Skip to content
12 changes: 9 additions & 3 deletions .github/workflows/ci.yml
Original file line number Diff line number Diff line change
Expand Up @@ -485,17 +485,23 @@ jobs:
steps:
- uses: actions/checkout@v6

- name: Clean out unused stuff to save space
run: |
sudo rm -rf /usr/local/lib/android /usr/share/dotnet /opt/ghc /opt/hostedtoolcache/CodeQL
sudo apt-get clean

- name: Add NVHPC Repo
run: |
echo 'deb [trusted=yes] https://developer.download.nvidia.com/hpc-sdk/ubuntu/amd64 /' | \
sudo tee /etc/apt/sources.list.d/nvhpc.list

- name: Install 🐍 3 & NVHPC
run: |
sudo apt-get update -y && \
sudo apt-get install -y cmake environment-modules git python3-dev python3-pip python3-numpy && \
sudo apt-get install -y --no-install-recommends nvhpc-25-11 && \
sudo apt-get update -y
sudo apt-get install -y cmake environment-modules git python3-dev python3-pip python3-numpy
sudo apt-get install -y --no-install-recommends nvhpc-25-11
sudo rm -rf /var/lib/apt/lists/*
apt-cache depends nvhpc-25-11
python3 -m pip install --upgrade pip
python3 -m pip install --upgrade pytest

Expand Down
2 changes: 1 addition & 1 deletion include/pybind11/detail/class.h
Original file line number Diff line number Diff line change
Expand Up @@ -207,7 +207,7 @@ extern "C" inline PyObject *pybind11_meta_call(PyObject *type, PyObject *args, P

/// Cleanup the type-info for a pybind11-registered type.
extern "C" inline void pybind11_meta_dealloc(PyObject *obj) {
with_internals([obj](internals &internals) {
with_internals_if_internals([obj](internals &internals) {
auto *type = (PyTypeObject *) obj;

// A pybind11-registered type will:
Expand Down
102 changes: 75 additions & 27 deletions include/pybind11/detail/internals.h
Original file line number Diff line number Diff line change
Expand Up @@ -103,7 +103,7 @@ class thread_specific_storage {
// However, in GraalPy (as of v24.2 or older), TSS is implemented by Java and this call
// requires a living Python interpreter.
#ifdef GRAALVM_PYTHON
if (!Py_IsInitialized() || _Py_IsFinalizing()) {
if (Py_IsInitialized() == 0 || _Py_IsFinalizing() != 0) {
return;
}
#endif
Expand Down Expand Up @@ -195,6 +195,14 @@ struct override_hash {

using instance_map = std::unordered_multimap<const void *, instance *>;

inline bool is_interpreter_alive() {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am sad to report that (AFAICT) this is still not quite correct.

  • Py_IsInitialized and Py_IsFinalizing check whether the whole runtime is initialized/finalizing, not whether a specific interpreter is. It can be invalid to DECREF an object from a particular interpreter even though the runtime is still active.
  • Py_IsFinalizing remains true even after Py_Finalize returns. It does not reset to false unless Py_Initialize is later called again.

I think the correct solution is a internals::leak_detach() method which will set the relevant members to NULL, without decref'ing them, so that a later destructor invocation won't call into the CPython API. Then internals_pp_manager::destroy() can call that method if the internals_pp still contains a valid pointer, before destroying it.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'll point Cursor to this, the pybind11 sources, and the CPython 3.14 sources, and ask it to look very thoroughly to figure out what is the best achievable solution in terms of avoiding UB but leaking as little as possible. I'll report here what it finds.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

d'ooh, I should've caught this before merging it. I just opened #5965 for a possible solution to this by checking the current interpreter instead of the "is alive" check.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

All details and the results are posted here: PR #5966

@b-pass please take what's useful for your PR #5965 (I haven't looked there yet)

#if PY_VERSION_HEX < 0x030D0000
return Py_IsInitialized() != 0 || _Py_IsFinalizing() != 0;
#else
return Py_IsInitialized() != 0 || Py_IsFinalizing() != 0;
#endif
}

#ifdef Py_GIL_DISABLED
// Wrapper around PyMutex to provide BasicLockable semantics
class pymutex {
Expand Down Expand Up @@ -308,7 +316,17 @@ struct internals {
internals(internals &&other) = delete;
internals &operator=(const internals &other) = delete;
internals &operator=(internals &&other) = delete;
~internals() = default;
~internals() {
// Normally this destructor runs during interpreter finalization and it may DECREF things.
// In odd finalization scenarios it might end up running after the interpreter has
// completely shut down, In that case, we should not decref these objects because pymalloc
// is gone.
if (is_interpreter_alive()) {
Py_CLEAR(instance_base);
Py_CLEAR(default_metaclass);
Py_CLEAR(static_property_type);
}
}
};

// the internals struct (above) is shared between all the modules. local_internals are only
Expand All @@ -325,6 +343,16 @@ struct local_internals {

std::forward_list<ExceptionTranslator> registered_exception_translators;
PyTypeObject *function_record_py_type = nullptr;

~local_internals() {
// Normally this destructor runs during interpreter finalization and it may DECREF things.
// In odd finalization scenarios it might end up running after the interpreter has
// completely shut down, In that case, we should not decref these objects because pymalloc
// is gone.
if (is_interpreter_alive()) {
Py_CLEAR(function_record_py_type);
}
}
};

enum class holder_enum_t : uint8_t {
Expand Down Expand Up @@ -569,7 +597,7 @@ inline object get_python_state_dict() {
// The bool follows std::map::insert convention: true = created, false = existed.
template <typename Payload>
std::pair<Payload *, bool> atomic_get_or_create_in_state_dict(const char *key,
bool clear_destructor = false) {
void (*dtor)(PyObject *) = nullptr) {
error_scope err_scope; // preserve any existing Python error states

auto state_dict = reinterpret_borrow<dict>(get_python_state_dict());
Expand All @@ -586,16 +614,13 @@ std::pair<Payload *, bool> atomic_get_or_create_in_state_dict(const char *key,
// Use unique_ptr for exception safety: if capsule creation throws, the storage is
// automatically deleted.
auto storage_ptr = std::unique_ptr<Payload>(new Payload{});
// Create capsule with destructor to clean up when the interpreter shuts down.
auto new_capsule = capsule(
storage_ptr.get(),
// The destructor will be called when the capsule is GC'ed.
// - If our capsule is inserted into the dict below, it will be kept alive until
// interpreter shutdown, so the destructor will be called at that time.
// - If our capsule is NOT inserted (another thread inserted first), it will be
// destructed when going out of scope here, so the destructor will be called
// immediately, which will also free the storage.
/*destructor=*/[](void *ptr) -> void { delete static_cast<Payload *>(ptr); });
auto new_capsule
= capsule(storage_ptr.get(),
// The destructor will be called when the capsule is GC'ed.
// If the insert below fails (entry already in the dict), then this
// destructor will be called on the newly created capsule at the end of this
// function, and we want to just release this memory.
/*destructor=*/[](void *v) { delete static_cast<Payload *>(v); });
// At this point, the capsule object is created successfully.
// Release the unique_ptr and let the capsule object own the storage to avoid double-free.
(void) storage_ptr.release();
Expand All @@ -613,17 +638,16 @@ std::pair<Payload *, bool> atomic_get_or_create_in_state_dict(const char *key,
throw error_already_set();
}
created = (capsule_obj == new_capsule.ptr());
if (clear_destructor && created) {
// Our capsule was inserted.
// Remove the destructor to leak the storage on interpreter shutdown.
if (PyCapsule_SetDestructor(capsule_obj, nullptr) < 0) {
// - If key already existed, our `new_capsule` is not inserted, it will be destructed when
// going out of scope here, and will call the destructor set above.
// - Otherwise, our `new_capsule` is now in the dict, and it owns the storage and the state
// dict will incref it. We need to set the caller's destructor on it, which will be
// called when the interpreter shuts down.
if (created && dtor) {
if (PyCapsule_SetDestructor(capsule_obj, dtor) < 0) {
throw error_already_set();
}
}
// - If key already existed, our `new_capsule` is not inserted, it will be destructed when
// going out of scope here, which will also free the storage.
// - Otherwise, our `new_capsule` is now in the dict, and it owns the storage and the state
// dict will incref it.
}

// Get the storage pointer from the capsule.
Expand Down Expand Up @@ -707,14 +731,27 @@ class internals_pp_manager {
internals_pp_manager(char const *id, on_fetch_function *on_fetch)
: holder_id_(id), on_fetch_(on_fetch) {}

static void internals_shutdown(PyObject *capsule) {
auto *pp = static_cast<std::unique_ptr<InternalsType> *>(
PyCapsule_GetPointer(capsule, nullptr));
if (pp) {
pp->reset();
}
// We reset the unique_ptr's contents but cannot delete the unique_ptr itself here.
// The pp_manager in this module (and possibly other modules sharing internals) holds
// a raw pointer to this unique_ptr, and that pointer would dangle if we deleted it now.
//
// For pybind11-owned interpreters (via embed.h or subinterpreter.h), destroy() is
// called after Py_Finalize/Py_EndInterpreter completes, which safely deletes the
// unique_ptr. For interpreters not owned by pybind11 (e.g., a pybind11 extension
// loaded into an external interpreter), destroy() is never called and the unique_ptr
// shell (8 bytes, not its contents) is leaked.
// (See PR #5958 for ideas to eliminate this leak.)
}

std::unique_ptr<InternalsType> *get_or_create_pp_in_state_dict() {
// The `unique_ptr<InternalsType>` output is leaked on interpreter shutdown. Once an
// instance is created, it will never be deleted until the process exits (compare to
// interpreter shutdown in multiple-interpreter scenarios).
// Because we cannot guarantee the order of destruction of capsules in the interpreter
// state dict, leaking avoids potential use-after-free issues during interpreter shutdown.
Comment on lines -711 to -715
Copy link
Contributor

@XuehaiPan XuehaiPan Jan 21, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Freeing the unique_ptr content is causing use-after-free segmentation faults. Because the order of GC is not guaranteed.

https://github.com/metaopt/optree/actions/runs/21203776041/job/60995252956?pr=245#step:18:229

libc++abi: terminating due to uncaught exception of type pybind11::cast_error: Unable to cast Python instance of type <class 'optree.PyTreeSpec'> to C++ type '?'
  #7  0x00007fb9cc4a5a55 in std::terminate() () from /lib/x86_64-linux-gnu/libstdc++.so.6
  No symbol table info available.
  #8  0x00007fb9cc4bb391 in __cxa_throw () from /lib/x86_64-linux-gnu/libstdc++.so.6
  No symbol table info available.
  #9  0x00007fb9cc7b3990 in pybind11::detail::load_type<optree::PyTreeSpec, void> (conv=..., 
      handle=...) at /usr/include/c++/13/bits/allocator.h:184
  No locals.
  #10 0x00007fb9cc7cdfca in pybind11::detail::load_type<optree::PyTreeSpec&> (handle=...)
      at /tmp/pip-build-env-5okmu3ac/overlay/lib/python3.13t/site-packages/pybind11/include/pybind11/cast.h:1655
          conv = {<pybind11::detail::type_caster_base<optree::PyTreeSpec>> = {<pybind11::detail::type_caster_generic> = {typeinfo = 0x0, cpptype = 0x7fb9cc83a580 <typeinfo for optree::PyTreeSpec>, 
                value = 0x0}, static name = {text = "%"}}, <No data fields>}
  #11 0x00007fb9cc7cdffb in pybind11::cast<optree::PyTreeSpec&, 0> (handle=...)
      at /tmp/pip-build-env-5okmu3ac/overlay/lib/python3.13t/site-packages/pybind11/include/pybind11/cast.h:1680
          is_enum_cast = false
  #12 0x00007fb9cc7ce04d in thread_safe_cast<optree::PyTreeSpec&> (handle=...)
      at /home/runner/work/optree/optree/include/optree/synchronization.h:182
  No locals.
  #13 0x00007fb9cc7f5876 in optree::PyTreeSpec::PyTpTraverse (self_base=0x20002ce0630, 
      visit=0x7fb9ceba61c2 <visit_decref>, arg=0x0)
      at /home/runner/work/optree/optree/src/treespec/gc.cpp:40
          self = <optimized out>
          __PRETTY_FUNCTION__ = "static int optree::PyTreeSpec::PyTpTraverse(PyObject*, visitproc, void*)"

See also:

https://github.com/pybind/pybind11/blob/master/docs/advanced/classes.rst#custom-type-setup

py::cast<CppClass&> is used in tp_traverse and tp_clear, while py::cast is using internals.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@XuehaiPan could you help by extracting a reproducer from your situation? That'll be worth gold. Without reproducers these mind-bending issues will just keep coming back.

Copy link
Collaborator Author

@b-pass b-pass Jan 22, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Running this same stack with -fno-omit-frame-pointer might also help?

My guess is that the py::cast failed and threw an exception because the type is no longer registered with internals (because internals has already destructed).

This is not a use-after-free segv, this is an unhandled/uncaught exception.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I have found that a version bump is necessary for the internal version (used in the key in the interpreter state dict).

In [1]: import env

In [2]: from pybind11_tests import custom_type_setup as m

In [3]: import torch

In [4]: obj = m.OwnsPythonObjects()

In [5]: obj.value = obj

In [6]: exit
[1]    85789 segmentation fault  ipython

torch is built against the old Pybind11 with the same internal version.

Copy link
Contributor

@XuehaiPan XuehaiPan Jan 22, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The problem is that internals and local_internals are freed by pp->reset() before py::cast is called during garbage collection. The sequence of events is:

  1. On interpreter shutdown, everything is garbage collected, including user objects and the capsule that holds internals. However, the GC order is not guaranteed.
  2. The capsule destructor calls pp->reset(), which destroys the internals/local_internals content (the unique_ptr now holds nullptr).
  3. During GC, tp_traverse/tp_clear is invoked on pybind11-wrapped objects, which calls py::cast.
  4. py::cast calls get_type_info, which calls get_internals().
  5. get_internals() finds that *pp is nullptr and creates a new empty internals (with no type registrations).
  6. Type lookup fails, py::cast throws cast_error, and the program terminates.

Each interpreter must not have more than two internals. Now we have an unexpected extra empty internals.


Here is my patch:

diff --git a/include/pybind11/detail/internals.h b/include/pybind11/detail/internals.h
index e2dee77e..26cc8cb7 100644
--- a/include/pybind11/detail/internals.h
+++ b/include/pybind11/detail/internals.h
@@ -20,6 +20,7 @@
 #include <atomic>
 #include <cstdint>
 #include <exception>
+#include <fstream>
 #include <limits>
 #include <mutex>
 #include <thread>
@@ -294,6 +295,11 @@ struct internals {
     internals()
         : static_property_type(make_static_property_type()),
           default_metaclass(make_default_metaclass()) {
+        // Debug logging
+        {
+            std::ofstream dbg("debug.txt", std::ios::app);
+            dbg << "internals::internals() ctor called, this=" << this << "\n";
+        }
         tstate.set(nullptr); // See PR #5870
         PyThreadState *cur_tstate = PyThreadState_Get();
 
@@ -317,6 +323,11 @@ struct internals {
     internals &operator=(const internals &other) = delete;
     internals &operator=(internals &&other) = delete;
     ~internals() {
+        // Debug logging
+        {
+            std::ofstream dbg("debug.txt", std::ios::app);
+            dbg << "internals::~internals() dtor called, this=" << this << "\n";
+        }
         // Normally this destructor runs during interpreter finalization and it may DECREF things.
         // In odd finalization scenarios it might end up running after the interpreter has
         // completely shut down, In that case, we should not decref these objects because pymalloc

Output:

internals::internals() ctor called, this=0xca4c508c0    # Original created
internals::~internals() dtor called, this=0xca4c508c0   # Destroyed by pp->reset()
internals::internals() ctor called, this=0xca1e3f480    # NEW empty one created!

Copy link
Collaborator Author

@b-pass b-pass Jan 22, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is expected behavior of this change whenever something accesses internals late during interpreter shutdown.

We could do a PyGC_Collect() before releasing the internals to try to make this a little less likely, but it could certainly still happen.

We could decide to back out some of this change and stay with a slightly more leaky behavior of pybind11 internals. We did not address all of the leaks yet, default_metaclass is still leaking because it doesn't implement tp_traverse.

auto result = atomic_get_or_create_in_state_dict<std::unique_ptr<InternalsType>>(
holder_id_, /*clear_destructor=*/true);
holder_id_, &internals_shutdown);
auto *pp = result.first;
bool created = result.second;
// Only call on_fetch_ when fetching existing internals, not when creating new ones.
Expand Down Expand Up @@ -834,6 +871,17 @@ inline auto with_internals(const F &cb) -> decltype(cb(get_internals())) {
return cb(internals);
}

template <typename F>
inline void with_internals_if_internals(const F &cb) {
auto &ppmgr = get_internals_pp_manager();
auto &internals_ptr = *ppmgr.get_pp();
if (internals_ptr) {
auto &internals = *internals_ptr;
PYBIND11_LOCK_INTERNALS(internals);
cb(internals);
}
}

template <typename F>
inline auto with_exception_translators(const F &cb)
-> decltype(cb(get_internals().registered_exception_translators,
Expand Down
Loading