Bug: Issues with write barrier for __newindex

In 'luaV_finishset', there is an update on a table that is a field on
another table. If the first table is the same as the one with the field
(e.g., after 't.__newindex = t'), the update can change the value on
that field (e.g., there may be a collision and the field is moved, or
the field being updated is '__newindex' itself). After that, the
barrier is called with the table stored in that field, which is not
the correct table anymore.

lvm.c

@@ -360,13 +360,19 @@ void luaV_finishset (lua_State *L, const TValue *t, TValue *key,
       luaT_callTM(L, tm, t, key, val);
       return;
     }
-    t = tm;  /* else repeat assignment over 'tm' */
-    luaV_fastset(t, key, val, hres, luaH_pset);
-    if (hres == HOK) {
-      luaV_finishfastset(L, t, val);
-      return;  /* done */
+    t = tm;  /* else must repeat assignment over 'tm' */
+    /* do the equivalent to 'luaV_fastset', but saving 'h' */
+    if (!ttistable(t))
+      hres = HNOTATABLE;
+    else {
+      Table *h = hvalue(t);  /* next call can change the value at 't' */
+      hres = luaH_pset(h, key, val);
+      if (hres == HOK) {
+        luaC_barrierback(L, obj2gco(h), val);  /* luaV_finishfastset */
+        return;  /* done */
+      }
     }
-    /* else 'return luaV_finishset(L, t, key, val, slot)' (loop) */
+    /* else 'return luaV_finishset(L, t, key, val, hres)' (loop) */
   }
   luaG_runerror(L, "'__newindex' chain too long; possible loop");
 }

testes/events.lua

@@ -390,6 +390,18 @@ do
   for i=1, 10 do t[i] = 1 end
 end
 
+
+do   -- bug since 5.4
+  local parent = {}
+  parent.__newindex = parent
+  collectgarbage()
+  local child = setmetatable({}, parent)
+  child.__newindex = {x = "hello"}
+  collectgarbage("step")
+  assert(parent.__newindex.x == "hello")
+end
+
+
 -- concat metamethod x numbers (bug in 5.1.1)
 c = {}
 local x