244366 - [FIXr]Invalidate event processing should flush reflows

Status: RESOLVED FIXED

(Core :: Web Painting, defect, P1)

Description

[Boris Zbarsky [:bzbarsky]]

The short story is that the fix for bug 36849 introduced a hack where we flush
reflows and repaints on mousemove events to prevent flickering.  It seems like a
more comprehensive solution would be as following (in the viewmanager):

1) We get ready to process invalidates
2) We tell the presshell to flush reflows (which incidentally will do all its
   invalidating in a view batch, if we do it right).
3) We process our invalidate queue. 

During step 2, posting of an invalidate event is suppressed (since we're about
to process one anyway).

One possible issue with this approach is that it may reduce responsiveness...
Would need to test.

Comment 1

[Boris Zbarsky [:bzbarsky]]

Attachment: First cut at a patch

This is quite raw; the event state manager change, in particular, is just
wrong.

Comment 2

[Boris Zbarsky [:bzbarsky]]

There's a problem here, by the way... we could flush reflows during invalidate
processing, but new reflows could happen before we go to paint, without the new
invalidates having happened yet.  So we really want to flush all the reflows and
invalidates when we get the paint event from widget or something...  Except that
call gives us the area to paint, and that area would need to be changed in the
call itself.

Robert, am I missing something, or does that sum up what needs to happen?

Comment 3

[Boris Zbarsky [:bzbarsky]]

Attachment: Slightly more advanced

Comment 4

[Boris Zbarsky [:bzbarsky]]

Attachment: Proposed patch

Summary of changes:

1)  Add a WillPaint notification to nsIViewObserver.
2)  Call this from two places: FlushPendingInvalidates, in the hope that this
will coalesce the invalidates we already have pending and the invalidates the
reflow will generate, and WM_PAINT event handling.  The latter is definitely
needed to solve some jitter and blinking problems we had.
3)  Remove the FlushPendingEvents calls we had on every mousemove (but keep
them while we're in a drag-drop gesture).
4)  Change presshell code that could tear down the presshell from inside
ProcessReflowCommands to not do that (add a check like the nsObjectFrame code
had, and consolidate the identical code in a prescontext method).  I've been
meaning to fix this anyway, since it could cause crashes; these just became
more likely when we were reflowing from inside ESM code and WM_PAINT code and
such.  I put the method on nsPresContext instead of nsIPresShell because that's
a more "layout-private" class as far as I can tell, but I don't have a strong
attachment to this placement.

Comment 5

[Boris Zbarsky [:bzbarsky]]

I's sort of like to get this in for beta, but it's a little scary...

Comment 6

[Martijn Wargers (dead)]

With the Proposed patch, I crash by just following bug 277653. Without it, I
don't crash with my Firefox non-debug build.
The testcase and the url in bug 277653 work nice, though. :)

Comment 7

[Boris Zbarsky [:bzbarsky]]

Martijn, what are the exact steps to reproduce that crash you're seeing?  Can
you get a stack, possibly?

Either put that info in this bug or mail me?

Comment 8

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 170800 [details] [diff] [review]
Proposed patch

This has a problem (the one Martijn pointed out; the upshot is that editor
actually forces layout and paint from inside frame construction, so we end up
reflowing in the middle of frame construction, which is bad).

Comment 9

[Martijn Wargers (dead)]

Attachment: Testcase that crashes Mozilla when "Proposed patch" is built in

This is a minimal testcase that makes Mozilla crash, when you've applied the
"Proposed patch" patch.

Comment 10

[Boris Zbarsky [:bzbarsky]]

Attachment: Slightly more bulletproofed

Changes, in addition to those listed in comment 4:

5)  Make editor end its view update batch as DEFERRED when it's not doing sync
reflow/painting.  This is what actually fixes the crash Martijn ran into (which
I didn't see in my build because it had unrelated text control changes).  The
comment here explains why this is the right thing to do in any case.
6)  Fix a bug in view/viewmanager where we'd bail out of
nsView::ResetWidgetBounds on the assumption that the end of the batch would
process the change, but didn't let the viewmanager know that it had pending
changes.  This caused tooltips to not appear, for example, since with this
patch we always check mHasPendingUpdates (which I have accordingly renamed)
before calling ProcessPendingUpdates.
7)  Even though the editor change fixed the crash Martijn saw, I decided to add
some more bulletproofing and not reflow on paint while in the middle of things
bracketed by WillCauseReflow/DidCauseReflow (mostly frame construction; the
other stuff bracketed by these was already protected by mIsReflowing).	While I
was there, I also disabled flushing reflows in the same circumstances.

I've tested this on a bunch of DHTML testcases now, as well as on various
normal pages, with and without plugins, in a vanilla build with nothing but
this patch applied.  So far, so good.  I'll keep testing it, just in case....

Comment 11

[Robert O'Callahan (:roc) (email my personal email if necessary)]

Comment on attachment 171103 [details] [diff] [review]
Slightly more bulletproofed

go for it.

Comment 12

[Boris Zbarsky [:bzbarsky]]

Checked in.  The only effect on perf metrics was that Tdhtml went up a bit.  I
looked into it, and a few of the tests got slower (the rest did not).  Testing
in a patched and unpatched build, all the tests that got slower were skipping a
lot more animation frames without this patch -- this patch makes the animation
much smoother.

Comment 13

[Boris Zbarsky [:bzbarsky]]

This seems to have caused at least bug 279308 and bug 279205.

Comment 14

[Boris Zbarsky [:bzbarsky]]

Bug 279205 is probably caused by this change to nsViewManager in this patch:

-        nsCOMPtr<nsIViewObserver> vobs;
-        vVM->GetViewObserver(*getter_AddRefs(vobs));
+        // Hold a refcount to the observer. The continued existence of the
observer will
+        // delay deletion of this view hierarchy should the event want to cause its
+        // destruction in, say, some JavaScript event handler.
+        nsCOMPtr<nsIViewObserver> vobs = GetViewObserver();

Note that this gets the wrong view observer.

Comment 15

[Boris Zbarsky [:bzbarsky]]

This also caused bug 279786

Comment 16

[Marco Perez [:mpk]]

The fix seems to have caused bug #280214 as well.

Comment 17

[Boris Zbarsky [:bzbarsky]]

Adding regressions caused or maybe caused by this to the "depends on" list.

Comment 18

[Steve England [:stevee]]

This patch may have caused bug 281173

Comment 19

[Scott MacGregor]

I believe this patch caused a regression in the mozilla application suite and
for Thunderbird regarding excessive black flicker. I narrowed the regression
window down to the day this change went in. See Bug #283123 for details.