Creating windows above just-destroyed windows causes newly created windows to receive invalid stack positions

There's a race condition in the stacking code which can cause windows to receive invalid stack positions.

This sample program shows how that race condition might happen over two display connections:

#include <X11/Xlib.h>
#include <stdio.h>

int main ()
{
    Display *d = XOpenDisplay (NULL);
    Display *other = XOpenDisplay (NULL);
    Window root = DefaultRootWindow (d);
    Window container = XCreateSimpleWindow (d, root, 100, 100, 100, 100, 0, 0, 0);

    Window w1 = XCreateSimpleWindow (d, container, 0, 0, 10, 10, 0, 0xffff00, 0xffff00);
    Window w2 = XCreateSimpleWindow (d, container, 1, 1, 10, 10, 0, 0xff0000, 0xff0000);
    Window w3 = XCreateSimpleWindow (d, container, 2, 2, 10, 10, 0, 0x00ff00, 0x00ff00);

    /* Map all three */
    XMapWindow (d, w1);
    XMapWindow (d, w2);
    XMapWindow (d, w3);

    XMapWindow (d, container);

    /* Destroy 1 */
    XDestroyWindow (d, w1);
    XSync (d, 0);

    /* Stack 3 above 1 */
    XWindowChanges xwc;
    xwc.stack_mode = Above;
    xwc.sibling = w1;
    XConfigureWindow (other, w3, CWSibling | CWStackMode, &xwc);

    Window *children = NULL;
    unsigned int nchildren;
    Window dummy;

    XQueryTree (other, container, &dummy, &dummy, &children, &nchildren);
    int i = 0;
    for (i = 0; i < nchildren; ++i)
    {
 printf ("0x%x\n", (unsigned int) children[i]);
    }

    XCloseDisplay (d);
    XCloseDisplay (other);
    return 0;
}

The problem is that configuring a window to be stacked relative to a destroyed window is an invalid operation, however, there's a race condition where if we are between handling events, we may find a suitable candidate to stack above, however we may not have received a DestroyNotify for that candidate yet. If we issue a ConfigureWindow request to stack above that candidate, it will fail with a BadWindow error when we later flush the output buffer, however we didn't really have a way of knowing it was going to fail (and furthermore, we ignore X errors, a bad thing in itself, but a separate bug), so that would cause us to mark the window as having successfully being restacked internally whereas in reality it wasn't restacked at all. Then when we restack other windows relative to the window that failed the restack, they are going to receive positions which are relative to its /original/ position (and not its changed position), which is usually above everything else and invalid.

From a user point of view, I can reproduce this by opening ~20 or so gnome-terminals and restarting compiz in the classic session. The terminals will end up above the panels.