Windows can be freed at some point after they are unmanaged - because
there is an effect in progress, because a language binding is holding
a reference. Therefore, we need to clean up the later to associate
the xwayland and wayland windows deterministically in an "unamanaged"
handler.
https://bugzilla.gnome.org/show_bug.cgi?id=736694
g_idle_add() makes no guarantee about when it will be run - if Mutter
is busy drawing and blocking glXSwapBuffers() it could happen only
minutes later. Use meta_later_add (META_LATER_BEFORE_REDRAW) instead -
this will deterministically be run after the Wayland socket is read
from but before the next frame is painted.
https://bugzilla.gnome.org/show_bug.cgi?id=736694
Putting X windows and pointers to MetaWindows into a union had a number of
problems:
- It caused awkward initialization and conditionalization
- There was no way to refer to Wayland windows (represented by
MetaWindow *) in the past, which is necessary for the MetaStackTracker
algorithms
- We never even cleaned up old MetaStackWindow so there could be
records in MetaStackWindow pointing to freed MetaWindow.
Replace MetaStackWindow with a 64-bit "stack ID" which is:
- The XID for X Windows
- a "window stamp" for Wayland windows - window stamps are assigned
for all MetaWindow and are unique across the life of the process.
https://bugzilla.gnome.org/show_bug.cgi?id=736559
Add private functions for the test framework to use to find out the
wayland and x11 display names, so they can set up the environment for
children.
https://bugzilla.gnome.org/show_bug.cgi?id=736505
It's possible for a released pointer to have repick / set_focus on it as
part of sync_input_focus. When the pointer is actually re-init'd, it
will memset 0, which can cause corruption as our destroy listener has
already been added.
Released devices should be idempotent, so just make sure method calls on
them don't have any effect.
A wl_surface may have a wl_subsurface interface, but no buffers attached
yet, even though the geometry calculation code for surfaces/subsurfaces
assumes everything has already a buffer.
Just skip subsurfaces that don't have a buffer, those can't be set
a geometry yet, and right now it's crashing accessing the texture from
the NULL surface->buffer.
https://bugzilla.gnome.org/show_bug.cgi?id=735452
This makes it so that MetaSurfaceActorWayland is effectively just a
wrapper actor around MetaShapedTexture with some extra scaling. I think
the MetaSurfaceActor subclassing was a bad idea -- we really should have
these abstractions in much higher levels in the stack than the
compositor.
It doesn't make sense to update it in the surface actor. It's also
theoretically wrong to update the buffer's texture on surface commit,
too, because it's buffer state, not surface state, but I don't think
there's any place we use a wl_buffer without a wl_surface.
The initialization sequence before was quite icky, initializing Clutter
in a few different places depending on what was going on.
Put that all back into main.c
MetaGrabOp is painful and tedious to work with, because it's a
sequential series of values, meaning we have to use a giant unreadable
switch statement to figure out some basic things about the value.
To solve this, modify the encoding for MetaGrabOp and for the specific
window grab operations so that they're a set of bitflags that we can
easily check.
We've long used a switch statement on the grab operation to determine
where events should go. The issue with MetaGrabOp is that it's a mixture
of a few different things, including event routing, state management,
and the behavior to choose during operations.
This leads to poorly defined event routing and hard-to-follow logic,
since it's sometimes unclear what should point where, and our utility
methods for determining grab operations apart can be poorly named.
To fix this, establish the concept of a "event route", which describes
where events should be routed to.
Refuse to create a touch resource if we don't have the capability
(for misbehaving clients), and don't attempt to use touch data
structures that are not initialized.
This is a terrible hack. We need to figure out a better way to do
interactive resizes.
This fixes weird resizing from the left bugs when using GTK+, which is
really slow at acking configures.
This is an easy way to get into an infinite loop where we're constantly
re-sending stuff to the window. If it worked once, it probably won't
work again.
We assume in meta_window_wayland_move_resize that the next commit that
changes the geometry will always be for our next pending operation, so
if we have a move pending on a resize, the next commit will trigger the
move. This is, of course, fundamentally wrong.
We broke this assumption even more now that we don't fizzle out calls to
meta_window_move_resize_internal and now call it on every commit, which
means that a simple damage and then commit would complete a pending
move.
This was even broken by apps like weston-terminal, which, when clicking
on the maximize button, first redraws the terminal with the maximize
button state back on hover on press, and would only redraw when it got
the configure event with the coordinates.
To track the correct commit to apply the move for, we implement the
ack_configure request and ignore all move/resizes that happen before
that.
Right now, we actually fizzle out the entire move/resize if there's a
future pending configure we're waiting on.
The grabbing state is now checked for both pointer/touch devices
within the seat, and the grab start coordinates returned by
meta_wayland_seat_get_grab_info().
https://bugzilla.gnome.org/show_bug.cgi?id=733631
The capability flags are determined from the device types of the slave devices
that are currently attached. This also happens whenever a device is added or
removed, so the capabilities are kept up to date, and clients know about these.
On VT switch, all slave devices are temporarily removed, so the cascade of
signals will make the seat end up with capabililities=0 while input is suspended.
https://bugzilla.gnome.org/show_bug.cgi?id=733563
Anytime the keymap is changed, either directly, or indirectly through the
keyboard capability being released/initialized, there should be a
notification of the modifiers being changed too.
https://bugzilla.gnome.org/show_bug.cgi?id=733563
Otherwise the focus_surface_listener list element becomes stale, and then
mangled if the devices' data is initialized again, and the memory memset().
https://bugzilla.gnome.org/show_bug.cgi?id=733563
This doesn't match what Weston does. I don't know of any apps that this
fixes (we don't have any apps that even use non-zero dx/dy, I don't
think), but this is part of a cleanup for window geometry.
When frame extents change, we might not update the frame rect, but the
buffer rect still needs to be updated. Split out the check for this to
be independent of the check for the frame rect.
This fixes issues that could happen when the window was maximized while
it was in the top-left corner.
The output_id is more of an opaque identifier for the monitor, based on
its underlying ID from the windowing system. Since we also use the term
"output_id" for the output's index, rename our use of the opaque cookie
"output_id" to "winsys_id".
When we changed the setting of the buffer rect to be inside the moving
code to make sure it was updated in places we were moving directly
without any round-trip needed, I removed a code to set the buffer rect
without remembering that's where the size of it was updated.
Add back the code to update the buffer rect.
This fixes Wayland windows not appearing.
It returns FALSE when button_count is not 0. But grabbing for
move/resize is activated by clicking the button, so this condition
disallows the wayland clients to be moved/resized.
https://bugzilla.gnome.org/show_bug.cgi?id=731237
Rather than calculate it speculatively with the current properties
which may be too new or too out of date, make sure it always fits
with the proper definition. We update it when we update the toplevel
window for X11, and when a Wayland surface is committed with a newly
attached buffer.
There is no way this value will ever be read, because we set the
cursor_surface to NULL, this is set at the same time as cursor_surface,
and it's only read if cursor_surface is non-NULL.
Clutter touch events are translated into events being sent down
the interface resource, with the exception of FRAME/CANCEL events,
which are handled directly via an evdev event filter.
The seat now announces invariably the WL_SEAT_CAPABILITY_TOUCH
capability, this should be eventually updated as devices come and
go.
The creation of MetaWaylandTouchSurface structs is dynamic, attached
to the lifetime of first/last touch on the client surface, and only
if the surface requests the wl_touch interface. MetaWaylandTouchInfo
structs are created to track individual touches, and are locked to
a single MetaWaylandTouchSurface (the implicit grab surface) determined
on CLUTTER_TOUCH_BEGIN.
https://bugzilla.gnome.org/show_bug.cgi?id=724442
Smooth scroll event vectors from clutter have the same dimensions as the
ones from from Xi2, i.e. where 1.0 is 1 discrete scroll step. To scale
these to the coordinate space used by wl_pointer.axis
vertical/horizontal scroll events, multiply the vector by 10.
https://bugzilla.gnome.org/show_bug.cgi?id=729601
The last commit added support for the "appmenu" button in decorations,
but didn't actually implement it. Add a new MetaWindowMenuType parameter
to the show_window_menu () functions and use it to ask the compositor
to display the app menu when the new button is activated.
https://bugzilla.gnome.org/show_bug.cgi?id=730752
The requested_rect is a strange name for it, because it's not actually
the rect that the user or client requested all the time: in the case of
a simple move or a simple resize, we calculate some of the fields
ourselves.
To the MetaWindow subclass implementations, it just means "the rect
before we constrained it", so just use the name unconstrained_rect.
This also makes it match the name of the MetaWindow field.
Realistically, the user rect contains the unconstrained window
rectangle coordinates that we want to be displaying, in case
something in the constraints change.
Rename it to the "unconstrained_rect", and change the code to always
save it, regardless of current state.
When metacity was originally being built, the purpose of the user
rect was a lot less clear. The code only saved it on user actions,
with various other calls to save_user_window_placement() and a force
mechanism sprinkled in to avoid windows being snapped back to odd
places when constraints changed.
This could lead to odd bugs. For instance, if the user uses some
extension which automatically tiles windows and didn't pass
user_action=TRUE, and then the struts changed, the window would be
placed back at the last place a user moved it to, rather than where
the window was tiled to.
The META_IS_USER_ACTION flag is still used in the constraints code
to determine whether we should allow shoving windows offscreen, so
we can't remove it completely, but we should think about splitting
out the constrainment policies it commands for a bit more
fine-grained control.
https://bugzilla.gnome.org/show_bug.cgi?id=726714
The default focus interface uses the button count to determine
whether we should update the pointer focused surface. When releasing
an implicit grab, we need to send the button release events to the
implicitly grabbed surface, so we can't reset the focus surface too
soon. We already explicitly set the focus at the end of implicit
grabs, so counting the buttons after is perfectly fine.
If we send out a configure notify for a window and then have some
other kind of state change, we need to make sure that we continue
to send out that new size, rather than the last size the client
sent us a buffer for.
In particular, a client might give us a 250x250 buffer and then
immediately request fullscreen. We send out a configure for the
monitor size and a state that tells it it's full-screen, but then
it takes focus, and since the client hasn't sent us a buffer for
the new size, we tell it it's fullscreen at 250x250.
Fix this.
If we attach to a MetaWindow that disappears before the idle fires,
we'll notice that we can't associate the window properly again and
try to access data on the MetaWindow struct, which might crash.
Install a weak ref that ties the lifetime of the idle to the lifetime
of the MetaWindow.
It seems every GTK+ app does this for some reason at startup. This
is really unfortunate, since we'll have to create and destroy a new
MetaWindow really quickly.
Scale surfaces based on output scale and the buffer scale set by them.
We pick the scale factor of the monitor there are mostly on.
We only handle native i.e non xwayland / legacy clients yet.
https://bugzilla.gnome.org/show_bug.cgi?id=728902
Advertise the scale factor on the output and transform pointer and damage
events as well as input and opaque regions for clients
that scale up by themselves i.e use set_buffer_scale.
We do not scale any 'legacy' apps yet.
https://bugzilla.gnome.org/show_bug.cgi?id=728902
Ugh. So in the fullscreen case, we need to make sure to specify that
it's a MOVE_ACTION so that we move to the saved position, but we
can't do that in the resizing case since we need to use the resized
rectangle.
The flags are really hurting us here. Perhaps we should make it the
client's responsibility to specify a complete rectangle which we
could resize to; then the weird-o logic would be self-contained in
each front-end.
I'm not convinced this covers all cases, especially when we could have
a dangling weird state pointer, but it fixes our existing two testcases.
Restoring the position in our move_resize_internal implementation
is too late. We need to do it at ack-time, before we hand off the
new position to the constraints code.
For the server-initiated resize case, like unmaximize or some forms
of tiling, we dropped the x/y of the server-assigned rectangle on the
floor, which meant the surface didn't move to where it needed to be in
that case. Now, save it internally, and combine it with the dx/dy passed
in during attaches to figure out where we actually need to be.
Make sure to only use it for when we send out a configure notify. We
should use the passed in rectangle for other scenarios, like a
client-initiated resize.
This fixes incorrect surface placement after unmaximization.
For the server-initiated resize case, like unmaximize or some forms
of tiling, we dropped the x/y of the server-assigned rectangle on the
floor, which meant the surface didn't move to where it needed to be in
that case. Now, save it internally, and combine it with the dx/dy passed
in during attaches to figure out where we actually need to be.
This fixes incorrect surface placement after unmaximization.
Looking at the code paths where is_mouse / is_keyboard are used,
all of them should never be run when dealing with a COMPOSITOR
grab op, since they're filtered out above or the method is just
never run during that time.
It's confusing that COMPOSITOR is in here, and requires us to
be funny with other places in code, so just take it out.
pointer->current needs to always be the surface under the pointer,
even when we have a grab. We do need to make sure we keep the focus
surface the same even when we have a grab, though, so add logic
for that.
In order to correctly fix the issue to make sure we only set the
focused surface to NULL during a grab, but not the current surface,
we need to merge update_current_surface back into repick_for_event
so we have more control over the behavior here.
... not when we do an update.
We only repick when we handle events, not when we update. Perhaps
this is a mistake.
Since update runs before handle_event, this means that when we
drop a grab, update will notice the NULL surface, since we haven't
repicked after the event, and then we'll repick the correct surface.
The end result is that you see a root cursor after a grab ends,
rather than the correct window cursor.
This doesn't fix it, since the current surface becomes NULL when
we start the grab. But it does make the code here more correct when
we fix that bug.
I was talking with other people and they became confused at the
term "double-buffered", since we were also talking about
double-buffering in general, e.g. swapping between two buffers.
Instead, we'll adapt the "pending state" nomenclature that we
already use for the field / variable names.
If we have a focused surface, we need to eat up key events, not
just if we have a non-empty focus resource list. The latter would
happen if we have a focused client but it never called get_keyboard.
The latest Xorg / Xwayland has support for -displayfd being used
in conjunction with an explicit display number. Use that to know
when the X server is ready, rather than UNIX signals, because
they're UNIX signals.
If we're sending out a configure event, we can't immediately move the
window; we need to instead wait to apply the new position when the
client sends a new buffer.
dx/dy should be against the regular window's rect, and need to
be ignored when we're resizing. Instead, we use gravity to anchor
the window's new rectangle when resizing.
Sophisticated clients, like those using ClutterGtk, will have more
than one focused resource per client, as both Clutter and GDK will
ask for a wl_pointer / wl_keyboard. Support this naturally using
the same "hack" as Weston: multiple resource lists, where we move
elements from one to the other.
In order to support multiple pointers for the same client, we're
going to need to kill it.
This will cause crashes for now, but will be fixed by the next
commit.
default_grab_focus tries to add implicit grab semantics where
focus won't take effect if there's a pointer button down. This
is not what we want for popup grabs at all, as it's perfectly
valid to want to drag on a menu while there's a button down.
The idea here is that while we take a WM-side grab, like a compositor
grab or a resizing grab, we need to remove the focus from the Wayland
client.
We make a special exception for CLICKING operations, because these
are really an internal state machine while you're pressing on a button
inside a frame, and in this case, we need to not kill the focus.
Really, it is a special case. When the subsurface is synchronous,
commit changes meaning from being applied immediately to being
queued up for replay later. Handle this explicit special case
with an explicit special case in the code.
This means that in all other paths, we can unconditionally
apply the actor immediately.
Even when it doesn't have a role.
This fixes cursors not quite working right, as they're a "detached"
surface without a role since nobody called set_cursor on them yet.
Instead of using commit_attached_buffer / actor_surface_commit.
We want to kill the return values of these methods because we
really should always be calling them, even if the surface doesn't
have a role.
This is also something that we did upstream. Since we want to
introduce an explicit "xdg_transient" window type for tooltips
and popovers, and since "transient_for" is a confusing dumb
80s term lifted from the ICCCM spec, just rename it.
This was changed upstream a little while ago for C++ compatibility.
It's also the more common term for the operation: you close a window,
you don't delete one. In fact, a delete event might seem like it
would be about resource management instead.
Since we get the ClientMessage after the surface is created, there's
no good way to synchronize the two streams. In this case, what we
need to do is delay the surface commit until after we get the
ClientMessage. Ideally, we'd be using a better surface system overall
where committing the surface didn't depend on what type it is, but
oh well, this is a good short-term hack for now.
This is effectively the same, but since we lose the xserver.xml protocol
in the new XWayland DDX, we have to use SIGUSR1 anyway, so might as well
switch over now.
The make_toplevel / window_unmanaging interface has never made
a lot of sense to me. Replace it with set_window, which does
effectively the same thing.
It's still not perfect in the case of XWayland, but I don't think
XWayland will ever make me happy.
It should be META_TYPE_WAYLAND_STAGE, not META_WAYLAND_TYPE_STAGE.
Well, actually, it *should* be META_TYPE_NATIVE_STAGE, because it's
not related to Wayland at all. But that comes later :)
I'm a bit tired of hearing about this when I launch mutter-wayland
nested. Ideally, this would be part of display server integration,
not GNOME integration, so we could simply not make the call when
nested, but oh well.
Previously, a sequence like this would crash a client:
=> surface.attach(buffer)
=> buffer.destroy()
The correct behavior is to wait until we release the buffer before
destroying it.
=> surface.attach(buffer)
=> surface.attach(buffer2)
<= buffer.release()
=> buffer.destroy()
The protocol upstream says that "the surface contents are undefined"
in a case like this. Personally, I think that this is broken behavior
and no client should ever do it, so I explicitly killed any client
that tried to do this.
But unfortunately, as we're all well aware, XWayland does this.
Rather than wait for XWayland to be fixed, let's just allow this.
Technically, since we always copy SHM buffers into GL textures, we
could release the buffer as soon as the Cogl texture is made.
Since we do this copy, the semantics we apply are that the texture is
"frozen" in time until another newer buffer is attached. For simple
clients that simply abort on exit and don't wait for the buffer event
anyhow, this has the added bonus that we'll get nice destroy animations.
If we have a CLICKING grab op we still need to send events to xwayland
so that we get them back for gtk+ to process thus we can't steer
wayland input focus away from it.
https://bugzilla.gnome.org/show_bug.cgi?id=726123
meta_wayland_seat_repick() can be called in various cases while mutter
has a GRAB_OP ongoing which means we could be sending wrong pointer
enter/leave events.
https://bugzilla.gnome.org/show_bug.cgi?id=726123
In particular we need to know about all key events to keep the xkb
state reliable even if the event is then consumed by a global shortcut
or grab and never reaches any wayland client.
We also need to keep track of all pressed keys at all times so that we
can send an updated set or pressed keys to the focused client when a
grab ends.
https://bugzilla.gnome.org/show_bug.cgi?id=722847
Any given clutter event carries the modifier state as it was before it
occured but, for the wayland modifiers event, we want the state
including the current event.
To fix this, we'll keep our xkb_state instance around instead of the
serialized mods.
https://bugzilla.gnome.org/show_bug.cgi?id=722847
Which is used for Wayland popup grabs.
The issue here is that we don't want the code that raises or focuses
windows based on mouse ops to run while a client has a grab.
We still keep the "old" grab infrastructure in place for now, but
ideally we'd replace it eventually with a better grab-op infrastructure.
Do to a bad mixup, the surface listener was never actually fired.
This was accidentally fixed as part of a refactoring in a27fb19,
but the surface listener was broken, and we started crashing. To
fix, just remove the surface listener, as we've mostly been testing
without it.
This is not needed since the instance is being destroyed and in fact
actively harmful when code called from other handlers disconnects us
for other reasons. In that case we might crash because the
disconnection doesn't prevent other handlers from running in the
current signal emission and thus we try to remove ourselves from an
empty list.
This changes the user data of all surface extensions resources to be
the MetaWaylandSurface instead of the MetaWaylandSurfaceExtension,
which means that we no longer need all these pesky wl_container_ofs
in implementations.
Don't set the surface actor to a new buffer if it's becoming unmapped.
This is also technically wrong since we'll send out the release event,
but oh well.
We should probably decouple MetaWaylandBuffer from the CoglTexture
at some point, so we can send out releases on-demand.
We need a MetaWaylandSurface to build a MetaSurfaceActor, but
we don't have one until we get the set_window_xid() call from
XWayland. On the other hand, plugins expect to see the window
actor right from when the window is created, so we need this
empty state.
Based on a patch by Jasper St. Pierre.
This time, to make way for MetaSurfaceActorEmpty. This also fixes
destroy effects as a side effect. It still has issues if we try
to re-assign an actor that's already toplevel (e.g. somebody
re-popping up a menu that's already being destroyed), but this
will be fixed soon.
The idea here is that MetaWindowActor will do the unparenting of
the surface actor when it itself is destroyed. To prevent bad issues
with picking, we only make the surface actor reactive when it's
toplevel.
The rendering logic before was somewhat complex. We had three independent
cases to take into account when doing rendering:
* X11 compositor. In this case, we're a traditional X11 compositor,
not a Wayland compositor. We use XCompositeNameWindowPixmap to get
the backing pixmap for the window, and deal with the COMPOSITE
extension messiness.
In this case, meta_is_wayland_compositor() is FALSE.
* Wayland clients. In this case, we're a Wayland compositor managing
Wayland surfaces. The rendering for this is fairly straightforward,
as Cogl handles most of the complexity with EGL and SHM buffers...
Wayland clients give us the input and opaque regions through
wl_surface.
In this case, meta_is_wayland_compositor() is TRUE and
priv->window->client_type == META_WINDOW_CLIENT_TYPE_WAYLAND.
* XWayland clients. In this case, we're a Wayland compositor, like
above, and XWayland hands us Wayland surfaces. XWayland handles
the COMPOSITE extension messiness for us, and hands us a buffer
like any other Wayland client. We have to fetch the input and
opaque regions from the X11 window ourselves.
In this case, meta_is_wayland_compositor() is TRUE and
priv->window->client_type == META_WINDOW_CLIENT_TYPE_X11.
We now split the rendering logic into two subclasses, which are:
* MetaSurfaceActorX11, which handles the X11 compositor case, in that
it uses XCompositeNameWindowPixmap to get the backing pixmap, and
deal with all the COMPOSITE extension messiness.
* MetaSurfaceActorWayland, which handles the Wayland compositor case
for both native Wayland clients and XWayland clients. XWayland handles
COMPOSITE for us, and handles pushing a surface over through the
xf86-video-wayland DDX.
Frame sync is still in MetaWindowActor, as it needs to work for both the
X11 compositor and XWayland client cases. When Wayland's video display
protocol lands, this will need to be significantly overhauled, as it would
have to work for any wl_surface, including subsurfaces, so we would need
surface-level discretion.
https://bugzilla.gnome.org/show_bug.cgi?id=720631
If the client destroys the pointer resource, we shouldn't unfocus the
surface, and we should regrab it when the client gets the pointer
resource again.
This also fixes a crash at surface destruction because of the unchecked
wl_link_remove that will happen on both pointer and surface destroy.
