Replace RenderFullScreen with top layer

Replace RenderFullScreen with top layer

This patch makes the Fullscreen API use the top layer, as per the latest
Fullscreen spec.

The previous attempt apparently broke plugin reloading on setAttribute,
but there is not a simple test case for it (empirically, a plugin-based
app broke after the change). The cause was the m_protectWidgetDuringReattach
mechanism that saves a plugin from reloading when entering/exiting fullscreen.
In this case, the m_protectWidgetDuringReattach mechanism inappropriately
kicked in during a reattach triggered by recalcStyle.

The fix is to only set m_protectWidgetDuringReattach in setIsInTopLayer, rather
than in detach.

Previously landed as r165904, r165710, 166304
Codereviews:
https://codereview.chromium.org/134753003/
https://codereview.chromium.org/139743005/
https://codereview.chromium.org/142653003/

BUG=240576, 246077

[falken, 2014-02-05 09:38:56]

eseidel: could you review this in esprehn's absence?

This makes a small change from the reverted patch
(https://codereview.chromium.org/142653003/). You can see the delta from
patchset #2 to see the change.

[esprehn, 2014-02-11 00:25:15]

This patch is not correct, what if you full screen an ancestor of the plugin?

https://codereview.chromium.org/138833005/diff/200001/Source/core/html/HTMLPl...
File Source/core/html/HTMLPlugInElement.cpp (right):

https://codereview.chromium.org/138833005/diff/200001/Source/core/html/HTMLPl...
Source/core/html/HTMLPlugInElement.cpp:302: m_protectWidgetDuringReattach =
existingRenderWidget()->widget();
This doesn't seem right, what if you have <div><object ...></div>

Now the plugin will reload since it's not the thing that's going into the top
layer, its parent is.

[falken, 2014-02-11 14:39:57]

https://codereview.chromium.org/138833005/diff/200001/Source/core/html/HTMLPl...
File Source/core/html/HTMLPlugInElement.cpp (right):

https://codereview.chromium.org/138833005/diff/200001/Source/core/html/HTMLPl...
Source/core/html/HTMLPlugInElement.cpp:302: m_protectWidgetDuringReattach =
existingRenderWidget()->widget();
On 2014/02/11 00:25:16, esprehn wrote:
> This doesn't seem right, what if you have <div><object ...></div>
> 
> Now the plugin will reload since it's not the thing that's going into the top
> layer, its parent is.

Ah, good point. How about adding "topLayerChanged" to AttachContext, or else
making Element::setIsInTopLayer traverse descendants and have any
HTMLPlugInElements protect their widgets?

From the breakage caused by the previous patch, it seems reloading the plugin
even when needsWidgetUpdate is false during a reattach is necessary in some
cases. So HTMLPlugInElement somehow needs to know when the reattach was caused
by something entering/exiting top layer.

[esprehn, 2014-03-06 06:59:30]

On 2014/02/11 at 14:39:57, falken wrote:
>
https://codereview.chromium.org/138833005/diff/200001/Source/core/html/HTMLPl...
> File Source/core/html/HTMLPlugInElement.cpp (right):
> 
>
https://codereview.chromium.org/138833005/diff/200001/Source/core/html/HTMLPl...
> Source/core/html/HTMLPlugInElement.cpp:302: m_protectWidgetDuringReattach =
existingRenderWidget()->widget();
> On 2014/02/11 00:25:16, esprehn wrote:
> > This doesn't seem right, what if you have <div><object ...></div>
> > 
> > Now the plugin will reload since it's not the thing that's going into the
top
> > layer, its parent is.
> 
> Ah, good point. How about adding "topLayerChanged" to AttachContext, or else
making Element::setIsInTopLayer traverse descendants and have any
HTMLPlugInElements protect their widgets?
> 
> From the breakage caused by the previous patch, it seems reloading the plugin
even when needsWidgetUpdate is false during a reattach is necessary in some
cases. So HTMLPlugInElement somehow needs to know when the reattach was caused
by something entering/exiting top layer.

I've been trying to understand this more, how did the old full screen not
reattach everything? The wrapping for full screen happened inside
NodeRenderingContext (now RenderTreeBuilder) so it's not clear why plugins
didn't restart then?

We might instead add an RAII thing that we use when toggling the top layer, so
you do:

ex.

goFullScreen() {
  ProtectPluginScope scope;
  updateStyleIfNeeded();
}

but that does mean forcing the full screen update immediately.

[falken, 2014-03-06 16:15:59]

On 2014/03/06 06:59:30, esprehn wrote:
> On 2014/02/11 at 14:39:57, falken wrote:
> >
>
https://codereview.chromium.org/138833005/diff/200001/Source/core/html/HTMLPl...
> > File Source/core/html/HTMLPlugInElement.cpp (right):
> > 
> >
>
https://codereview.chromium.org/138833005/diff/200001/Source/core/html/HTMLPl...
> > Source/core/html/HTMLPlugInElement.cpp:302: m_protectWidgetDuringReattach =
> existingRenderWidget()->widget();
> > On 2014/02/11 00:25:16, esprehn wrote:
> > > This doesn't seem right, what if you have <div><object ...></div>
> > > 
> > > Now the plugin will reload since it's not the thing that's going into the
> top
> > > layer, its parent is.
> > 
> > Ah, good point. How about adding "topLayerChanged" to AttachContext, or else
> making Element::setIsInTopLayer traverse descendants and have any
> HTMLPlugInElements protect their widgets?
> > 
> > From the breakage caused by the previous patch, it seems reloading the
plugin
> even when needsWidgetUpdate is false during a reattach is necessary in some
> cases. So HTMLPlugInElement somehow needs to know when the reattach was caused
> by something entering/exiting top layer.
> 
> I've been trying to understand this more, how did the old full screen not
> reattach everything? The wrapping for full screen happened inside
> NodeRenderingContext (now RenderTreeBuilder) so it's not clear why plugins
> didn't restart then?

Looking at http://trac.webkit.org/changeset/95371

It looks like by directly doing remove() then addChild(), a reattach doesn't
occur.

Another possibility I was considering is if we could somehow make
entering/exiting top layer not require a reattach...