bpo-33895: Relase GIL while calling functions that acquire Windows loader lock (GH-7789)

LoadLibrary, GetProcAddress, FreeLibrary and GetModuleHandle acquire the system loader lock. Calling these while holding the GIL will cause a deadlock on the rare occasion that another thread is detaching and needs to destroy its thread state at the same time.

4 files changed, 31 insertions, 6 deletions

diff --git a/Modules/_ctypes/_ctypes.c b/Modules/_ctypes/_ctypes.c
index cc4aab7389..0d95d2b6f7 100644
--- a/Modules/_ctypes/_ctypes.c
+++ b/Modules/_ctypes/_ctypes.c
@@ -678,7 +678,9 @@ CDataType_in_dll(PyObject *type, PyObject *args)
     }
 
 #ifdef MS_WIN32
+    Py_BEGIN_ALLOW_THREADS
     address = (void *)GetProcAddress(handle, name);
+    Py_END_ALLOW_THREADS
     if (!address) {
         PyErr_Format(PyExc_ValueError,
                      "symbol '%s' not found",
@@ -3243,18 +3245,23 @@ static PyGetSetDef PyCFuncPtr_getsets[] = {
 #ifdef MS_WIN32
 static PPROC FindAddress(void *handle, const char *name, PyObject *type)
 {
+    PPROC address;
 #ifdef MS_WIN64
     /* win64 has no stdcall calling conv, so it should
        also not have the name mangling of it.
     */
-    return (PPROC)GetProcAddress(handle, name);
+    Py_BEGIN_ALLOW_THREADS
+    address = (PPROC)GetProcAddress(handle, name);
+    Py_END_ALLOW_THREADS
+    return address;
 #else
-    PPROC address;
     char *mangled_name;
     int i;
     StgDictObject *dict;
 
+    Py_BEGIN_ALLOW_THREADS
     address = (PPROC)GetProcAddress(handle, name);
+    Py_END_ALLOW_THREADS
     if (address)
         return address;
     if (((size_t)name & ~0xFFFF) == 0) {
@@ -3275,7 +3282,9 @@ static PPROC FindAddress(void *handle, const char *name, PyObject *type)
         return NULL;
     for (i = 0; i < 32; ++i) {
         sprintf(mangled_name, "_%s@%d", name, i*4);
+        Py_BEGIN_ALLOW_THREADS
         address = (PPROC)GetProcAddress(handle, mangled_name);
+        Py_END_ALLOW_THREADS
         if (address)
             return address;
     }
diff --git a/Modules/_ctypes/callproc.c b/Modules/_ctypes/callproc.c
index bed5364020..9bcc9557ec 100644
--- a/Modules/_ctypes/callproc.c
+++ b/Modules/_ctypes/callproc.c
@@ -1286,7 +1286,10 @@ static PyObject *load_library(PyObject *self, PyObject *args)
     if (!name)
         return NULL;
 
+    Py_BEGIN_ALLOW_THREADS
     hMod = LoadLibraryW(name);
+    Py_END_ALLOW_THREADS
+
     if (!hMod)
         return PyErr_SetFromWindowsErr(GetLastError());
 #ifdef _WIN64
@@ -1303,9 +1306,15 @@ Free the handle of an executable previously loaded by LoadLibrary.\n";
 static PyObject *free_library(PyObject *self, PyObject *args)
 {
     void *hMod;
+    BOOL result;
     if (!PyArg_ParseTuple(args, "O&:FreeLibrary", &_parse_voidp, &hMod))
         return NULL;
-    if (!FreeLibrary((HMODULE)hMod))
+
+    Py_BEGIN_ALLOW_THREADS
+    result = FreeLibrary((HMODULE)hMod);
+    Py_END_ALLOW_THREADS
+
+    if (!result)
         return PyErr_SetFromWindowsErr(GetLastError());
     Py_RETURN_NONE;
 }
diff --git a/Modules/overlapped.c b/Modules/overlapped.c
index ef4390b825..e5a209bf75 100644
--- a/Modules/overlapped.c
+++ b/Modules/overlapped.c
@@ -127,8 +127,10 @@ initialize_function_pointers(void)
     closesocket(s);
 
     /* On WinXP we will have Py_CancelIoEx == NULL */
+    Py_BEGIN_ALLOW_THREADS
     hKernel32 = GetModuleHandle("KERNEL32");
     *(FARPROC *)&Py_CancelIoEx = GetProcAddress(hKernel32, "CancelIoEx");
+    Py_END_ALLOW_THREADS
     return 0;
 }
 
diff --git a/Modules/posixmodule.c b/Modules/posixmodule.c
index 0f6bd14904..931c0d3f3a 100644
--- a/Modules/posixmodule.c
+++ b/Modules/posixmodule.c
@@ -7749,9 +7749,13 @@ check_CreateSymbolicLink(void)
     /* only recheck */
     if (Py_CreateSymbolicLinkW)
         return 1;
+
+    Py_BEGIN_ALLOW_THREADS
     hKernel32 = GetModuleHandleW(L"KERNEL32");
     *(FARPROC*)&Py_CreateSymbolicLinkW = GetProcAddress(hKernel32,
                                                         "CreateSymbolicLinkW");
+    Py_END_ALLOW_THREADS
+
     return Py_CreateSymbolicLinkW != NULL;
 }
 
@@ -11288,7 +11292,6 @@ check_ShellExecute()
            the system SHELL32.DLL, even if there is another SHELL32.DLL
            in the DLL search path. */
         hShell32 = LoadLibraryW(L"SHELL32");
-        Py_END_ALLOW_THREADS
         if (hShell32) {
             *(FARPROC*)&Py_ShellExecuteW = GetProcAddress(hShell32,
                                             "ShellExecuteW");
@@ -11296,6 +11299,7 @@ check_ShellExecute()
         } else {
             has_ShellExecute = 0;
         }
+        Py_END_ALLOW_THREADS
     }
     return has_ShellExecute;
 }
@@ -11909,11 +11913,12 @@ os_cpu_count_impl(PyObject *module)
     /* Vista is supported and the GetMaximumProcessorCount API is Win7+
        Need to fallback to Vista behavior if this call isn't present */
     HINSTANCE hKernel32;
-    hKernel32 = GetModuleHandleW(L"KERNEL32");
-
     static DWORD(CALLBACK *_GetMaximumProcessorCount)(WORD) = NULL;
+    Py_BEGIN_ALLOW_THREADS
+    hKernel32 = GetModuleHandleW(L"KERNEL32");
     *(FARPROC*)&_GetMaximumProcessorCount = GetProcAddress(hKernel32,
         "GetMaximumProcessorCount");
+    Py_END_ALLOW_THREADS
     if (_GetMaximumProcessorCount != NULL) {
         ncpu = _GetMaximumProcessorCount(ALL_PROCESSOR_GROUPS);
     }