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.

1 files changed, 8 insertions, 3 deletions

diff --git a/Modules/posixmodule.c b/Modules/posixmodule.c
index 0f6bd14..931c0d3 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);
     }