[RFC xserver 0/3] Allow XWM to control Xwayland commits

Mon Nov 28 13:47:11 UTC 2016

On Fri, 25 Nov 2016 12:30:18 +0200
Pekka Paalanen <ppaalanen at gmail.com> wrote:

> On Fri, 25 Nov 2016 03:47:56 -0500 (EST)
> Olivier Fourdan <ofourdan at redhat.com> wrote:
> 
> > Hi Pekka,
> >   
> > > this is probably the first time I'm sending patches for the xserver, so
> > > pointing out API misuse, coding style issues etc. would be appreciated. The
> > > last patch also has some XXX comments with questions.
> > > 
> > > The first patch, refactoring, is not absolutely necessary (anymore - I used
> > > to
> > > have a use for it while brewing this), but I think it makes things look
> > > nicer.
> > > 
> > > The fundamental problem is that Xwayland and the Wayland compositor (window
> > > manager) are racing, particularly when windows are being mapped. This can
> > > lead
> > > to glitches. The specific glitch I bumped into is described at
> > > https://phabricator.freedesktop.org/T7622 .
> > > 
> > > The proposed solution allows the WM to temporarily prohibit commits.
> > > 
> > > It would be awkward for Weston to postpone mapping certain windows since it
> > > would leak quite much Xwayland into the heart of the window manager. A
> > > Wayland
> > > compositor also cannot ignore commits, because Xwayland will wait for frame
> > > callbacks until it commits again. To not get stuck, one would need to fire
> > > frame callbacks outside of their normal path. It seems much simpler to just
> > > control Xwayland's commits, which also ensures that the first commit will
> > > carry
> > > fully drawn window decorations too.
> > > 
> > > This series is the Xwayland side of the fix to T7622. The Weston side is
> > > still
> > > brewing, but I was able to test that the Xwayland code works.    
> > 
> > I've taken a look at https://phabricator.freedesktop.org/T7622 but I
> > am not sure how that proposal would play with apps that implement the
> > _NET_WM_SYNC protocol, using _NET_WM_SYNC_REQUEST and sync counters
> > as described in EMWH [1]  
> 
> Hi,
> 
> I have no idea. I don't know how that protocol works and Weston does
> not use it. Yet.
> 
> > GNOME (mutter,gtk3) go a bit farther even and use an extended version
> > of this protocol described by Owen [2].
> > 
> > I suspect these make the de-syncronization with the Xwayland toplevel
> > decorated by the X window manager even more visible under Wayland, as
> > the repaint is scheduled according to the app content and not sync'ed
> > with Xwayland and repaint of the shadows by the XWM.
> > 
> > Would your proposal help there?   
> 
> I would hope so, but given that I don't know how the sync protocol
> works, I can't say. So far I'm just following the plan Daniel made.
> 
> From what Daniel wrote about the sync protocol in T7622, it sounds like
> "my" proposal is the first step toward making the sync protocol really
> successful.
> 
> I've been wondering though whether the client content sync
> protocol should be implemented in the WM or could/should it be
> implemented in Xwayland? Somehow it would feel natural to me, being a
> foreigner to X11, that Xwayland would throttle its wl_surface.commits
> based on both the client content sync protocol and the WM sync protocol
> (this series). I don't think we would want to route every single commit
> through the WM.
> 
> 
> Thanks,
> pq
> 
> > [1]
> > https://specifications.freedesktop.org/wm-spec/wm-spec-latest.html#idm140200472538288
> > [2] http://fishsoup.net/misc/wm-spec-synchronization.html

Hi,

having read the above two specs, it is very much not obvious how to
connect all the dots. It needs experimenting.

Would it be acceptable to build the knowledge of all these sync
protocols into Xwayland server?

The thing there is that Xwayland is issuing the wl_surface.commits, and
having XWM to explicitly trigger all commits always seems like
unnecessary overhead. However, Xwayland does not know when to commit if
it doesn't understand the X11 sync protocols. Either every commit
for every window has to roundtrip through XWM, or Xwayland needs to
understand stuff. Or the third option, which I like the least, is to
have the Wayland compositor special-case all surface from Xwayland and
put them through a completely different and new code paths not shared
with anything.

The basic frame counter only throttles resizes, so that would be mostly
between XWM and the X11 app. Xwayland could automatically stop
committing when the counter indicates the app is busy reacting to a
resize, and resume committing when the app catches up.

The extended frame counter seems fairly straightforward at first hand
to integrate as a commit prohibitor, like _XWAYLAND_ALLOW_COMMITS.

I'm not sure if _NET_WM_FRAME_DRAWN should correspond to Wayland frame
callbacks (which Xwayland could then handle on its own) or sent by XWM
when it has drawn the decorations. The former would be more
straightforward, the latter might allow using FRAME_DRAWN for
triggering the commit but somehow I have a feeling it might not be a
good idea...

_NET_WM_FRAME_TIMINGS looks like something we could provide with the
Presentation feedback protocol straight from Xwayland.

Anyway, we have three different kinds of X11 apps:
- no sync
- basic sync
- extended sync

For the no-sync kind I think we still need something like my patch
series. The other sync types would build on top, either by extending or
replacing _XWAYLAND_ALLOW_COMMITS - or even just leaving it to XWM as a
master switch of commits.

Hence, at the moment, I believe this series would be a useful step
forward, if the idea of the server inspecting properties is acceptable.

Mind, I am mostly thinking this in Weston XWM terms, which draws the
window decorations through X11 like a normal window manager. I'm not
sure how things would change if the Wayland compositor drew the
decorations directly, internally, as a part of composition, but I would
hope it just means some steps in the sync protocols simply become
no-ops.

CC'd Owen in case he has opinions.

Thanks,
pq

> > [3]
> > https://bugzilla.gnome.org/show_bug.cgi?id=767212 [4]
> > https://bugs.freedesktop.org/show_bug.cgi?id=81275  
> 

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 801 bytes
Desc: OpenPGP digital signature
URL: <https://lists.x.org/archives/xorg-devel/attachments/20161128/ded7cc27/attachment.sig>