Fix c_stack_hard_limit: add _PyOS_STACK_MARGIN_BYTES

Author: vstinner

Modules/_testinternalcapi.c

@@ -2425,7 +2425,7 @@ check_threadstate_set_stack(PyThreadState *tstate, void *start, size_t size)
     assert(!PyErr_Occurred());
 
     _PyThreadStateImpl *ts = (_PyThreadStateImpl *)tstate;
-    assert(ts->c_stack_hard_limit == (uintptr_t)start);
+    assert(ts->c_stack_hard_limit == (uintptr_t)start + _PyOS_STACK_MARGIN_BYTES);
     assert(ts->c_stack_top == (uintptr_t)start + size);
     assert(ts->c_stack_soft_limit >= ts->c_stack_hard_limit);
     assert(ts->c_stack_soft_limit < ts->c_stack_top);

Python/ceval.c

@@ -446,9 +446,10 @@ tstate_set_stack(PyThreadState *tstate,
     assert(stack_size > 0);
     assert(stack_size >= (_PyOS_STACK_MARGIN_BYTES * 3));
 
+    uintptr_t start = (uintptr_t)stack_start_addr;
     _PyThreadStateImpl *ts = (_PyThreadStateImpl *)tstate;
-    ts->c_stack_hard_limit = (uintptr_t)stack_start_addr;
-    ts->c_stack_top = (uintptr_t)stack_start_addr + stack_size;
+    ts->c_stack_hard_limit = start + _PyOS_STACK_MARGIN_BYTES;
+    ts->c_stack_top = start + stack_size;
 
     uintptr_t soft_limit = ts->c_stack_hard_limit;
 #ifdef _Py_THREAD_SANITIZER
@@ -498,7 +499,7 @@ PyUnstable_ThreadState_ResetStack(PyThreadState *tstate)
     ULONG guarantee = 0;
     SetThreadStackGuarantee(&guarantee);
 
-    uintptr_t start = (uintptr_t)low + guarantee + _PyOS_STACK_MARGIN_BYTES;
+    uintptr_t start = (uintptr_t)low + guarantee;
     size_t size = (uintptr_t)high - start;
     tstate_set_stack(tstate, (void*)start, size);
 
@@ -525,7 +526,7 @@ PyUnstable_ThreadState_ResetStack(PyThreadState *tstate)
     }
     if (err == 0) {
         uintptr_t base = ((uintptr_t)stack_addr) + guard_size;
-        uintptr_t start = base + _PyOS_STACK_MARGIN_BYTES;
+        uintptr_t start = base;
         size_t pystack_size = (base + stack_size) - start;
         tstate_set_stack(tstate, (void*)start, pystack_size);
     }
@@ -534,7 +535,7 @@ PyUnstable_ThreadState_ResetStack(PyThreadState *tstate)
     {
         uintptr_t here_addr = _Py_get_machine_stack_pointer();
         uintptr_t top = _Py_SIZE_ROUND_UP(here_addr, 4096);
-        uintptr_t start = top - (Py_C_STACK_SIZE + _PyOS_STACK_MARGIN_BYTES);
+        uintptr_t start = top - Py_C_STACK_SIZE;
         size_t pystack_size = top - start;
         tstate_set_stack(tstate, (void*)start, pystack_size);
     }