diff --git a/src/conversions/chrono.rs b/src/conversions/chrono.rs
index 5ea8904fca2..bd97d3409fc 100644
--- a/src/conversions/chrono.rs
+++ b/src/conversions/chrono.rs
@@ -593,7 +593,7 @@ fn py_time_to_naive_time(py_time: &Bound<'_, PyAny>) -> PyResult<NaiveTime> {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::{types::PyTuple, BoundObject};
+    use crate::{test_utils::assert_warnings, types::PyTuple, BoundObject};
     use std::{cmp::Ordering, panic};
 
     #[test]
@@ -875,7 +875,6 @@ mod tests {
 
             check_utc("regular", 2014, 5, 6, 7, 8, 9, 999_999, 999_999);
 
-            #[cfg(not(Py_GIL_DISABLED))]
             assert_warnings!(
                 py,
                 check_utc("leap second", 2014, 5, 6, 7, 8, 59, 1_999_999, 999_999),
@@ -915,7 +914,6 @@ mod tests {
 
             check_fixed_offset("regular", 2014, 5, 6, 7, 8, 9, 999_999, 999_999);
 
-            #[cfg(not(Py_GIL_DISABLED))]
             assert_warnings!(
                 py,
                 check_fixed_offset("leap second", 2014, 5, 6, 7, 8, 59, 1_999_999, 999_999),
@@ -1101,7 +1099,6 @@ mod tests {
 
             check_time("regular", 3, 5, 7, 999_999, 999_999);
 
-            #[cfg(not(Py_GIL_DISABLED))]
             assert_warnings!(
                 py,
                 check_time("leap second", 3, 5, 59, 1_999_999, 999_999),
@@ -1163,10 +1160,10 @@ mod tests {
             .unwrap()
     }
 
-    #[cfg(not(any(target_arch = "wasm32", Py_GIL_DISABLED)))]
+    #[cfg(not(any(target_arch = "wasm32")))]
     mod proptests {
         use super::*;
-        use crate::tests::common::CatchWarnings;
+        use crate::test_utils::CatchWarnings;
         use crate::types::IntoPyDict;
         use proptest::prelude::*;
         use std::ffi::CString;
diff --git a/src/conversions/num_bigint.rs b/src/conversions/num_bigint.rs
index f9b4f11fa69..0f8924bbf12 100644
--- a/src/conversions/num_bigint.rs
+++ b/src/conversions/num_bigint.rs
@@ -317,7 +317,7 @@ fn int_n_bits(long: &Bound<'_, PyInt>) -> PyResult<usize> {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::tests::common::generate_unique_module_name;
+    use crate::test_utils::generate_unique_module_name;
     use crate::types::{PyAnyMethods as _, PyDict, PyModule};
     use indoc::indoc;
     use pyo3_ffi::c_str;
diff --git a/src/conversions/num_complex.rs b/src/conversions/num_complex.rs
index edcc0ef2d54..cdb7ad3fc42 100644
--- a/src/conversions/num_complex.rs
+++ b/src/conversions/num_complex.rs
@@ -196,7 +196,7 @@ complex_conversion!(f64);
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::tests::common::generate_unique_module_name;
+    use crate::test_utils::generate_unique_module_name;
     use crate::types::PyAnyMethods as _;
     use crate::types::{complex::PyComplexMethods, PyModule};
     use crate::IntoPyObject;
diff --git a/src/err/mod.rs b/src/err/mod.rs
index 16265865283..fe8d3755d4a 100644
--- a/src/err/mod.rs
+++ b/src/err/mod.rs
@@ -853,6 +853,7 @@ mod tests {
     use super::PyErrState;
     use crate::exceptions::{self, PyTypeError, PyValueError};
     use crate::impl_::pyclass::{value_of, IsSend, IsSync};
+    use crate::test_utils::assert_warnings;
     use crate::{ffi, PyErr, PyTypeInfo, Python};
 
     #[test]
@@ -1052,7 +1053,6 @@ mod tests {
             warnings.call_method0("resetwarnings").unwrap();
 
             // First, test the warning is emitted
-            #[cfg(not(Py_GIL_DISABLED))]
             assert_warnings!(
                 py,
                 { PyErr::warn(py, &cls, ffi::c_str!("I am warning you"), 0).unwrap() },
@@ -1072,7 +1072,6 @@ mod tests {
                 .unwrap();
 
             // This has the wrong module and will not raise, just be emitted
-            #[cfg(not(Py_GIL_DISABLED))]
             assert_warnings!(
                 py,
                 { PyErr::warn(py, &cls, ffi::c_str!("I am warning you"), 0).unwrap() },
diff --git a/src/instance.rs b/src/instance.rs
index 3bbbb4729b5..f9347f2b2ca 100644
--- a/src/instance.rs
+++ b/src/instance.rs
@@ -2171,7 +2171,7 @@ impl Py<PyAny> {
 #[cfg(test)]
 mod tests {
     use super::{Bound, IntoPyObject, Py};
-    use crate::tests::common::generate_unique_module_name;
+    use crate::test_utils::generate_unique_module_name;
     use crate::types::{dict::IntoPyDict, PyAnyMethods, PyCapsule, PyDict, PyString};
     use crate::{ffi, Borrowed, PyAny, PyResult, Python};
     use pyo3_ffi::c_str;
diff --git a/src/lib.rs b/src/lib.rs
index 34b4a6ae156..bdf0dd7cd5c 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -410,7 +410,8 @@ pub use inventory; // Re-exported for `#[pyclass]` and `#[pymethods]` with `mult
 /// Tests and helpers which reside inside PyO3's main library. Declared first so that macros
 /// are available in unit tests.
 #[cfg(test)]
-#[macro_use]
+mod test_utils;
+#[cfg(test)]
 mod tests;
 
 // Macro dependencies, also contains macros exported for use across the codebase and
diff --git a/src/test_utils.rs b/src/test_utils.rs
new file mode 100644
index 00000000000..32b74997bcb
--- /dev/null
+++ b/src/test_utils.rs
@@ -0,0 +1,6 @@
+// Brings in `test_utils` from the `tests` directory
+//
+// to make that file function (lots of references to `pyo3` within it) need
+// re-bind `crate` as pyo3
+use crate as pyo3;
+include!("../tests/test_utils/mod.rs");
diff --git a/src/tests/mod.rs b/src/tests/mod.rs
index 136236d7db0..3b72764301c 100644
--- a/src/tests/mod.rs
+++ b/src/tests/mod.rs
@@ -1,10 +1,3 @@
-#[macro_use]
-pub(crate) mod common {
-    #[cfg(not(Py_GIL_DISABLED))]
-    use crate as pyo3;
-    include!("./common.rs");
-}
-
 /// Test macro hygiene - this is in the crate since we won't have
 /// `pyo3` available in the crate root.
 #[cfg(all(test, feature = "macros"))]
diff --git a/src/types/any.rs b/src/types/any.rs
index fb3d97b55ce..665fb607480 100644
--- a/src/types/any.rs
+++ b/src/types/any.rs
@@ -1624,7 +1624,7 @@ mod tests {
     use crate::{
         basic::CompareOp,
         ffi,
-        tests::common::generate_unique_module_name,
+        test_utils::generate_unique_module_name,
         types::{IntoPyDict, PyAny, PyAnyMethods, PyBool, PyInt, PyList, PyModule, PyTypeMethods},
         Bound, BoundObject, IntoPyObject, PyTypeInfo, Python,
     };
diff --git a/src/types/typeobject.rs b/src/types/typeobject.rs
index 58c94d46cf4..953a78bc6aa 100644
--- a/src/types/typeobject.rs
+++ b/src/types/typeobject.rs
@@ -244,7 +244,7 @@ impl<'py> PyTypeMethods<'py> for Bound<'py, PyType> {
 
 #[cfg(test)]
 mod tests {
-    use crate::tests::common::generate_unique_module_name;
+    use crate::test_utils::generate_unique_module_name;
     use crate::types::{PyAnyMethods, PyBool, PyInt, PyModule, PyTuple, PyType, PyTypeMethods};
     use crate::PyAny;
     use crate::Python;
diff --git a/tests/test_arithmetics.rs b/tests/test_arithmetics.rs
index 3e0e7712ba4..bf3f99c3462 100644
--- a/tests/test_arithmetics.rs
+++ b/tests/test_arithmetics.rs
@@ -4,8 +4,7 @@ use pyo3::class::basic::CompareOp;
 use pyo3::py_run;
 use pyo3::{prelude::*, BoundObject};
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct UnaryArithmetic {
diff --git a/tests/test_buffer.rs b/tests/test_buffer.rs
index 4d975f36479..7ffa3e58864 100644
--- a/tests/test_buffer.rs
+++ b/tests/test_buffer.rs
@@ -9,8 +9,7 @@ use std::{
 };
 
 #[macro_use]
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 enum TestGetBufferError {
     NullShape,
diff --git a/tests/test_buffer_protocol.rs b/tests/test_buffer_protocol.rs
index d5f50d02e5d..5a2baa33d05 100644
--- a/tests/test_buffer_protocol.rs
+++ b/tests/test_buffer_protocol.rs
@@ -13,8 +13,7 @@ use std::ptr;
 use std::sync::atomic::{AtomicBool, Ordering};
 use std::sync::Arc;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct TestBufferClass {
@@ -98,8 +97,8 @@ fn test_buffer_referenced() {
 #[test]
 #[cfg(all(Py_3_8, not(Py_GIL_DISABLED)))] // sys.unraisablehook not available until Python 3.8
 fn test_releasebuffer_unraisable_error() {
-    use common::UnraisableCapture;
     use pyo3::exceptions::PyValueError;
+    use test_utils::UnraisableCapture;
 
     #[pyclass]
     struct ReleaseBufferError {}
diff --git a/tests/test_bytes.rs b/tests/test_bytes.rs
index 7df25c2fe58..0caaf2a37ff 100644
--- a/tests/test_bytes.rs
+++ b/tests/test_bytes.rs
@@ -3,8 +3,7 @@
 use pyo3::prelude::*;
 use pyo3::types::PyBytes;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyfunction]
 fn bytes_pybytes_conversion(bytes: &[u8]) -> &[u8] {
diff --git a/tests/test_class_attributes.rs b/tests/test_class_attributes.rs
index b1a019c0457..e8d104d9d10 100644
--- a/tests/test_class_attributes.rs
+++ b/tests/test_class_attributes.rs
@@ -3,8 +3,7 @@
 use pyo3::prelude::*;
 use pyo3::py_run;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct Foo {
@@ -154,8 +153,8 @@ fn recursive_class_attributes() {
 #[test]
 #[cfg(all(Py_3_8, not(Py_GIL_DISABLED)))] // sys.unraisablehook not available until Python 3.8
 fn test_fallible_class_attribute() {
-    use common::UnraisableCapture;
     use pyo3::exceptions::PyValueError;
+    use test_utils::UnraisableCapture;
 
     #[pyclass]
     struct BrokenClass;
diff --git a/tests/test_class_basics.rs b/tests/test_class_basics.rs
index 027f050320e..1dda580c2a4 100644
--- a/tests/test_class_basics.rs
+++ b/tests/test_class_basics.rs
@@ -4,8 +4,7 @@ use pyo3::prelude::*;
 use pyo3::types::PyType;
 use pyo3::{py_run, PyClass};
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct EmptyClass {}
@@ -619,12 +618,12 @@ fn access_frozen_class_without_gil() {
 #[cfg(all(Py_3_8, not(Py_GIL_DISABLED)))] // sys.unraisablehook not available until Python 3.8
 #[cfg_attr(target_arch = "wasm32", ignore)]
 fn drop_unsendable_elsewhere() {
-    use common::UnraisableCapture;
     use std::sync::{
         atomic::{AtomicBool, Ordering},
         Arc,
     };
     use std::thread::spawn;
+    use test_utils::UnraisableCapture;
 
     #[pyclass(unsendable)]
     struct Unsendable {
diff --git a/tests/test_class_comparisons.rs b/tests/test_class_comparisons.rs
index 2f716ca431a..e8509d1e266 100644
--- a/tests/test_class_comparisons.rs
+++ b/tests/test_class_comparisons.rs
@@ -2,8 +2,7 @@
 
 use pyo3::prelude::*;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass(eq)]
 #[derive(Debug, Clone, PartialEq)]
diff --git a/tests/test_class_conversion.rs b/tests/test_class_conversion.rs
index c451a37ac35..2ff0f56a9b3 100644
--- a/tests/test_class_conversion.rs
+++ b/tests/test_class_conversion.rs
@@ -3,8 +3,7 @@
 use pyo3::prelude::*;
 
 #[macro_use]
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 #[derive(Clone, Debug, PartialEq)]
diff --git a/tests/test_class_formatting.rs b/tests/test_class_formatting.rs
index 976e6da4e5e..1b4add063d4 100644
--- a/tests/test_class_formatting.rs
+++ b/tests/test_class_formatting.rs
@@ -4,8 +4,7 @@ use pyo3::prelude::*;
 use pyo3::py_run;
 use std::fmt::{Display, Formatter};
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass(eq, str)]
 #[derive(Debug, PartialEq)]
diff --git a/tests/test_coroutine.rs b/tests/test_coroutine.rs
index 868caf7e69b..24e007ab5a3 100644
--- a/tests/test_coroutine.rs
+++ b/tests/test_coroutine.rs
@@ -14,8 +14,7 @@ use pyo3::{
 #[cfg(target_has_atomic = "64")]
 use std::sync::atomic::{AtomicBool, Ordering};
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 fn handle_windows(test: &str) -> String {
     let set_event_loop_policy = r#"
diff --git a/tests/test_declarative_module.rs b/tests/test_declarative_module.rs
index f3a01001952..57800cdce2e 100644
--- a/tests/test_declarative_module.rs
+++ b/tests/test_declarative_module.rs
@@ -7,8 +7,7 @@ use pyo3::exceptions::PyException;
 use pyo3::prelude::*;
 use pyo3::sync::OnceLockExt;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 mod some_module {
     use pyo3::create_exception;
diff --git a/tests/test_default_impls.rs b/tests/test_default_impls.rs
index e5735f5dd2a..19dc0773a80 100644
--- a/tests/test_default_impls.rs
+++ b/tests/test_default_impls.rs
@@ -2,8 +2,7 @@
 
 use pyo3::prelude::*;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 // Test default generated __repr__.
 #[pyclass(eq, eq_int)]
diff --git a/tests/test_enum.rs b/tests/test_enum.rs
index f1a2ab3eff6..cdc9935a6df 100644
--- a/tests/test_enum.rs
+++ b/tests/test_enum.rs
@@ -4,8 +4,7 @@ use pyo3::prelude::*;
 use pyo3::py_run;
 use pyo3::types::PyString;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass(eq, eq_int)]
 #[derive(Debug, PartialEq, Eq, Clone)]
diff --git a/tests/test_exceptions.rs b/tests/test_exceptions.rs
index 081041ed12d..97b5466d205 100644
--- a/tests/test_exceptions.rs
+++ b/tests/test_exceptions.rs
@@ -7,8 +7,7 @@ use std::fmt;
 #[cfg(not(target_os = "windows"))]
 use std::fs::File;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyfunction]
 #[cfg(not(target_os = "windows"))]
@@ -101,9 +100,9 @@ fn test_exception_nosegfault() {
 #[test]
 #[cfg(all(Py_3_8, not(Py_GIL_DISABLED)))]
 fn test_write_unraisable() {
-    use common::UnraisableCapture;
     use pyo3::{exceptions::PyRuntimeError, ffi, types::PyNotImplemented};
     use std::ptr;
+    use test_utils::UnraisableCapture;
 
     Python::attach(|py| {
         let capture = UnraisableCapture::install(py);
diff --git a/tests/test_frompy_intopy_roundtrip.rs b/tests/test_frompy_intopy_roundtrip.rs
index e8e41c14ef2..c4b120c5292 100644
--- a/tests/test_frompy_intopy_roundtrip.rs
+++ b/tests/test_frompy_intopy_roundtrip.rs
@@ -6,8 +6,7 @@ use std::collections::HashMap;
 use std::hash::Hash;
 
 #[macro_use]
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[derive(Debug, Clone, IntoPyObject, IntoPyObjectRef, FromPyObject)]
 pub struct A<'py> {
diff --git a/tests/test_frompyobject.rs b/tests/test_frompyobject.rs
index 35cc813a603..9a72e57284c 100644
--- a/tests/test_frompyobject.rs
+++ b/tests/test_frompyobject.rs
@@ -5,8 +5,7 @@ use pyo3::prelude::*;
 use pyo3::types::{IntoPyDict, PyDict, PyList, PyString, PyTuple};
 
 #[macro_use]
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 /// Helper function that concatenates the error message from
 /// each error in the traceback into a single string that can
diff --git a/tests/test_gc.rs b/tests/test_gc.rs
index 8f39c1baca1..f97868fec9f 100644
--- a/tests/test_gc.rs
+++ b/tests/test_gc.rs
@@ -13,8 +13,7 @@ use std::sync::atomic::{AtomicBool, Ordering};
 use std::sync::Once;
 use std::sync::{Arc, Mutex};
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass(freelist = 2)]
 struct ClassWithFreelist {}
diff --git a/tests/test_getter_setter.rs b/tests/test_getter_setter.rs
index 4d15627ea51..582cfd89c5f 100644
--- a/tests/test_getter_setter.rs
+++ b/tests/test_getter_setter.rs
@@ -7,8 +7,7 @@ use pyo3::py_run;
 use pyo3::types::PyString;
 use pyo3::types::{IntoPyDict, PyList};
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct ClassWithProperties {
diff --git a/tests/test_inheritance.rs b/tests/test_inheritance.rs
index 0500c0c2f83..99c26ae1041 100644
--- a/tests/test_inheritance.rs
+++ b/tests/test_inheritance.rs
@@ -6,8 +6,7 @@ use pyo3::py_run;
 use pyo3::ffi;
 use pyo3::types::IntoPyDict;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass(subclass)]
 struct BaseClass {
diff --git a/tests/test_intopyobject.rs b/tests/test_intopyobject.rs
index 58795d7c7f6..72ecc0f234d 100644
--- a/tests/test_intopyobject.rs
+++ b/tests/test_intopyobject.rs
@@ -6,8 +6,7 @@ use std::collections::HashMap;
 use std::hash::Hash;
 
 #[macro_use]
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[derive(Debug, IntoPyObject)]
 pub struct A<'py> {
diff --git a/tests/test_macro_docs.rs b/tests/test_macro_docs.rs
index fec11e51904..e27277a9326 100644
--- a/tests/test_macro_docs.rs
+++ b/tests/test_macro_docs.rs
@@ -4,8 +4,7 @@ use pyo3::prelude::*;
 use pyo3::types::IntoPyDict;
 
 #[macro_use]
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 /// The MacroDocs class.
diff --git a/tests/test_macros.rs b/tests/test_macros.rs
index 72b5b45ef7f..e7b1997145a 100644
--- a/tests/test_macros.rs
+++ b/tests/test_macros.rs
@@ -5,8 +5,7 @@
 use pyo3::prelude::*;
 
 #[macro_use]
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 macro_rules! make_struct_using_macro {
     // Ensure that one doesn't need to fall back on the escape type: tt
diff --git a/tests/test_mapping.rs b/tests/test_mapping.rs
index bab7c0a418a..df52372d143 100644
--- a/tests/test_mapping.rs
+++ b/tests/test_mapping.rs
@@ -10,8 +10,7 @@ use pyo3::types::PyList;
 use pyo3::types::PyMapping;
 use pyo3::types::PySequence;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass(mapping)]
 struct Mapping {
diff --git a/tests/test_methods.rs b/tests/test_methods.rs
index 06544f221e9..7fbc1a138b6 100644
--- a/tests/test_methods.rs
+++ b/tests/test_methods.rs
@@ -2,7 +2,6 @@
 
 #[cfg(not(Py_LIMITED_API))]
 use pyo3::exceptions::PyWarning;
-#[cfg(not(Py_GIL_DISABLED))]
 use pyo3::exceptions::{PyFutureWarning, PyUserWarning};
 use pyo3::prelude::*;
 use pyo3::py_run;
@@ -11,8 +10,7 @@ use pyo3::types::{IntoPyDict, PyDict, PyList, PySet, PyString, PyTuple, PyType};
 use pyo3::BoundObject;
 use pyo3_macros::pyclass;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct InstanceMethod {
@@ -1223,7 +1221,6 @@ impl UserDefinedWarning {
 }
 
 #[test]
-#[cfg(not(Py_GIL_DISABLED))] // FIXME: enable once `warnings` is thread-safe
 fn test_pymethods_warn() {
     // We do not test #[classattr] nor __traverse__
     // because it doesn't make sense to implement deprecated methods for them.
@@ -1424,7 +1421,6 @@ fn test_pymethods_warn() {
 }
 
 #[test]
-#[cfg(not(Py_GIL_DISABLED))] // FIXME: enable once `warnings` is thread-safe
 fn test_py_methods_multiple_warn() {
     #[pyclass]
     struct MultipleWarnContainer {}
diff --git a/tests/test_module.rs b/tests/test_module.rs
index 2f5766aadaf..5e6244420e0 100644
--- a/tests/test_module.rs
+++ b/tests/test_module.rs
@@ -8,8 +8,7 @@ use pyo3::types::{IntoPyDict, PyDict, PyTuple};
 use pyo3::BoundObject;
 use pyo3_ffi::c_str;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct AnonClass {}
@@ -168,7 +167,7 @@ fn test_module_from_code_bound() {
             py,
             c_str!("def add(a,b):\n\treturn a+b"),
             c_str!("adder_mod.py"),
-            &common::generate_unique_module_name("adder_mod"),
+            &test_utils::generate_unique_module_name("adder_mod"),
         )
         .expect("Module code should be loaded");
 
diff --git a/tests/test_multiple_pymethods.rs b/tests/test_multiple_pymethods.rs
index ccbda42b865..1a788425113 100644
--- a/tests/test_multiple_pymethods.rs
+++ b/tests/test_multiple_pymethods.rs
@@ -4,8 +4,7 @@ use pyo3::prelude::*;
 use pyo3::types::PyType;
 
 #[macro_use]
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct PyClassWithMultiplePyMethods {}
diff --git a/tests/test_proto_methods.rs b/tests/test_proto_methods.rs
index e8df772faf1..94a8ddd368a 100644
--- a/tests/test_proto_methods.rs
+++ b/tests/test_proto_methods.rs
@@ -6,8 +6,7 @@ use pyo3::{prelude::*, py_run};
 use std::iter;
 use std::sync::Mutex;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct EmptyClass;
diff --git a/tests/test_pyfunction.rs b/tests/test_pyfunction.rs
index 54447b92940..07bd9f5482c 100644
--- a/tests/test_pyfunction.rs
+++ b/tests/test_pyfunction.rs
@@ -7,7 +7,6 @@ use std::collections::HashMap;
 use pyo3::buffer::PyBuffer;
 #[cfg(not(Py_LIMITED_API))]
 use pyo3::exceptions::PyWarning;
-#[cfg(not(Py_GIL_DISABLED))]
 use pyo3::exceptions::{PyFutureWarning, PyUserWarning};
 use pyo3::ffi::c_str;
 use pyo3::prelude::*;
@@ -18,8 +17,7 @@ use pyo3::types::PyFunction;
 use pyo3::types::{self, PyCFunction};
 use pyo3_macros::pyclass;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyfunction(name = "struct")]
 fn struct_function() {}
@@ -669,7 +667,6 @@ impl UserDefinedWarning {
 }
 
 #[test]
-#[cfg(not(Py_GIL_DISABLED))] // FIXME: enable once `warnings` is thread-safe
 fn test_pyfunction_warn() {
     #[pyfunction]
     #[pyo3(warn(message = "this function raises warning"))]
@@ -724,7 +721,6 @@ fn test_pyfunction_warn() {
 }
 
 #[test]
-#[cfg(not(Py_GIL_DISABLED))] // FIXME: enable once `warnings` is thread-safe
 fn test_pyfunction_multiple_warnings() {
     #[pyfunction]
     #[pyo3(warn(message = "this function raises warning"))]
diff --git a/tests/test_pyself.rs b/tests/test_pyself.rs
index 3f7d3844fce..b9b9b0ea4d2 100644
--- a/tests/test_pyself.rs
+++ b/tests/test_pyself.rs
@@ -5,8 +5,7 @@ use pyo3::prelude::*;
 use pyo3::types::{PyBytes, PyString};
 use std::collections::HashMap;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 /// Assumes it's a file reader or so.
 /// Inspired by https://github.com/jothan/cordoba, thanks.
diff --git a/tests/test_sequence.rs b/tests/test_sequence.rs
index ca1b15ae9a5..970f143503f 100644
--- a/tests/test_sequence.rs
+++ b/tests/test_sequence.rs
@@ -6,8 +6,7 @@ use pyo3::{ffi, prelude::*};
 
 use pyo3::py_run;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct ByteSequence {
diff --git a/tests/test_static_slots.rs b/tests/test_static_slots.rs
index b11ecd08fb9..c00a8815270 100644
--- a/tests/test_static_slots.rs
+++ b/tests/test_static_slots.rs
@@ -6,8 +6,7 @@ use pyo3::types::IntoPyDict;
 
 use pyo3::py_run;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct Count5();
diff --git a/tests/test_string.rs b/tests/test_string.rs
index 6ac6615d8a2..1648e067760 100644
--- a/tests/test_string.rs
+++ b/tests/test_string.rs
@@ -2,8 +2,7 @@
 
 use pyo3::prelude::*;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyfunction]
 fn take_str(_s: &str) {}
diff --git a/tests/test_text_signature.rs b/tests/test_text_signature.rs
index d1decc9bf3b..88057b05b6a 100644
--- a/tests/test_text_signature.rs
+++ b/tests/test_text_signature.rs
@@ -4,8 +4,7 @@ use pyo3::prelude::*;
 use pyo3::types::{PyDict, PyTuple};
 use pyo3::{types::PyType, wrap_pymodule};
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[test]
 fn class_without_docs_or_signature() {
diff --git a/src/tests/common.rs b/tests/test_utils/mod.rs
similarity index 75%
rename from src/tests/common.rs
rename to tests/test_utils/mod.rs
index df5af81c247..523bc1ede6f 100644
--- a/src/tests/common.rs
+++ b/tests/test_utils/mod.rs
@@ -1,21 +1,28 @@
+// Common macros and helpers for tests
+//
+// This file is used in two different ways, which makes it a bit of a pain to build:
+// - as a module `include!`-ed from `src/test_utils.rs`
+// - as a module `mod test_utils` in various integration tests
+//
 // the inner mod enables the #![allow(dead_code)] to
-// be applied - `test_utils.rs` uses `include!` to pull in this file
+// be applied - `src/test_utils.rs` uses `include!` to pull in this file
 
-/// Common macros and helpers for tests
-#[allow(dead_code)] // many tests do not use the complete set of functionality offered here
+#[allow(dead_code, unused_macros)] // many tests do not use the complete set of functionality offered here
 #[allow(missing_docs)] // only used in tests
 #[macro_use]
 mod inner {
 
-    #[allow(unused_imports)] // pulls in `use crate as pyo3` in `test_utils.rs`
+    #[allow(unused_imports)]
+    // pulls in `use crate as pyo3` in `src/test_utils.rs`, no function in integration tests
     use super::*;
 
-    #[cfg(not(Py_GIL_DISABLED))]
     use pyo3::prelude::*;
 
-    #[cfg(not(Py_GIL_DISABLED))]
+    use pyo3::sync::MutexExt;
     use pyo3::types::{IntoPyDict, PyList};
 
+    use std::sync::{Mutex, PoisonError};
+
     use uuid::Uuid;
 
     #[macro_export]
@@ -76,31 +83,22 @@ mod inner {
             py_expect_warning!($py, *d, $code, [$(($warning_msg, $warning_category)),+])
         }};
         ($py:expr, *$dict:expr, $code:expr, [$(($warning_msg:literal, $warning_category:path)),+] $(,)?) => {{
-            let code_lines: Vec<&str> = $code.lines().collect();
-            let indented_code: String = code_lines.iter()
-                .map(|line| format!("    {}", line))  // add 4 spaces indentation
-                .collect::<Vec<String>>()
-                .join("\n");
-
-            let wrapped_code = format!(r#"
-import warnings
-with warnings.catch_warnings(record=True) as warning_record:
-{}
-"#, indented_code);
-
-            $py.run(&std::ffi::CString::new(wrapped_code).unwrap(), None, Some(&$dict.as_borrowed())).expect("Failed to run warning testing code");
-            let expected_warnings = [$(($warning_msg, <$warning_category as pyo3::PyTypeInfo>::type_object($py))),+];
-            let warning_record: Bound<'_, pyo3::types::PyList> = $dict.get_item("warning_record").expect("Failed to capture warnings").expect("Failed to downcast to PyList").extract().unwrap();
-
-            assert_eq!(warning_record.len(), expected_warnings.len(), "Expecting {} warnings but got {}", expected_warnings.len(), warning_record.len());
-
-            for ((index, warning), (msg, category)) in warning_record.iter().enumerate().zip(expected_warnings.iter()) {
-                let actual_msg = warning.getattr("message").unwrap().str().unwrap().to_string_lossy().to_string();
-                let actual_category = warning.getattr("category").unwrap();
-
-                assert_eq!(actual_msg, msg.to_string(), "Warning message mismatch at index {}, expecting `{}` but got `{}`", index, msg, actual_msg);
-                assert!(actual_category.is(category), "Warning category mismatch at index {}, expecting {:?} but got {:?}", index, category, actual_category);
-            }
+            $crate::test_utils::CatchWarnings::enter($py, |warning_record| {
+                $py.run(&std::ffi::CString::new($code).unwrap(), None, Some(&$dict.as_borrowed())).expect("Failed to run warning testing code");
+                let expected_warnings = [$(($warning_msg, <$warning_category as pyo3::PyTypeInfo>::type_object($py))),+];
+
+                assert_eq!(warning_record.len(), expected_warnings.len(), "Expecting {} warnings but got {}", expected_warnings.len(), warning_record.len());
+
+                for ((index, warning), (msg, category)) in warning_record.iter().enumerate().zip(expected_warnings.iter()) {
+                    let actual_msg = warning.getattr("message").unwrap().str().unwrap().to_string_lossy().to_string();
+                    let actual_category = warning.getattr("category").unwrap();
+
+                    assert_eq!(actual_msg, msg.to_string(), "Warning message mismatch at index {}, expecting `{}` but got `{}`", index, msg, actual_msg);
+                    assert!(actual_category.is(category), "Warning category mismatch at index {}, expecting {:?} but got {:?}", index, category, actual_category);
+                }
+
+                Ok(())
+            }).expect("failed to test warnings");
         }};
     }
 
@@ -166,17 +164,23 @@ with warnings.catch_warnings(record=True) as warning_record:
         }
     }
 
-    #[cfg(not(Py_GIL_DISABLED))]
     pub struct CatchWarnings<'py> {
         catch_warnings: Bound<'py, PyAny>,
     }
 
-    #[cfg(not(Py_GIL_DISABLED))]
+    /// catch_warnings is not thread-safe, so only one thread can be using this struct at
+    /// a time.
+    static CATCH_WARNINGS_MUTEX: Mutex<()> = Mutex::new(());
+
     impl<'py> CatchWarnings<'py> {
         pub fn enter<R>(
             py: Python<'py>,
             f: impl FnOnce(&Bound<'py, PyList>) -> PyResult<R>,
         ) -> PyResult<R> {
+            // NB this is best-effort, other tests could always call the warnings API directly.
+            let _mutex_guard = CATCH_WARNINGS_MUTEX
+                .lock_py_attached(py)
+                .unwrap_or_else(PoisonError::into_inner);
             let warnings = py.import("warnings")?;
             let kwargs = [("record", true)].into_py_dict(py)?;
             let catch_warnings = warnings
@@ -188,7 +192,6 @@ with warnings.catch_warnings(record=True) as warning_record:
         }
     }
 
-    #[cfg(not(Py_GIL_DISABLED))]
     impl Drop for CatchWarnings<'_> {
         fn drop(&mut self) {
             let py = self.catch_warnings.py();
@@ -198,11 +201,9 @@ with warnings.catch_warnings(record=True) as warning_record:
         }
     }
 
-    #[cfg(not(Py_GIL_DISABLED))]
-    #[macro_export]
     macro_rules! assert_warnings {
         ($py:expr, $body:expr, [$(($category:ty, $message:literal)),+] $(,)? ) => {{
-            $crate::tests::common::CatchWarnings::enter($py, |w| {
+            $crate::test_utils::CatchWarnings::enter($py, |w| {
                 use $crate::types::{PyListMethods, PyStringMethods};
                 $body;
                 let expected_warnings = [$((<$category as $crate::type_object::PyTypeInfo>::type_object($py), $message)),+];
@@ -222,6 +223,8 @@ with warnings.catch_warnings(record=True) as warning_record:
         }};
     }
 
+    pub(crate) use assert_warnings;
+
     pub fn generate_unique_module_name(base: &str) -> std::ffi::CString {
         let uuid = Uuid::new_v4().simple().to_string();
         std::ffi::CString::new(format!("{base}_{uuid}")).unwrap()
diff --git a/tests/test_variable_arguments.rs b/tests/test_variable_arguments.rs
index 2c7851914aa..d186e8def3b 100644
--- a/tests/test_variable_arguments.rs
+++ b/tests/test_variable_arguments.rs
@@ -3,8 +3,7 @@
 use pyo3::prelude::*;
 use pyo3::types::{PyDict, PyTuple};
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct MyClass {}
diff --git a/tests/test_various.rs b/tests/test_various.rs
index c0ed072f18e..cf2ae82b5ff 100644
--- a/tests/test_various.rs
+++ b/tests/test_various.rs
@@ -6,8 +6,7 @@ use pyo3::types::PyTuple;
 
 use std::fmt;
 
-#[path = "../src/tests/common.rs"]
-mod common;
+mod test_utils;
 
 #[pyclass]
 struct MutRefArg {