[RFC wayland] Add wl_proxy destruction callbacks

[Daniel Vetter]

On Mon, May 30, 2016 at 01:10:42PM +0200, Miguel Angel Vico wrote:
> Hi all,
> 
> A few days ago, I had a little chat over IRC with Pekka about addition
> of proxy objects destruction callbacks to the wayland client protocol.
> 
> 
> Summing up, we recently ran into some applications where native objects
> (wl_surface, wl_egl_window, wl_display) used by EGL are destroyed/made
> invalid before destroying the corresponding EGL objects. This sometimes
> causes crashes of the EGL driver, which is not nice. We have seen this
> with the NVIDIA EGL implementation, but I assume the Mesa EGL
> implementation is similarly exposed.
> 
> I agree this is in fact an application bug, but the EGL spec states that
> functions such as makeCurrent or swapBuffers should return error (not
> crash) if the native objects become invalid. I also agree the spec
> should have been clearer and probably allowed "undefined behavior", but
> it is not the case.
> 
> Having an objects destruction notification mechanism such as destruction
> callbacks would allow us to satisfy the spec.
> 
> Also, such functionality would also make life way easier under certain
> circumstances. I'm basically thinking about multi-threaded applications,
> where several threads make use of the same native objects, and for some
> reason one of the threads has to destroy one or more of them due to some
> sort of error happening.
> 
> Of course, this can still be considered an application bug, and the
> application could still make sure none of the threads is going to use
> the native objects before destroying them, but again, specs allow users
> to do many non-recommended things.
> 
> I think we should try to do our best to gracefully handle those
> non-desirable API usages, and avoid crashes whenever is possible.
> 
> 
> Pekka did not see this as something crazy to have, but wanted to hear
> from some of the toolkits guys before making the decision of whether
> changing the wl_proxy ABI is a good idea.
> 
> 
> As an alternative, destruction callbacks could be hung off of
> wl_egl_window. In a similar way we support wl_egl_window_resize
> callbacks, we could support wl_egl_window_destroy callbacks.
> 
> However, this isn't as foolproof as adding wl_proxy destruction
> callbacks, since destruction of wl_surface or wl_display objects before
> wl_egl_window would lead to same issues.
> 
> 
> I really think adding destruction callbacks to wl_proxy would be an
> improvement worth making, but others' thoughts must be heard first.

Just a quick comment: I guess a piglit testcase to demonstrate the failure
(or well, just any minimal test) would be awesome. That way folks can
quickly figure out what all goes wrong without this.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch