CSD X11 clients and Wayland clients don't have a window frame drawn by
the compositor to flash. So instead of flashing the whole screen when
configured to just flash the window, flash just the window region.
https://bugzilla.gnome.org/show_bug.cgi?id=763284
Some windows, like Chromium and Steam, are technically CSD in that they
don't want a system titlebar and draw their own, but we should still
provide them with a shadow.
If a queued event is being processed after the surface is
destroyed, trying to access the window associated with the surface
will lead to a segmentation fault.
This patch avoids the crash by first checking if the surface is not null.
https://bugzilla.gnome.org/show_bug.cgi?id=754715
Some backgrounds don't fully fill the screen. For those backgrounds
it's important to paint a color behind them to fill in the gaps.
This commit checks whether or not the background image textures take
up the entire monitor, and in the event they don't, draws a color
behind them (such as it would do if the background were
translucent).
https://bugzilla.gnome.org/show_bug.cgi?id=754476
meta_background_get_texture only draws the bottom image texture
if
1) the blend factor leaves the top image translucent
or
2) the top image is translucent from alpha
The latter case doesn't actually matter since we're using REPLACE
on the top image texture.
This commit drops the unnecessary check for the second case and
applies demorgans law to the conditional for clarity.
https://bugzilla.gnome.org/show_bug.cgi?id=754476
Since mutter has two X connections and does damage handling on the
frontend while fence triggering is done on the backend, we have a race
between XDamageSubtract() and XSyncFenceTrigger() causing missed
redraws in the GL_EXT_X11_sync_object path.
If the fence trigger gets processed first by the server, any client
drawing that happens between that and the damage subtract being
processed and is completely contained in the last damage event box
that mutter got, won't be included in the current frame nor will it
cause a new damage event.
A simple fix for this would be XSync()ing on the frontend connection
after doing all the damage subtracts but that would add a round trip
on every frame again which defeats the asynchronous design of X
fences.
Instead, if we move fence handling to the frontend we automatically
get the right ordering between damage subtracts and fence triggers.
https://bugzilla.gnome.org/show_bug.cgi?id=728464
The spec says:
"A server should avoid signalling the frame callbacks if the surface is not
visible in any way, e.g. the surface is off-screen, or completely obscured
by other opaque surfaces."
We actually do have the information to do that but we are always calling
the frame callbacks in after_stage_paint. So fix that to only call when
when the surface gets drawn on screen.
https://bugzilla.gnome.org/show_bug.cgi?id=739163
The compositor maintains a ring of shared fences with the X server in order to
properly synchronize rendering between the X server and the compositor's GPU
channel. When all of the fences have been used, the compositor needs to reset
one so that it can be reused. It does this by first waiting on the CPU for the
fence to become triggered, and then sending a request to the X server to reset
the fence.
If the compositor's GPU channel is busy processing other work (e.g. the desktop
switcher animation), then the X server may process the reset request before the
GPU has consumed the fence. This causes the GPU channel to hang.
Fix the problem by having the compositor's GPU channel trigger its own fence
after waiting for the X server's fence. Wait for that fence on the CPU before
sending the reset request to the X server. This ensures that the GPU has
consumed the X11 fence before the server resets it.
Signed-off-by: Aaron Plattner <aplattner@nvidia.com>
https://bugzilla.gnome.org/show_bug.cgi?id=728464
If GL advertises this extension we'll use it to synchronize X with GL
rendering instead of relying on the XSync() behavior with open source
drivers.
Some driver bugs were uncovered while working on this so if we have
had to reboot the ring a few times, something is probably wrong and
we're likely to just make things worse by continuing to try. Let's
err on the side of caution, disable ourselves and fallback to the
XSync() path in the compositor.
https://bugzilla.gnome.org/show_bug.cgi?id=728464
When a client sets an input region or a opaque region to NULL, it
should still be considered a change to the corresponding region on the
actor. This patch makes sure this state is properly forwarded.
https://bugzilla.gnome.org/show_bug.cgi?id=753222
Keep the active position state in its original coordinate space, and
synchronize the surface actor with it when it changes and when
synchronizing the rest of the surface state, in case the surface scale
had changed.
https://bugzilla.gnome.org/show_bug.cgi?id=745655
Whenever a MetaSurfaceActor is painted, update the list of what outputs
the surface is being drawed upon. Since we do this on paint, we
effectively avoids this whenever the surface is not drawn, for example
being minimized, on a non-active workspace, or simply outside of the
damage region of a frame.
DND icons and cursors are not affected by this patch, since they are not
drawn as MetaSurfaceActors. If a MetaSurfaceActor or a parent is cloned,
then we'll check the position of the original actor again when the clone is
drawn, which is slightly expensive, but harmless. If the MetaShapedTexture
instead is cloned, as GNOME Shell does in many cases, then these clones
will not cause duplicate position checks.
https://bugzilla.gnome.org/show_bug.cgi?id=744453
We may access it during painting even if it has been freed. For now,
manually unset it during the MetaWaylandSurface cleanup; in the future
make MetaWaylandSurface a GObject and make the surface pointer a weak
reference.
https://bugzilla.gnome.org/show_bug.cgi?id=744453
The elementary guys would like this as an API, and I don't see any
reason to refuse -- this is quite nice shadow painting code :)
For some reason, gobject-introspection can't seem to cope with
MetaWindowShape. I'll look into it a bit later, but for now, mark
the function it has trouble with as (skip).
It seems that when translated, paint_offset and actor_offset will always
be the same, so our translation of the clip group won't work. For now,
until I figure out what's going on here, just use the painting offset,
since that what seems to make sense to me.
I didn't write this code, though, so I don't know why the actor's
allocation was involved in this computation at all.
I tested briefly with clones (magnifier, manual cloning through the
looking glass) and couldn't find any other artifacts, so I'm going to do
this for now.
A much less hacky version of maximize / unmaximize is reimplemented
in terms of this, but it could also eventually be used for fullscreen /
unfullscreen, and tile / untile.
The comment explains it better, but Clutter tries to be smart and
repaint actors when their allocations change. Since the window group's
allocation changes when windows move around, this means that moving a
window will always cause a full-stage repaint, which is super slow.
Hack around this for now.
While nothing will completely fix X11's artifacts, this tends to look a
bit better, *especially* with mask textures that have black at the
edges (which are most of them).
It's also faster for GPUs to manage.
When we're unredirected, we don't have a pixmap, and thus our allocation
becomes 0x0. So when events come in, they pass right through our actor,
going to the one underneath in the stack.
Fix this by having a fallback size on the shaped texture actor when
we're unredirected, causing it to always have a valid allocation.
This fixes clicking on stuff in sloppy / mouse mode focus.
Implicit conversion from int to float is not supported by
GLSL ES.
Fixes:
(gnome-shell:8954): Cogl-WARNING **: Shader compilation failed:
1:2: P0004: High precision not supported, instead compiling high precision as medium precision
4:17: S0001: Type mismatch in arithmetic operation between 'int' and 'float'
when one trigger the overview mode on Mali 400 r1p1 GPU.
https://bugzilla.gnome.org/show_bug.cgi?id=745442
In order to switch to the correct surface actor scale given the monitor
the surface is on, without relying on the client committing a new state
given some other side effect, sync the surface actor state when the main
monitor associated with the corresponding window changed.
https://bugzilla.gnome.org/show_bug.cgi?id=744933
Since the surface actor knows more about how it draws itself, instead of
pushing texture state (buffer and scale), input region and opaque region
from MetaWaylandSurface after having transformed into what the surface
actor expects, make the surface actor set its own state given what state
the Wayland surface is in.
https://bugzilla.gnome.org/show_bug.cgi?id=744933
The current ordering updates the clip shape of the composite overlay
window after unredirecting the target window. This has the effect of
forcing X to clear the target window and sending an expose to the
application to repaint - causing an unsightly flash. If we update the
shape first, then unredirect, X restores the background of the root
window (sending no expose events as no one is interested) and the
background is typically NONE for the root window. Then the unredirect
paints the contents of the composite backing pixmap over top without
requiring a round trip and waiting for the client to repaint - thus no
flashing.
Fixes regression from
commit d6282716b2
Author: Jasper St. Pierre <jstpierre@mecheye.net>
Date: Fri Dec 6 17:10:44 2013 -0500
compositor: Simplify the unredirected window management code
Cc: Jasper St. Pierre <jstpierre@mecheye.net>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
https://bugzilla.gnome.org/show_bug.cgi?id=743858
This just exposes the type and the singleton getter necessary to make
it available to introspection. We'll expose more functionality as it
becomes needed.
https://bugzilla.gnome.org/show_bug.cgi?id=743745
If the app finished multiple frames before we sent _NET_WM_FRAME_DRAWN,
we could add the send_frame_messages_timer multiple times. In the rare
case that the app immediately closed the window, the older timeout
could potentially then run on the freed actor.
https://bugzilla.gnome.org/show_bug.cgi?id=738686
* Use -1 rather than 0 as a flag for pending queue entries; 0 is
a valid frame_counter value from Cogl.
* Consistently handle the fact we can have more than one pending
entry. It's app misbehavior to submit a new frame before
_NET_WM_FRAME_DRAWN is received; but we accept such frame messages,
so we can't just leak them.
* If we remove send_frame_message_timer, assign the current frame counter
to pending entries.
* To try to avoid regressing on this, when sending _NET_WM_FRAME_TIMINGS
messages, if we have stale messages, or messages with no frame drawn
time, warn and remove them from the queue rather than just accumulating.
* Improve commenting.
https://bugzilla.gnome.org/show_bug.cgi?id=738686
The parent pick() implementation in ClutterActor only recurses if the
vfunc is untouched, which means it's up to the MetaWaylandSurface
implementation to actually recurse, just the same as if an input mask
applied.
https://bugzilla.gnome.org/show_bug.cgi?id=738890
WindowActors can outlive their corresponding window to animate unmap.
Unredirecting the actor does not make sense in that case, so make
sure to not request it.
https://bugzilla.gnome.org/show_bug.cgi?id=740133
If the actor surface has an input mask, custom picking is implemented
for the portions affected by the mask, although the child actors (most
usually subsurfaces) are left out.
https://bugzilla.gnome.org/show_bug.cgi?id=738890
Damage coordinates are relative to the drawable not to the screen. So we
have to check whether x and y are 0 and not window_rect.x/y otherwise the
herustic will never trigger for windows on monitors whos x and y are not 0.
https://bugzilla.gnome.org/show_bug.cgi?id=738271
This actor is a subclass of MetaFeedbackActor that additionally
implements the "drag failed" animation, snapping back to the drag
origin position in a surface.
This actor is a non-reactive container that autoembeds itself into
the feedback window group in the compositor. The API is meant to
help on creating things attached to pointer/touchpoints, with an
X/Y attachment offset, and following the position of certain events.
Although not strictly a window group... This ClutterActor is
meant to stay always on top, and only show non-reactive actors
created by Mutter itself. Two possible usecases for this layer
are DnD surfaces, and touch spots.
We might also want to move cursors out of an overlay in MetaStage
into here at some point.
According to the documentation, the method returns "whether the X window
that the actor was displaying has been destroyed" - that is very much
true when we delay the actual actor destruction for a destroy animation,
so update the method accordingly.
https://bugzilla.gnome.org/show_bug.cgi?id=735927
When a window is destroyed, the corresponding actor may still be
kept around for the destroy effect. But as the actor is removed
from the compositor's stack list immediately, the compositor will
always stack it above "valid" window actors - this is not what we
want, so only update the compositor's list when the actor is
actually destroyed.
https://bugzilla.gnome.org/show_bug.cgi?id=735927
Not having a paint volume causes every single paint to turn into
full-stage redraw, since otherwise culling won't properly work.
Since we don't paint outside of our allocation, just use the simple
default implementation, but also return TRUE inside it.
Make the vignette options properties so they can be animated;
modify the function-call API for meta_background_actor_set_vignette()
to correspond more closely to the 3 properties.
https://bugzilla.gnome.org/show_bug.cgi?id=735637
Without GLSL, we didn't apply the vignetting, which not only made the
background uniform in color, it made it much lighter. Adjust for this
and make the average brightness with the vignette effect the same
with or without GLSL.
https://bugzilla.gnome.org/show_bug.cgi?id=735637
The old requirement that multiple MetaBackgroundActor objects be
layered on top of each to produce blended backgrounds resulted in
extremely inefficient drawing since the entire framebuffer had
to be read and written multiple times.
* Replace the MetaBackground ClutterContent with a plain GObject
that serves to hold the background parameters and prerender
textures to be used to draw the background. It handles
colors, gradients, and blended images, but does not handle
vignetting
* Add vignetting to MetaBackgroundActor directly.
* Add MetaBackgroundImage and MetaBackgroundImageCache to allow
multiple MetaBackground objects to share the same images
By removing the usage of ClutterContent, the following optimizations
were easy to add:
Blending is turned off when the actor is fully opaque
Nearest-neighbour filtering is used when drawing 1:1
The GLSL vignette code is slightly improved to use a vertex shader
snippet for computing the texture coordinate => position in actor
mapping.
https://bugzilla.gnome.org/show_bug.cgi?id=735637
The old check for using NEAREST by checking clutter_actor_is_in_clone_paint()
and meta_actor_is_untransformed (actor) doesn't work properly since
clutter_actor_is_in_clone_paint() does not look at ancestors of the
actor; it only applies to a direct clone of the actor. Using
meta_actor_painting_untransformed() allows us to check exactly what we
care about rather than using tricky approximations.
https://bugzilla.gnome.org/show_bug.cgi?id=735632
The painting_untransformed() function in MetaWindowGroup is useful
elsewhere, in particular if we want to check whether we can avoid
bilinear filtering when painting a texture 1:1.
https://bugzilla.gnome.org/show_bug.cgi?id=735632
meta_surface_actor_is_argb32 assumes that lack of stex means that a window is
ARGB32. When we unredirect a window we detach the texture so we end up without
a texture. Given that should_unredirect returns FALSE when a window is argb32,
we know that this window is indeed not ARGB32.
Returing TRUE in that case causes us to flip between redirected and
unredirected on every paint.
So fix that by returning FALSE in that case.
When the blended region was empty, meaning we didn't have to paint
anything blended -- the case for an app update -- was drawing the
entire window blended, because of a think-o in the complex and
complicated logic.
Fix this so that we don't draw anything for the blended region when
empty.
region first
If we're going to render the entire texture blended, then don't bother
painting the unblended stuff, since we're just going to draw on top
anyway.
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.
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.
This allows creating the stage much earlier than it otherwise would have
been. Our initialization sequence has always been a bit haphazard, with
first the MetaBackend created, then the MetaDisplay, and inside of that,
the MetaScreen and MetaCompositor.
Refactor this out so that the MetaBackend creates the Clutter
stage. Besides the clarity of early initialization, we now have much
easier access to the stage, allowing us to use it for things such as
key focus and beyond.
If we for some reason have an error trying to allocate the framebuffer,
we'll still mark the tower as revalidated. Move the validation to the
end of the actual revalidation code to solve this.
It's a deprecated API that can surprise us. Namely, when the internal
format passed is COGL_PIXEL_FORMAT_ANY, it will *always* allocate an
RGBA8888 pixel format texture, even if we only passed it a RGB format
or even an A8 format.
cogl_texture_2d_new_with_data is the newer, better API and doesn't have
these warts.
Connecting to size-changed is wrong -- size-changed tells us when
we *told* the X server or resize the window. For X11, we're sort of
guaranteed that the surface will be updated at some point before the
next frame, but for Xwayland, we can't be sure that the new surface is
attached at this point.
This fixes weird artifacts when resizing apps like xclock.
This was wrong for subsurfaces that extend beyond the parent's shape,
since the paint volume would be wrong in this case. Instead of using the
shape region which can be out of date and wrong, just use the union of
our children's volumes, which is a lot easier to manage.
Use connect_after() to accomodate code in GNOME Shell that,
when benchmarking drawing performance, connects to ::after-paint
and calls glFinish(). The timing information from that will be
more accurate if we hold off until that completes before we signal
apps to begin drawing the next frame. If there are no other
connections to ::after-paint, connect() vs. connect_after() doesn't
matter.
https://bugzilla.gnome.org/show_bug.cgi?id=732350
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".
This signal is emitted the first time a frame of contents of the
window is completed by the application and has been drawn on the
screen. This is meant to be used for performance measurement of
application startup.
https://bugzilla.gnome.org/show_bug.cgi?id=732343
With get_input_region existing, get_input_rect is a misnomer. Really,
it's about the geometry of the output surface, and it's only used that
way in the compositor code.
Way back when in GNOME 3.2, get_input_rect was added when we added
invisible borders. get_outer_rect was always synonymous with server-side
geometry of the toplevel. get_outer_rect was used for both user-side
policy (the "frame rect") and to get the geometry of the window.
Invisible borders were meant to extend the input region of the frame
window silently. Since most users of get_outer_rect cared about the
frame rect, we kept that the same and added a new method, get_input_rect
to get the full rect of the framed window with all invisible borders for
input kept on.
As time went on and CSD and Wayland became a reality, the relationship
between the server-side geometry and the "frame rect" became more
complicated, as can be evidenced by the recent commits. Since clients
don't tend to be framed anymore, they set their own input region.
get_buffer_rect is also sort of a poor name, since X11 doesn't really
have buffers, but we don't really have many other alternatives.
This doesn't change any of the code, nor the meaning. It will always
refer to the rectangle where the toplevel should be placed.
The smallest possible spread corresponds to an unblurred shadow, which
neither grows nor shrinks - thus the spread should be zero not negative
as returned by our current calculation.
https://bugzilla.gnome.org/show_bug.cgi?id=731353
Avoid populating *_VERSION constants through cflags in pkg-config-file
which could be overridden by the project using it. Properly prefix the
defines with META_ to make gi-scanner happy.
When opening the window menu without an associated control - e.g.
by right-clicking the titlebar or by keyboard - using coordinates
for the menu position is appropriate. However when the menu is
associated with a window button, the expected behavior in the
shell can be implemented much easier with the full button geometry:
the menu will point to the center of the button's bottom edge
rather than align to the left/right side of the titlebar as it
does now, and the clickable area where a release event does not
dismiss the menu will match the actual clickable area in mutter.
So add an additional show_window_menu_for_rect() function and
use it when opening the menu from a button.
https://bugzilla.gnome.org/show_bug.cgi?id=731058
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
It looks weird to have Alt+Space pop up under the cursor instead
of the top-left corner of the window, and the Wayland request will
pass through the coordinates as well.
Add it to the compositor interface, and extend the
_GTK_SHOW_WINDOW_MENU ClientMessage to support it as well.
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
Talking it over with Owen, we weren't sure why this was here.
At one point, we were creating a foreign stage window, so potentially
Clutter didn't select for its own events, but now we're using a standard
stage window, so this seems weird.
Why we did it on the COW, nobody knows. Maybe copy/paste bugginess?
Each level in the tower is initialized by binding the texture for that
level to an offscreen framebuffer and rendering the previous level as a
textured rectangle. The problem was that we were blending the previous
level with undefined data so argb32 windows with transparencies would
result in artefacts. This makes sure to disable blending when drawing
the textured rectangle.
The code here before was completely wrong. Not only did it mix up
coordinate spaces of "client rect" vs. "frame rect", but it used
meta_frame_get_frame_bounds, which is specifically for the *visible*
bounds of a window!
In the case that we don't have a bounding or input shape region at
all on the client window, the input shape that we should apply is
the surface's natural shape. So, set the region to NULL to get the
natural rect picking semantics.
Compositors haven't been able to manage more than one screen for
quite a while. Merge MetaCompScreen into MetaCompositor, and update
the API to match.
We still keep MetaScreen in the public compositor API for compatibility
purposes.
We previously separated out MetaDisplay and MetaScreen. mutter
would only manage one screen, but we still kept a list of screens
for simplicity.
With Wayland support, we no longer care about the ability to
manage more than one screen at a time. Remove this by killing
the list of screens, in favor of having just one MetaScreen
in MetaDisplay.
We also kill off active_screen at the same time, since it's
not necessary anymore.
A future cleanup should merge MetaDisplay and MetaScreen. To avoid
breaking API, we should probably keep MetaScreen around as a dummy
type.
cogl_texture_get_components can be used on both X11 and Wayland
backends. Technically, the detection is different: we actually
check the actual RENDER format in the old code, while Cogl simply
assumes that any pixmap with a depth >= 32 is ARGB32. Since Cogl
already seems to be working with its internal checks, it makes
more sense to use Cogl's check rather than keeping our own.
is_argb32 can be called at any time, including times when we don't
have a texture. In that case, just assume we're ARGB32. The value
really shouldn't be important though.
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
This ensures that we send the proper leave and enter events to wayland
clients.
Particularly, this solves a bug in SSD xwayland windows where clicking
and dragging on the title bar to move the window only works on the odd
turn (unless the pointer moves away from the title bar between
tries). This happens because xwayland gets a button press but doesn't
see the release so when it gets the next button press it discards it
because its pointer button tracking logic says that the button is
already pressed. Sending the proper wayland pointer leave event fixes
it since wayland clients must forget about button state at that point.
https://bugzilla.gnome.org/show_bug.cgi?id=726123
Creating a new cogl texture may fail, in which case the intent to
free it will crash. While something is clearly wrong (insanely
large window, oom, ...), crashing the WM is harsh and we should
try to avoid it if at all possible, so carry on.
https://bugzilla.gnome.org/show_bug.cgi?id=722266
All WM events (passive button grabs and passive keyboard grabs)
are handled through clutter now, so we must make sure we spoof
them even if they happen on frames (because that's where we
grab on)
Weirdly, clutter stopped segfaulting when we call clutter_x11 methods
and the backend is not right, but this is correct anyway, and
probably fixes some BadDrawable errors in mutter-wayland on x11,
caused by mixing windows of the outer X and windows of Xwayland.
Mouse event handling was duplicated, resulting in weird interactions
if clutter was allowed to see certain events (for example under
wayland, where it gets all events). Because now clutter sees all
X events, even when running as an x11 compositor, we can handle
everything using the clutter variants.
At the same time, rewrite a little the passive button grab code,
to make it clear what is being matched on what and why.
