Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix focus handling in multi-window applications #489

Open
wants to merge 2 commits into
base: main
Choose a base branch
from

Conversation

timboudreau
Copy link
Contributor

@timboudreau timboudreau commented Jun 5, 2024

This took some time to debug - starting with an application that opens a popup window, wherein the user should be able to tab between items and press a key to select and close the popup. Focus handling did not work at all in the popup window.

It turned out to be three separate issues:

  1. window_handle.process_central_messages() would discard any messages for Views with a different root than its own, so they would never be processed, so focus requests would be discarded if the main window queried the queue first
  2. ViewId attempted to cache the root view id, but would just cache the deepest parent found - so if you do something that generates an update message during construction of a tree of views, the cached value would not actually be the root view for the (not yet existing) window, and it would always return None from window_id() which relies on the (in this case bogus) root to map it to a window - so a call to request_focus() on a view too early would leave it with garbage as its root id, and no window handle would ever recognize and process it with the wrong root
  3. Not quite as strictly a bug, but I can't imagine why it would be by design: FocusGained and FocusLost events caused by a call to ViewId.request_focus()would be delivered to listeners, but a View would never receive them via event_before_children() or event_after_children() - so this one had a workaround, but the workaround - a view attaching listeners to itself - would be much less readable than a straightforward implementation of one of the event handler methods

The fix for 3. I am the least sure about:

  • It is a bit ugly (swapping an EventCx for an UpdateCx to call unconditional_view_event() and then recreating the UpdateCx)
  • It is not clear if there are other kinds of events which have the same problem of non-delivery to the view itself
  • ViewId.apply_event() might be a cleaner place to implement it (though it would mean a bunch of calling back and forth between WindowHandle and ViewId which might be less than ideal)
  • I do not stop propagation of events if event_*_children() returns EventPropagation::Stop - as long is this is about focus events, I think it is probably harmful to be able to block propagation of focus events at all - window focus events more so, since it's unlikely that one view in a tree knows for sure that none of its siblings need to know about the window losing focus - but the same logic applies more weakly to plain view focus events - it's a recipe for hard-to-diagnose focus bugs.

Since I'm a proud member of the 1990's println school of debugging, UpdateMessage now implements Debug - needed it to prove that retained messages really were eventually processed, and it will probably be useful in the future.

While this method *is* implementable in custom component by copy-paste programming,
it would be preferable for custom components to be able to keep up with any changes
in the default implementation by calling it directly instead.
This took some time to debug - starting with an application that opens a popup window,
wherein the user should be able to tab between items and press a key to select and close
the popup.

It turned out to be **three** separate issues:

1. `window_handle.process_central_messages()` would discard any messages for `View`s with a different root than its own, so they would never be processed.
2. `ViewId` attempted to cache the root view id, but would just cache the deepest parent found - so if you do something that generates an update message
during construction of a tree of views, the cached value would not actually be the root view for the (not yet existing) window
3. Not quite as strictly a bug, but I can't imagine why it would be by design: `FocusGained` and `FocusLost` events caused by a call to `ViewId.request_focus()`
*would* be delivered to listeners, but a `View` would never receive them via `event_before_children()` or `event_after_children()` - so this one had a
workaround, but the workaround - a view attaching listeners to itself - would be much less readable than a straightforward implementation of one of
the event handler methods

The fix for 3. I am the least sure about:

* It is a bit ugly (swapping an `EventCx` for an `UpdateCx` to call `unconditional_view_event()` and then recreating the `UpdateCx`)
* It is not clear if there are *other* kinds of events which have the same problem of non-delivery to the view itself
* `ViewId.apply_event()` might be a cleaner place to implement it (though it would mean a bunch of calling back and forth between `WindowHandle` and
`ViewId` which seems less than ideal)
* I do *not* stop propagation of events if `event_*_children()` returns `EventPropagation::Stop` - as long is this is about focus events, I think it
is probably harmful to be able to block propagation of focus events - window focus events more so, since it's unlikely that one view in a tree knows
for sure that none of its siblings need to know about the window losing focus - but the same logic applies more weakly to plain view focus events -
it's a recipe for hard-to-diagnose focus bugs.

Since I'm a proud member of the 1990's `println` school of debugging, `UpdateMessage` now implements `Debug` - needed it to prove that retained messages
really were eventually processed, and it will probably be useful in the future.
@timboudreau
Copy link
Contributor Author

Ignore the fact that Expose View::default_compute_layout appears to be one of the commits here - I did this on my branch containing that commit, but the commit that applies here reverts it.

@timboudreau
Copy link
Contributor Author

timboudreau commented Jun 5, 2024

Noted in comments, but will mention this here: Retaining messages for unrecognized view roots has the potential to leak ViewIds if there is any situation where it is normal to publish messages to a ViewId that either will never be owned by a window root, or which was but the window was closed before the message was encountered. I.e. if you just called ViewId::new().request_focus() you would add a focus request permanently to CENTRAL_UPDATE_MESSAGES.

You know more about the common use-cases for Floem than I do - is this likely to be a problem?

If so, the simplest gc-like solution would just be to wrap messages in a struct that keeps a counter, increment the counter each time the same message is seen in the queue but not processed, and if it crosses some magic threshold number, don't put it back. Any other solution would involve diffing collections and could get expensive, while if a message hangs out in the queue unprocessed for 100 events (or whatever number seems right), that's not going to have a big impact.

I think probably only implement a solution if we find out that it's actually a problem - fiddling with ViewIds and abandoning them is surely an application-level bug for the most part, although I could imagine a minor case or two where a view races with window removal (like, a click action dispatches request_focus but something else intercepts the click and decides to close the window).

@timboudreau
Copy link
Contributor Author

timboudreau commented Jun 5, 2024

Just in case you're wondering what on earth I'm doing that I'm running into these sorts of issues, here's a screen shot of a prototype of the DSP plugin UI:

(Actually plumbing this thing, given the need to interoperate with pointers handed to me over FFI as if they were RwSignal is a remaining challenge)
Screenshot 2024-06-05 at 7 36 27 PM

@jrmoulton
Copy link
Collaborator

  1. Not quite as strictly a bug, but I can't imagine why it would be by design: FocusGained and FocusLost events caused by a call to ViewId.request_focus()would be delivered to listeners, but a View would never receive them via event_before_children() or event_after_children()

I think is a regression after the ECS pr

Comment on lines +60 to +66
// Rewritten from the original looping version - unless we anticipate view ids
// deeper than the possible stack depth, this should be more efficient and
// require less cloning.
if let Some(p) = self.parent.get(id).unwrap_or(&None) {
self.root_view_id(*p)
} else {
Some(id)

This comment was marked as resolved.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Happy to leave it as-is too. Though I think if you have enough view ids to exhaust the stack, you might have some UI usability problems far worse than hitting that limit :-)

@@ -310,6 +310,9 @@ impl WindowHandle {
}
}
if was_focused != cx.app_state.focus {
// What is this for? If you call ViewId.request_focus(), this
// causes focus to be set to your view, then briefly set to
// None here and then back. Why?
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

the app_state focus will taken from it's option before a pointer down even is propagated. Once event propagation is done, if another view has set itself to have focus because of the pointer down event the focus_changed function will notify the old was_focused that it's focus has been lost and the new app state focus that it has gained focus

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What I found (I added dumping the stack to the focus method it calls if the value was None) is that, when someViewId.request_focus() was called with an existing focus target, focus goes old_id -> None -> new_id.

I didn't check if events were fired for focus going to None, since event propagation wasn't working at that point. As long as we're not doubling up event notifications, it's probably fine - just peculiar.

@timboudreau
Copy link
Contributor Author

  1. Not quite as strictly a bug, but I can't imagine why it would be by design: FocusGained and FocusLost events caused by a call to ViewId.request_focus()would be delivered to listeners, but a View would never receive them via event_before_children() or event_after_children()

I think is a regression after the ECS pr

Not familiar with that PR. Is there a less invasive (or at least less clunky) fix, then?

// event_before_children / event_after_children.
//
// This is a bit messy; could it be done in ViewId.apply_event()
// instead? Are there other cases of dropped events?
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Other events are from winit, so they would be dispatched by the event method, while Event::FocusGained and Event::FocusLost is kind of an "internal" event.

cx = UpdateCx {
app_state: ec.app_state,
};

cx.app_state.focus_changed(old, cx.app_state.focus);
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

unconditional_view_event probably needs to called in focus_changed.

E.g. when the focus got changed due to pointer click.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants