")
+    if flexiblas_controller:
+        pytest.skip(
+            "FlexiBLAS means there are two controllers for same OpenBLAS "
+            "(direct + FlexiBLAS) which causes test failures."
+        )
+
+
+def test_overlapping_context_manager_restores_correctly(request):
+    """A thread that failed to set should also not restore."""
+    _race_condition_skip_problem_platforms()
+
+    libs, lib, orig_limit = _require_resizable_non_openmp_openblas()
+    request.addfinalizer(_restore_openblas_thread_counts(libs))
+    openmp = ThreadpoolController().select(user_api="openmp")
+    original_openmp_info = openmp.info()
+
+    a_in, b_in, a_out = (Event() for _ in range(3))
+
+    a_limits_correct = []
+    b_limits_correct = []
+
+    def A():
+        with threadpool_limits(2):
+            a_in.set()
+            b_in.wait()
+        a_limits_correct.append(openmp.info() == original_openmp_info)
+        a_out.set()
+
+    def B():
+        a_in.wait()
+        with threadpool_limits(1):  # snapshot taken while A's limit is active
+            b_in.set()
+            a_out.wait()
+        b_limits_correct.append(openmp.info() == original_openmp_info)
+
+    ts = [Thread(target=A), Thread(target=B)]
+    for t in ts:
+        t.start()
+    for t in ts:
+        t.join()
+
+    assert a_limits_correct == [True]
+    assert b_limits_correct == [True]
+    assert lib.get_num_threads() == orig_limit
+    assert openmp.info() == original_openmp_info
+
+
+def test_overlapping_set_and_restore(request):
+    """A thread that failed to set should also not restore."""
+    _race_condition_skip_problem_platforms()
+
+    libs, lib, orig_limit = _require_resizable_non_openmp_openblas()
+    request.addfinalizer(_restore_openblas_thread_counts(libs))
+    openmp = ThreadpoolController().select(user_api="openmp")
+    original_openmp_info = openmp.info()
+
+    a_in, b_in, a_out = (Event() for _ in range(3))
+
+    a_limits_correct = []
+    b_limits_correct = []
+
+    def A():
+        limiter = threadpool_limits(2)
+        a_in.set()
+        b_in.wait()
+        limiter.restore_original_limits()
+        a_limits_correct.append(openmp.info() == original_openmp_info)
+        a_out.set()
+
+    def B():
+        a_in.wait()
+        limiter = threadpool_limits(1)  # snapshot taken while A's limit is active
+        b_in.set()
+        a_out.wait()
+        limiter.restore_original_limits()
+        b_limits_correct.append(openmp.info() == original_openmp_info)
+
+    ts = [Thread(target=A), Thread(target=B)]
+    for t in ts:
+        t.start()
+    for t in ts:
+        t.join()
+
+    assert a_limits_correct == [True]
+    assert b_limits_correct == [True]
+    assert lib.get_num_threads() == orig_limit
+    assert openmp.info() == original_openmp_info
+
+
+def test_get_current_info_race_condition(request):
+    """Test for race condition in multi-threaded OpenBLAS-pthreads usage.
+
+    1. Thread A sets limit L.
+    2. Thread B gets limit L in `_ThreadpoolLimiter.__init__`.
+    3. Thread A restores to original limit O and releases control.
+    4. Thread B continues `__init__`, setting a limit and taking control.
+    5. Thread B restores to limit L.
+    """
+    _race_condition_skip_problem_platforms()
+
+    libs, lib, orig_limit = _require_resizable_non_openmp_openblas()
+    request.addfinalizer(_restore_openblas_thread_counts(libs))
+
+    a_in, b_retrieved, a_out = (Event() for _ in range(3))
+
+    def A():
+        with threadpool_limits(2):
+            a_in.set()
+            b_retrieved.wait()
+        a_out.set()
+
+    class racy_threadpool_limits(threadpool_limits):
+        def _set_threadpool_limits(self):
+            # Tell A it can finish up:
+            b_retrieved.set()
+            # Wait until thread A is finished, so we can take control:
+            a_out.wait()
+            super()._set_threadpool_limits()
+
+    def B():
+        a_in.wait()
+        with racy_threadpool_limits(1):
+            pass
+
+    ts = [Thread(target=A), Thread(target=B)]
+    for t in ts:
+        t.start()
+    for t in ts:
+        t.join()
+
+    assert lib.get_num_threads() == orig_limit
+
+
 # skip test if not run in a azure pipelines job since it relies on a specific flexiblas
 # installation.
 @pytest.mark.skipif(