Previously in 3de1a39, it "worked by accident" in my testing since the
display being used in `map_to_output` was initialized first (the map
would not be applied because the display hadn't actually come online
yet), and was followed by a second display (at which point the map would
get applied for the first display).
Refs #5231
Fixes#4819.
This commit ensures that `seat_set_focus` is called to transfer focus
when a window is selected via a pen. Previously, it would race with
`node_at_coords`, and only properly transfer focus if its returned
`surface` was NULL.
Some input rules, like `map_to_output`, are dependent on a specific
screen being present. This currently does not work for hotplugged
outputs, or outputs that are processed after the input device is
initially probed.
This commit fixes both cases, by reconfiguring inputs on each output
addition.
Fixes#5231.
This commit refactors `cursor_handle_activity` to also take the idle
source, so that it can be reused for tablet and touch activity.
Previously, the timeouts would be tracked, but the cursor would never be
un-hidden for anything but pointers.
Fixes#5169.
If we started holding the tool tip down on a surface that accepts tablet
v2, we should notify that surface if it gets released over a surface
that doesn't support v2.
Since GTK supports tablet v2, this fixes the common case of starting a
drag over a GTK surface (e.g. scrollbar) and releasing it outside (e.g.
over the gaps between sway containers, or in a terminal).
Refs #5230.
See issue #5228. Currently, WL_OUTPUT_SUBPIXEL_NONE is ignored and
CAIRO_ANTIALIAS_SUBPIXEL is still set. This commit checks if subpixel is
set to none and if so, calls set_antialias with CAIRO_ANTIALIAS_GRAY.
This mirrors the functionality in Mako's
[PR261](https://github.com/emersion/mako/pull/261)
Instead of handling presses and releases on empty workspaces as setting
focus to the workspace, handle releases by notifying the seat of a
pointer action. This way DnDs are correctly released if the button is
released over an empty workspace. This is achieved by removing the early
return and letting the handle_button() call seat_pointer_notify_button()
at the very end.
Fixes#3932
Instead of removing the destroy listeners in the output destroy, remove
them in the damage destroy handler. Fixes the following use after free:
==646625==ERROR: AddressSanitizer: heap-use-after-free on address 0x61200017cab8 at pc 0x0000004f8f29 bp 0x7ffdf465ad30 sp 0x7ffdf465ad20
WRITE of size 8 at 0x61200017cab8 thread T0
#0 0x4f8f28 in wl_list_remove ../common/list.c:181
#1 0x43dd24 in handle_destroy ../sway/desktop/output.c:790
(`wl_list_remove(&output->damage_destroy.link);` here, 214e3030e1dce master branch)
#2 0x7f0e573a1c93 in wlr_signal_emit_safe ../util/signal.c:29
#3 0x7f0e57390954 in wlr_output_destroy ../types/wlr_output.c:365
#4 0x7f0e5735e37f in backend_destroy ../backend/x11/backend.c:128
#5 0x7f0e57348147 in wlr_backend_destroy ../backend/backend.c:47
#6 0x7f0e57356f75 in multi_backend_destroy ../backend/multi/backend.c:54
#7 0x7f0e5735710e in handle_display_destroy ../backend/multi/backend.c:107
#8 0x7f0e573f23e4 in wl_display_destroy (/lib64/libwayland-server.so.0+0x93e4)
#9 0x42f0b2 in server_fini ../sway/server.c:177
#10 0x42dd01 in main ../sway/main.c:414
#11 0x7f0e570f7041 in __libc_start_main (/lib64/libc.so.6+0x27041)
#12 0x40e3bd in _start (/opt/wayland/bin/sway+0x40e3bd)
0x61200017cab8 is located 120 bytes inside of 320-byte region [0x61200017ca40,0x61200017cb80)
freed by thread T0 here:
#0 0x7f0e57aa9357 in __interceptor_free (/lib64/libasan.so.6+0xb0357)
#1 0x7f0e5738b877 in wlr_output_damage_destroy ../types/wlr_output_damage.c:143
#2 0x7f0e5738b2b9 in output_handle_destroy ../types/wlr_output_damage.c:13
#3 0x7f0e573a1c93 in wlr_signal_emit_safe ../util/signal.c:29
#4 0x7f0e57390954 in wlr_output_destroy ../types/wlr_output.c:365
#5 0x7f0e5735e37f in backend_destroy ../backend/x11/backend.c:128
#6 0x7f0e57348147 in wlr_backend_destroy ../backend/backend.c:47
#7 0x7f0e57356f75 in multi_backend_destroy ../backend/multi/backend.c:54
#8 0x7f0e5735710e in handle_display_destroy ../backend/multi/backend.c:107
#9 0x7f0e573f23e4 in wl_display_destroy (/lib64/libwayland-server.so.0+0x93e4)
previously allocated by thread T0 here:
#0 0x7f0e57aa9887 in __interceptor_calloc (/lib64/libasan.so.6+0xb0887)
#1 0x7f0e5738b532 in wlr_output_damage_create ../types/wlr_output_damage.c:91
#2 0x43e4a7 in handle_new_output ../sway/desktop/output.c:875
#3 0x7f0e573a1c93 in wlr_signal_emit_safe ../util/signal.c:29
#4 0x7f0e57357261 in new_output_reemit ../backend/multi/backend.c:143
#5 0x7f0e573a1c93 in wlr_signal_emit_safe ../util/signal.c:29
#6 0x7f0e5736030a in wlr_x11_output_create ../backend/x11/output.c:253
#7 0x7f0e5735e309 in backend_start ../backend/x11/backend.c:113
#8 0x7f0e573480fb in wlr_backend_start ../backend/backend.c:36
#9 0x7f0e57356e61 in multi_backend_start ../backend/multi/backend.c:31
#10 0x7f0e573480fb in wlr_backend_start ../backend/backend.c:36
#11 0x42f4ba in server_start ../sway/server.c:205
#12 0x42dbd7 in main ../sway/main.c:394
#13 0x7f0e570f7041 in __libc_start_main (/lib64/libc.so.6+0x27041)
Fixes#5158
An if branch takes care of the case where the output needs to be turned
off (DPMS'ed or disabled). The other branch needs to unconditionally
enable the output.
output->current_mode is already taken care of in apply_config.
Sorry about that, probably made a bad change by mistake after my DRM testing.
Closes: https://github.com/swaywm/sway/issues/5193
This color, both in i3 and as described in sway(5), defaults to #888888.
However, the actual default also has an alpha of 88 instead of FF,
meaning it ends up significantly darker than intended.
With these changes, sway will respect positive exclusive zones of layer
surfaces anchored to one or three sides.
This matches the protocol, which states that a positive exclusive zone
should be respected, "if the surface is anchored to one edge or an
edge and both perpendicular edges". If the surfaces is "anchored to
only two perpendicular edges (a corner), anchored to only two
parallel edges or anchored to all edges a positive value will be
treated the same as zero".
This allows e.g. triggering one command while a key is held, then
triggering another to undo the change performed by it afterwards. One
use case for this is triggering push-to-talk functionality for VoIP
tools without granting them full access to all input events.
Fixes#3151
Before the default was "smart". This means a visible app could steal focus
whenever it wanted to. This can be an issue since having focus allows for
e.g. clipboard access.
This commit changes the default to "urgent" instead.
Closes: https://github.com/swaywm/sway/issues/5139
In the case of multiple overlapping floating windows, this commit fixes an issue where the wrong window would be focused in response to a cursor if one of the windows came from a different output (overhanging).
* check for workspace command name arg (fix#5131)
For the 'workspace <name> output <output>' command, output_location must
be greater than zero or the attempt to get the workspace name with
join_args will segfault or abort() (depending on the flavor of
sway_assert() in use). This checks and returns an error instead.
* put workspace output error string on one line
To ease grepping as requested
* check for name in workspace gaps command as well
A malformed command here will lead to the same result seen in #5131, so
add a check. Done inside the cmd_workspace_gaps() function itself, to
take advantage of the existing 'Expected...' string.
If a pad device for a tablet exists, reloading the configuration,
removing/reading the device or even suspending the system will recreate
the same Wayland input device multiple times.
Make sure we don't re-create the same Wayland device more than
necessary.
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Remove some unnecessary brackets in an error condition check identified
during review of the shortcuts inhibitor command code (#5021).
Signed-off-by: Michael Weiser <michael.weiser@gmx.de>
Add a command to influence keyboard shortcuts inhibitors. In its current
form it can be used to activate, deactivate or toggle an existing
inhibitor on the surface currently receiving input. This can be used to
define an escape shortcut such as:
bindsym --inhibited $mod+Escape seat - shortcuts_inhibitor deactivate
It also allows the user to configure a per-seat default of whether
keyboard inhibitors are honoured by default (the default) or not. Using
the activate/toggle command they can then enable the lingering inhibitor
at a later time of their choosing.
As a side effect this allows to specifically address a named seat for
actions as well, whatever use-case that might serve.
Signed-off-by: Michael Weiser <michael.weiser@gmx.de>
Adding support for the keyboard shortcuts inhibit protocol allows remote
desktop and virtualisation software to receive all keyboard input in
order to pass it through to their clients so users can fully interact
the their remote/virtual session. The software usually provides its own
key combination to release its "grab" to all keyboard input. The
inhibitor can be deactivated by the user by removing focus from the
surface using another input device such as the pointer.
Use support for the procotol in wlroots to add support to sway. Extend
the input manager with handlers for inhibitor creation and destruction
and appropriate bookkeeping. Attach the inhibitors to the seats they
apply to to avoid having to search the list of all currently existing
inhibitors on every keystroke and passing the inhibitor manager around.
Add a helper function to retrieve the inhibitor applying to the
currently focused surface of a seat, if one exists.
Extend bindsym with a flag for bindings that should be processed even if
an inhibitor is active. Conversely this disables all normal shortcuts if
an inhibitor is found for the currently focused surface in
keyboard::handle_key_event() since they don't have that flag set. Use
above helper function to determine if an inhibitor exists for the
surface that would eventually receive input.
Signed-off-by: Michael Weiser <michael.weiser@gmx.de>
This enables/disables adaptive synchronization on the output.
For now, the default is disabled because it might cause flickering on
some hardware if clients don't submit frames at regular enough
intervals. In the future an "auto" option will only enable adaptive sync
if a fullscreen client opts-in via a Wayland protocol.
This correct the description of the commmand:
workspace [--no-auto-back-and-forth] [number] <[num:]name>
Previously, the number and num pieces were being confused. This also
documents the behavior of the --no-auto-back-and-forth flag.
This can be used as a workaround to flag terminal windows as urgent when
commands are completed, until urgency is introduced in the Wayland
protocol.
Configure your shell to run `swaymsg "[pid=$PPID] urgent enable"` when
commands are completed, and use a terminal which uses one process per
window.
This adds support for input type configs to input_cmd_events. This works
similar to the wildcard handling that existed where configs for the
devices are stored and the type config is reset to INT_MIN so that it
does not override.
This also condenses the toggle_send_events and
toggle_wildcard_send_events functions into a single function to reduce
code duplication.
If a subsurface is created for a surface that is associated with a
scratchpad hidden view, do not attempt to send an enter to it. The
subsurface is not on any output and since there is no workspace
associated with the view, attempting to get the output for the NULL
workspace will result in a SIGSEGV.
This patch makes it so users that have configured their screen with a
transform don't have to update their config after the wlroots breaking
change.
References: https://github.com/swaywm/wlroots/pull/2023
The output manager config is created when the output is created. It is
updated when the mode, transform, scale, or layout for the output
changes, as well as, when the output is destroyed.
Since the output->enabled property was not being set before calling
apply_output_config, the output event handlers were early returning and
never updating the output manager config when the output state was
committed.
This fixes the issue by setting output->enabled in apply_output_config
below the output disabling section. There are also a few other minor
changes that are required to function.
Additionally, this renames output_enable to output_configure to better
describe the recent changes.
The condition in test_mark was checking the negation of list_seq_find.
This works fine for the first mark, but fails for every other mark. This
fixes the condition to check for anything other than -1, which is the
value returned from list_seq_find for index not found.
The only output_enable caller is now apply_output_config. Stop calling
apply_output_config from output_enable to simplify the code and avoid
the back-and-forth between these two functions.
output_enable is now the symmetric of output_disable: it just marks the
output as enabled and performs bookkeeping (e.g. creating teh default
workspace). It is called from apply_output_config after the output
commit, so that it can read the current output state and act
accordingly.
This change also allows us to avoid an extraneous wlr_output_commit.
References: https://github.com/swaywm/sway/issues/4921
apply_output_config will call output_enable if necessary.
This fixes a lone wlr_output_enable call (without a matching
wlr_output_commit call) which was a no-op.
This makes it so invalid configs will return the exit code 1 when the
validation flag is given. This also reduces the log level to SWAY_ERROR,
which makes it so only the errors are shown. If someone wants more
verbose output, the can use the -V/--verbose or -d/--debug flags.
Additionally, this also makes it so swaybg will not be spawned when
validating the config.
It looks like the code to drop privileges may have been broken via
commit 37f0e1f. That commit reverted the correct order from #911, which
first drops the gid then the uid. If setuid is called first then the
target user may not have the ability to setgid.
The container_at_tabbed and container_at_stacked container were checking
the bounds along the y-axis, but not the x-axis. This made it possible
to cause a segfault for specific resolution, horizontal gap, and
workspace children lengths. The issue is that child_index was -1 and was
resulting in a buffer underflow. Adding the x-axis bound checks for
early returns should prevent this from happening.
This fixes the following crash:
#0 0x00007f7daac3af25 in raise () at /usr/lib/libc.so.6
#1 0x00007f7daac24897 in abort () at /usr/lib/libc.so.6
#2 0x00007f7daac24767 in _nl_load_domain.cold () at /usr/lib/libc.so.6
#3 0x00007f7daac33526 in () at /usr/lib/libc.so.6
#4 0x0000555bfbc35029 in seat_set_focus_layer (seat=0x555bfd76d860, layer=0x555bfdda2ff0)
at ../sway/input/seat.c:1164
__PRETTY_FUNCTION__ = "seat_set_focus_layer"
__PRETTY_FUNCTION__ = "seat_set_focus_layer"
#5 0x0000555bfbc35029 in seat_set_focus_layer (seat=seat@entry=0x555bfd76d860, layer=0x555bfdda2ff0)
at ../sway/input/seat.c:1164
__PRETTY_FUNCTION__ = "seat_set_focus_layer"
#6 0x0000555bfbc25899 in handle_output_destroy (listener=0x555bfdb90688, data=<optimized out>)
at ../sway/desktop/layer_shell.c:263
layer = 0x555bfdd6b040
sway_layer = 0x555bfdb90610
seat = 0x555bfd76d860
client = 0x555bfdb76d70
set_focus = <optimized out>
#7 0x0000555bfbc5b669 in wl_signal_emit (data=0x555bfd795930, signal=0x555bfd795ae0)
at /usr/include/wayland-server-core.h:472
l = <optimized out>
next = 0x555bfdb6a3e8
__PRETTY_FUNCTION__ = "output_disable"
index = <optimized out>
#8 0x0000555bfbc5b669 in output_disable (output=output@entry=0x555bfd795930)
at ../sway/tree/output.c:263
__PRETTY_FUNCTION__ = "output_disable"
index = <optimized out>
#9 0x0000555bfbc3b890 in apply_output_config (oc=0x555bfd7d73d0, output=output@entry=0x555bfd795930)
at ../sway/config/output.c:321
wlr_output = 0x555bfd7afaf0
scale = <optimized out>
output_box = <optimized out>
#10 0x0000555bfbc28309 in handle_output_manager_apply
(listener=0x555bfbc7f148 <server+488>, data=0x555bfdca6eb0) at ../sway/desktop/output.c:936
wlr_output = <optimized out>
output = 0x555bfd795930
oc = <optimized out>
server = 0x555bfbc7ef60 <server>
config = 0x555bfdca6eb0
config_head = 0x555bfdb79350
ok = true
#11 0x00007f7dab4fbf7c in wlr_signal_emit_safe (signal=<optimized out>, data=0x555bfdca6eb0)
at ../subprojects/wlroots/util/signal.c:29
pos = 0x555bfbc7f148 <server+488>
l = 0x555bfbc7f148 <server+488>
cursor =
{link = {prev = 0x555bfbc7f148 <server+488>, next = 0x7fff238a8390}, notify = 0x7f7dab4fbef0 <handle_noop>}
end =
{link = {prev = 0x7fff238a8370, next = 0x555bfd7419f8}, notify = 0x7f7dab4fbef0 <handle_noop>}
#12 0x00007f7daa45469a in ffi_call_unix64 () at /usr/lib/libffi.so.6
#13 0x00007f7daa453fb6 in ffi_call () at /usr/lib/libffi.so.6
#14 0x00007f7daae6f82f in () at /usr/lib/libwayland-server.so.0
#15 0x00007f7daae6c193 in () at /usr/lib/libwayland-server.so.0
#16 0x00007f7daae6d7f2 in wl_event_loop_dispatch () at /usr/lib/libwayland-server.so.0
#17 0x00007f7daae6c39c in wl_display_run () at /usr/lib/libwayland-server.so.0
This crash happens because focus can only be set on mapped surfaces.
Closes#4929
Replaces criteria_get_views with criteria_get_containers which can
return containers without views when the criteria only contains
container properties.
Handle binding event types that cannot be encoded gracefully by dropping
the event. This prevents issues for binding types like BINDING_SWITCH,
where the event would cause a crash.
If applying an output config to an output fails, the output may be
destroyed. To be able to handle this situation correctly,
apply_output_config_to_outputs needs to use wl_list_for_each_safe.
This fixes a memory leak of oc (the output config) in handle_new_output.
Output configs returned from find_output_config are not stored and need
to be freed after use.
When a container was being made fullscreen and it is on the focused
workspace for a seat, focus was being set to the container. However,
when the container was on a non-focused workspace, the focus stack
wasn't being touched. When assigning a fullscreen container to a
workspace or moving a fullscreen container to a different workspace,
this would make it so the fullscreen container was never added to the
focus stack for the workspace thus preventing access to the workspace.
This adds the container to the top of the focus stack, behind the
container on the focused workspace.
Two changes were made:
- Bind the texture before glTexParameteri
- Set the scaling filter before each wlr_render_texture_with_matrix call
Logging in wlroots allows to check that the scaling filter is properly
set prior to rendering.
Fixes: 6968fb3123 ("add scale_filter output config option")
Closes: https://github.com/swaywm/sway/issues/4798
If a view is mapped to a workspace using an assign, the pid should still
be removed from the pid mapping list. This prevents child processes from
matching against it and mapping a view to a likely undesired workspace.
This fixes a crash when attempting to listen to a signal on a NULL
cursor image surface. If the surface is NULL, the listener is just
reinitialized using wl_list_init.
This adds a listener for the destroy event of the cursor image surface.
This prevents a use-after-free when the last visible image surface is
freed, there has not been a new cursor set, and the cursor is reshown.
This adds a check to make it so the indicator is only rendered on views
with a parent, which floating views do not. Since floating views do not
have a parent, the workspace layout was being incorrectly used to
determine whether to show the split indicator previously. This has no
impact on floating containers and a view within a floating container
will still have indicators rendered appropriately.
In case xcb-iccm is not installed on the system, size_hints will be
null. Handle this as if the get_constraints functions was not
implemented and return the defaults.
Fixes#4870
When gaps are resized for lack of space the calculation could result in
a gap size of non-integer pixels. This would result in containers
located at non-integer pixels which would be subtly broken.
Because the layout code rounds down the dimensions of the windows
resizing would often be off by one pixel. The width/height fraction
would not exactly reflect the final computed width and so the resize
code would end up calculating things wrong.
To fix this first snap the container size fractions to the pixel grid
and only then do the resize. Also use round() instead of floor() during
layout to avoid a slightly too small width. This applies in two cases:
1. For the container we are actually resizing using floor() might result
in being 1px too small.
2. For the other containers it might result in resizing them down by 1px
and then if the container being resized is the last all those extra
pixels would make the resize too large.
Fixes#4391
For i3 compatibility, allow the indicator and child_border colors values
to be optional. The indicator will fallback to sane defaults and
child_border will fallback to the background color for the class.
This is the third commit in a series of commits to refactor color
handling in sway. This removes add_color from commands.c. It was only
being used by bar_cmd_colors. This also changes the functions to use
parse_color which is used to validate rgb(a) colors throughout the code
base and is also what i3bar is using to parse the colors after they are
passed over ipc. After parsing the color and ensuring it is valid, the
rgba hex string is then generated using snprintf. This refactor also
ensures that all the colors for the command are valid before applying
any of them.
This is the second in a series of commits to refactor the color handling
in sway. This removes the duplicated color parsing code in
sway/commands/client.c. Additionally, this combines the parsing of
colors to float arrays with that in sway/config.c and introduces a
color_to_rgba function in commom/util.c.
As an added bonus, this also makes it so non of the colors in a border
color class will be changed unless all of the colors specified are
valid. This ensures that an invalid command does not get partially
applied.
When arranging layer-shell layers, verify that the currently focused
layer, if any, for each seat is still keyboard interactive. If the layer
is no longer keyboard interactive and there is not a keyboard
interactive overlay or top layer to change the focus to, refocus the
focus inactive node for the seat.
When many surfaces are created, sway can run out of file descriptors,
making wl_event_loop_add_timer (which creates a timerfd) fail and
return NULL. This patch posts a "no memory" error when that is the case,
and only removes the timer if it was created.
(Why "no memory"? It is not easy to distinguish between failures due to
running out of memory and failures due to running out of file
descriptors. Also, using the newer `wl_client_post_implementation_error`
function would lead to an increased version requirement for the
libwayland-server dependency.)
This removes `seat <seat> keyboard_grouping keymap` and replaces it with
`seat <seat> keyboard_grouping smart`. The smart keyboard grouping will
group based on both the keymap and repeat info. The reasoning for this
is that deciding what the repeat info should be for a group is either
arbitrary or non-deterministic when multiple keyboards in the group have
repeat info configured (unless somehow exposed to the user in a
reproducible uniquely identifiable fashion).
When clicking on the titlebar of a floating container (or descendant of
a floating container), the top-level floating container was being
focused and then allowing you to move the top-level floating container.
This made it so you couldn't switch to a different tab/stack within the
floating container. With this patch, the focus inactive view for the
container that the titlebar is associated with is focused, then the
traversal to the top-level floating container is performed to use with
the move floating operation.
Any descendant of a scratchpad container may be fullscreen so checking
to see if the top-level scratchpad container is fullscreen in
root_scratchpad_hide is not sufficient. This iterates through all
descendants of the scratchpad container
When an input becomes available, the input type config for that device
type will be merged underneath the input identifier config, provided
they both exist. If an input type config gets added or modified at a
later point, then those changes get merged onto the input identifier
configs for that type. However there was a missing case of the input
identifier config being added after the device is already available and
the input type config existing. This makes it so that the first time an
input identifier config gets stored, there will be a check to see if it
matched any of the available devices. If it does, then there will be a
search for the associated input type config, which will be merged
underneath the input identifier config if found.
This defers the destruction of wlr_keyboard_groups until idle. This is
to prevent the keyboard group's keyboard from being destroyed in the
middle of handling a keyboard event. This would occur when changing the
keymap of the last keyboard in a group with a keyboard binding. The
prevents crashing when attempting to update the xkb state of the
keyboard group's keyboard. The sway_keyboard_group is still immediately
destroyed so that the group is no longer used
This adds two missing calls to wl_list_remove to remove the key and
modifier listeners for the keyboard group's keyboard when destroying
the keyboard group. This fixes some crashes when changing the keymap of
the last keyboard in a group with a keyboard binding.
This adds seat configuration options which can be used to configure what
events affect the idle behavior of sway.
An example use-case is mobile devices: you would remove touch from the
list of idle_wake events. This allows the phone to stay on while you're
actively using it, but doesn't wake from idle on touch events while it's
sleeping in your pocket.
Showing a window in the scratchpad can move a visible scratchpad window
from another workspace to the current one. If the scratchpad window was
the last visible container in that workspace, the old workspace should
be destroyed.
If a layer is focused by any seat, it needs to be unfocused on unmap. If
the unmap was due to an output being disabled, there would not be a
sway_output and unmap would do an early return. This results in a
use-after-free if the layer was focused by any seat prior to being
unmapped. This change just moves the refocusing code above the early
returns.
Some wayland clients (mostly GTK3 apps) like eog or evince support
gestures like pinch-to-zoom. These gestures are given to clients
via the pointer_gestures_v1 protocol. This is already supported in
wlroots, so we just need to hook up the events here in sway.
Fixes#4724
When making the reload validation improvements, I forgot that input type
configs are stored in a separate list. This makes it so input type
configs are correctly applied on reload.
Repaint scheduling delays output render and frame done events from
output frame events, and block idle frame events from being scheduled in
between output frame done and output render in this period of time.
If a surface is committed after its frame done event, but before output
render, idle frame requests will be blocked, and the surface relies on
the upcoming render to schedule a frame.
If when the repaint timer expires, output render is deemed unnecessary,
no frame will be scheduled. This can lead to surfaces never having their
frame callbacks fire.
To fix this, we store that a surface has requested a frame in
surface_needs_frame. When the repaint expires, if no render is deemed
necessary, we check this flag and schedule an idle frame.
Fixes#4768
