Revert "[BE] Make PyObjectSlot use a global PyInterpreter (#162659)"

This reverts commit 29cbcbac4215e0d9070a1b7a07ddaec9a36bbd08.

Reverted https://github.com/pytorch/pytorch/pull/162659 on behalf of https://github.com/izaitsevfb due to reverted internally, see [D83214133](https://www.internalfb.com/diff/D83214133) ([comment](https://github.com/pytorch/pytorch/pull/162659#issuecomment-3369348172))

c10/core/TensorImpl.h

@@ -3244,7 +3244,7 @@ class C10_TensorImpl_Size_Check_Dummy_Class : private TensorImpl {
     are_equal<sizeof(autograd_meta_),      4,  FieldNameEnum::autograd_meta_>();
     are_equal<sizeof(extra_meta_),         4,  FieldNameEnum::extra_meta_>();
     are_equal<sizeof(version_counter_),    4,  FieldNameEnum::version_counter_>();
-    are_equal<sizeof(pyobj_slot_),         4,  FieldNameEnum::pyobj_slot_>();
+    are_equal<sizeof(pyobj_slot_),    8,  FieldNameEnum::pyobj_slot_>();
     is_le<sizeof(sizes_and_strides_),     88, FieldNameEnum::sizes_and_strides_>();
     are_equal<sizeof(storage_offset_),     8,  FieldNameEnum::storage_offset_>();
     are_equal<sizeof(numel_),              8,  FieldNameEnum::numel_>();
@@ -3269,7 +3269,7 @@ class C10_TensorImpl_Size_Check_Dummy_Class : private TensorImpl {
     is_le<sizeof(autograd_meta_),         16,  FieldNameEnum::autograd_meta_>();
     is_le<sizeof(extra_meta_),            16,  FieldNameEnum::extra_meta_>();
     are_equal<sizeof(version_counter_),    8,  FieldNameEnum::version_counter_>();
-    are_equal<sizeof(pyobj_slot_),         8,  FieldNameEnum::pyobj_slot_>();
+    are_equal<sizeof(pyobj_slot_),   16,  FieldNameEnum::pyobj_slot_>();
     are_equal<sizeof(sizes_and_strides_), 88,  FieldNameEnum::sizes_and_strides_>();
     are_equal<sizeof(storage_offset_),     8,  FieldNameEnum::storage_offset_>();
     are_equal<sizeof(numel_),              8,  FieldNameEnum::numel_>();

c10/core/impl/PyInterpreterHooks.h

@@ -13,10 +13,11 @@ struct C10_API PyInterpreterHooksInterface {
 
   // Get the PyInterpreter instance
   // Stub implementation throws error when Python is not available
-  // We return nullptr rather than throwing an error since there are bits of c10
-  // that expect an empty PyObjectSlot when python is not available.
   virtual PyInterpreter* getPyInterpreter() const {
-    return nullptr;
+    TORCH_CHECK(
+        false,
+        "PyTorch was compiled without Python support. "
+        "Cannot access Python interpreter from C++.");
   }
 };
 

c10/core/impl/PyObjectSlot.cpp

@@ -2,7 +2,7 @@
 
 namespace c10::impl {
 
-PyObjectSlot::PyObjectSlot() : pyobj_(nullptr) {}
+PyObjectSlot::PyObjectSlot() : pyobj_interpreter_(nullptr), pyobj_(nullptr) {}
 
 PyObjectSlot::~PyObjectSlot() {
   maybe_destroy_pyobj();
@@ -10,9 +10,9 @@ PyObjectSlot::~PyObjectSlot() {
 
 void PyObjectSlot::maybe_destroy_pyobj() {
   if (owns_pyobj()) {
-    TORCH_INTERNAL_ASSERT(getGlobalPyInterpreter() != nullptr);
+    TORCH_INTERNAL_ASSERT(pyobj_interpreter_ != nullptr);
     TORCH_INTERNAL_ASSERT(pyobj_ != nullptr);
-    (*getGlobalPyInterpreter())
+    (*pyobj_interpreter_.load(std::memory_order_acquire))
         ->decref(_unchecked_untagged_pyobj(), /*has_pyobj_slot*/ true);
     // NB: this destructor can only be entered when there are no
     // references to this C++ object (obviously), NOR any references
@@ -25,7 +25,7 @@ void PyObjectSlot::maybe_destroy_pyobj() {
 }
 
 PyInterpreter* PyObjectSlot::pyobj_interpreter() {
-  return getGlobalPyInterpreter();
+  return pyobj_interpreter_.load(std::memory_order_acquire);
 }
 
 PyObject* PyObjectSlot::_unchecked_untagged_pyobj() const {
@@ -35,7 +35,7 @@ PyObject* PyObjectSlot::_unchecked_untagged_pyobj() const {
 }
 
 PyInterpreter& PyObjectSlot::load_pyobj_interpreter() const {
-  auto interpreter = getGlobalPyInterpreter();
+  auto interpreter = pyobj_interpreter_.load(std::memory_order_acquire);
   if (interpreter) {
     return *interpreter;
   }

c10/core/impl/PyObjectSlot.h

@@ -6,17 +6,10 @@
 #include <c10/util/python_stub.h>
 #include <optional>
 
+#include <atomic>
+
 namespace c10::impl {
 
-// Function pointer type for getting the global interpreter
-using GetPyInterpreterFn = PyInterpreter* (*)();
-
-// Global function pointer (set by csrc initialization)
-C10_API extern GetPyInterpreterFn g_get_pyinterpreter_fn;
-
-// Helper function to get the global interpreter
-C10_API PyInterpreter* getGlobalPyInterpreter();
-
 struct C10_API PyObjectSlot {
  public:
   PyObjectSlot();
@@ -33,6 +26,8 @@ struct C10_API PyObjectSlot {
   // NB: THIS FUNCTION CAN RAISE AN EXCEPTION.  Make sure to clean up after
   // PyObject if necessary!
   void init_pyobj(PyObject* pyobj) {
+    pyobj_interpreter_.store(
+        getGlobalPyInterpreter(), std::memory_order_relaxed);
     pyobj_ = pyobj;
   }
 
@@ -60,15 +55,18 @@ struct C10_API PyObjectSlot {
 
   // @todo alban: I'm not too sure what's going on here, we can probably delete
   // it but it's worthwhile making sure
-  std::optional<PyObject*> check_pyobj() const {
-    impl::PyInterpreter* interpreter = getGlobalPyInterpreter();
-    if (interpreter == nullptr || pyobj_ == nullptr) {
+  std::optional<PyObject*> check_pyobj(bool ignore_hermetic_tls = false) const {
+    impl::PyInterpreter* interpreter =
+        pyobj_interpreter_.load(std::memory_order_acquire);
+    if (interpreter == nullptr) {
       return std::nullopt;
     }
-    if (c10::impl::HermeticPyObjectTLS::get_state()) {
+
+    if (!ignore_hermetic_tls && c10::impl::HermeticPyObjectTLS::get_state()) {
       return std::nullopt;
+    } else {
+      return _unchecked_untagged_pyobj();
     }
-    return _unchecked_untagged_pyobj();
   }
 
   PyInterpreter& load_pyobj_interpreter() const;
@@ -78,6 +76,30 @@ struct C10_API PyObjectSlot {
   void set_owns_pyobj(bool b);
 
  private:
+  // This field contains the interpreter tag for this object.  See
+  // Note [Python interpreter tag] for general context
+  //
+  // Note [Memory ordering on Python interpreter tag]
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // What memory_order do we need when accessing this atomic?  We don't
+  // need a single total modification order (as provided by
+  // memory_order_seq_cst) as pyobj_interpreter_ is monotonic: it can only
+  // transition from -1 to some positive integer and never changes afterwards.
+  // Because there is only one modification, it trivially already has a total
+  // modification order (e.g., we don't need fences or locked instructions on
+  // x86)
+  //
+  // In fact, one could make a reasonable argument that relaxed reads are OK,
+  // due to the presence of external locking (GIL) to ensure that interactions
+  // with other data structures are still correctly synchronized, so that
+  // we fall in the "Single-Location Data Structures" case as described in
+  // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p2055r0.pdf
+  // However, on x86, it doesn't matter if I use acquire or relaxed on the load
+  // as I get the same assembly in both cases.  So I just use the more
+  // conservative acquire (which will impede compiler optimizations but I don't
+  // care)
+  std::atomic<PyInterpreter*> pyobj_interpreter_;
+
   // This field contains a reference to a PyObject representing this Tensor.
   // If pyobj is nullptr, when we transfer Tensor to Python, we allocate a new
   // PyObject for it and set this field.  This field does not have to be