When the pointer crosses monitors, we account for a single motion event
resulting in the pointer moving across more than 2 monitors, in order
to correctly account each monitor scale and the distance traversed
across each monitor in the resulting relative motion vector.
However, memory on the direction is kept short, each iteration to
find the target view just remembers the direction it came from. This
brings a pathological case with 4 monitors with the same resolution
in a 2x2 grid, and a motion vector that crosses monitors at the
intersection of all 4 in a perfect diagonal. (Say, monitors are
all 1920x1080 and pointer moves from 1920,1080 to 1919,1079).
In that case, the intersection point at the crossing between 4
monitors (say, 1920,1080) will be considered to intersect with 2
edges of each view. Since there is always at least 2 directions to
try, the loop will always find the direction other than the one
it came from, and as a result endlessly jump across all 4 possible
choices.
In order to fix this, consider only the global v/h directions,
we already know if the pointer moves left/right or up/down, so
only consider those directions to jump across monitors.
For the case at hand, this will result in three monitors visited,
(either bottomright/bottomleft/topleft, or bottomright/topright/topleft)
with a total distance of 0,0 in the middle one, effectively
resulting in a correct diagonal motion.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/2598
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2803>
Refactor code so that variables don't depend the on motion line
content, but the other way around. This makes it clearer what each
vector means.
This has no functional changes.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2803>
That means before-update, prepare-paint, before-paint, paint-view, after-paint,
after-update. While yet to be used, it will be used as a transient frame
book keeping object, to maintain object and state that is only valid
during a frame dispatch.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2795>
Writing to fields (in this case the MetaColorDevice::pending_state) in
response to an asynchronous operation that was cancelled means we'll
write to an arbitrary memory location, potentially causing segmentation
faults or memory corruption.
Avoid these segfaults or memory corruption by only updating state if we
weren't cancelled. Also avoid trying to dereference the device pointer
if we're cancelled.
The memory corruption due to this has been causing test flakyness in the
monitor unit tests due, which should now hopefully be fixed.
Fixes: 19837796fe
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2794>
Recent versions of Xwayland can allow or disallow X11 clients from
different endianess to connect.
Add a setting to configure this feature from mutter, who spawns
Xwayland.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2785>
Commit 4e0ffba5c attempted to fix initialization of keyboard a11y,
but mousekeys do attempt to create a virtual input device at a
time that it is too early to try to create one.
Defer this operation until keyboard devices are added, so that
we are ensured to already have the seat input thread set up.
Fixes: 4e0ffba5c - backends/native: Initialize keyboard a11y on startup
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2778>
Commit bf84b24 created meta-enums.h but it's pretty empty so far, the
vast majority of enum definitions is still in common.h.
Move the Meta enum definitions to meta-enums.h as one would expect them
to be found.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2467>
This means we can eliminate the use of scattered singletons that isn't
added by the tests or the test framework itself.
tests: Don't get backend from old singleton getter
Either use the ownership chain, or the explicit test context instance
pointer.
tests/wayland: Pass context to test client constructor
So that we can get the Wayland compositor directly from the context.
tests: Don't get display from singleton
tests/client: Make test client carry a context pointer
tests/runner: Have test cases carry a context pointer
tests/wayland/test-driver: Get backend from context
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2718>
As with the backend commit, this means all objects can reach the
MetaContext by walking up the chain, thus can e.g. get the backend from
the context, instead of the global singleton.
This also is a squashed commit containing:
compositor: Get backend via the context
The MetaCompositor instance is owned by MetaDisplay, which is owned by
MetaContext. Get the backend via that chain of ownership.
dnd: Don't get backend from singleton
window-actor: Don't get backend from singleton
dnd: Don't get Wayland compositor via singleton
background: Don't get the monitor manager from the singleton
plugins: Don't get backend from singleton
This applies to MetaPlugin, it's manager class, and the default plugin.
feedback-actor: Pass a compositor pointer when constructing
This allows getting to the display.
later: Keep a pointer to the manager object
This allows using the non-singleton API in idle callbacks.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2718>
This means objects have an owner, where the chain eventually always
leads to a MetaContext. This also means that all objects can find their
way to other object instances via the chain, instead of scattered global
singletons.
This is a squashed commit originally containing the following:
cursor-tracker: Don't get backend from singleton
idle-manager: Don't get backend from singleton
input-device: Pass pointer to backend during construction
The backend is needed during construction to get the wacom database.
input-mapper: Pass backend when constructing
monitor: Don't get backend from singleton
monitor-manager: Get backend directly from monitor manager
remote: Get backend from manager class
For the remote desktop and screen cast implementations, replace getting
the backend from singletons with getting it via the manager classes.
launcher: Pass backend during construction
device-pool: Pass backend during construction
Instead of passing the (maybe null) launcher, pass the backend, and get
the launcher from there. That way we always have a way to some known
context from the device pool.
drm-buffer/gbm: Get backend via device pool
cursor-renderer: Get backend directly from renderer
input-device: Get backend getter
input-settings: Add backend construct property and getter
input-settings/x11: Don't get backend from singleton
renderer: Get backend from renderer itself
seat-impl: Add backend getter
seat/native: Get backend from instance struct
stage-impl: Get backend from stage impl itself
x11/xkb-a11y: Don't get backend from singleton
backend/x11/nested: Don't get Wayland compositor from singleton
crtc: Add backend property
Adding a link to the GPU isn't enough; the virtual CRTCs of virtual
monitors doesn't have one.
cursor-tracker: Don't get display from singleton
remote: Don't get display from singleton
seat: Don't get display from singleton
backend/x11: Don't get display from singleton
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2718>
While already cleaning up API, if this should ever be more non-static
than a constant, it's better if its a function on the monitor manager
instance than something static.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2718>
Since the Wacom panel rewrite, the "output" setting is handled as
a kind of tri-state for display-integrated tablets:
- If the setting is unset, the device is automatically mapped
to an output
- If the setting is set and not empty, the device is mapped to
the output defined by the EDID data
- If the setting is ['', '', ''], the device is mapped to the
span of all displays, like opaque tablets do.
This distinction for the unset setting fell through the cracks,
so both "Automatic" and "All displays" options were handled as
the former.
Add this distinction, so that display-integrated tablets can
be used like opaque tablets of sorts with no limitations.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2767>
These are the ones attached to a display, thus they are the ones that may need
help from this heuristic. Non-integrated tablets (e.g. Intuos) will default to
the span of all monitors.
Fixes mapping of opaque tablets if a display-integrated tablet of the same
brand is also plugged in.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2767>
The fields of 'priv->video_format.max_framerate' are all of type
uint32_t. Multiplying by G_USEC_PER_SEC can overflow, and equally,
dividing a large numerical type by uint32_t can err too.
Since the variable holding the result is int64_t, cast all uint32_t
fields to int64_t before doing any maths on it.
Spotted while trying to investigating an issue with framerates on
HDMI screencasts.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2762>
Add a helper function that ensures any queued virtual input events have
been flushed from the input thread. This works by posting a task to the
input thread, which will itself queue another callback back to the main
thread. Once the main thread callback is invoked, the flush call is
unblocked and the function returns. Upon this, any previously emitted
virtual input event should have already passed through the input thread
back into the main thread, however not necessarily fully processed.
For making sure it has been processed, one also have to make sure the
stage has been updated, e.g. via `meta_wait_for_paint()`.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2727>
meta_screen_cast_stream_src_set_cursor_sprite_metadata() receives
the cursor sprite, position, and scale, and with that it downloads
the cursor sprite by drawing it into a separate framebuffer, then
calls cogl_framebuffer_read_pixels() in it - this is the offscren
path that is very common when using screen capturing applications
such as OBS Studio.
There's a sneaky issue in this code path though: the 'scale' value
is a float. The cursor size is then determined by multiplying the
sprite width and height - two integer variables - by scale, and
this relies on standard float-to-int conversions. This is problematic
as sometimes the rounded values disagree with what is expected by
cogl_framebuffer_read_pixels(). If the packing of either the cursor
width or height is off by one, glReadPixels() will try to write into
off bounds, which crashes.
This can be reproduced by enabling fractional scaling, setting a 150%
zoom level, on a 4K screen, and opening any commit with an image diff
in gitlab.gnome.org, all while screencasting. When hovering the new
image, the cursor sprite will be such that it triggers this code path,
and reproduces this issue.
Fix this by always ceiling the cursor sprite sizes.
Closes https://gitlab.gnome.org/GNOME/mutter/-/issues/2542
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2736>
On hotplug, the events we receive from the kernel are async, and
connectors in the kernel come and go as they please. In practice, this
means that calling drmModeGetConnector() twice more or less directly
after each other, there is no guarantee that the latter call will return
anything if the former did.
When updating the connector in response to hotplugs, we'd first update
the list of existing connectors, and following that, query each and
every one again for their current state, to update our internal
representation; only the former handled drmModeGetConnector() returning
NULL, meaning if unlucky, we'd end up doing a null pointer dereference
when trying to update the state.
Handle this by querying the kernel for the current connector state only
once per connector, updating the list of connectors and their
corresponding state at the same time.
Fixes the following crash:
#0 meta_kms_connector_read_state at ../src/backends/native/meta-kms-connector.c:684
#1 meta_kms_connector_update_state at ../src/backends/native/meta-kms-connector.c:767
#2 meta_kms_impl_device_update_states at ../src/backends/native/meta-kms-impl-device.c:916
#3 meta_kms_device_update_states_in_impl at ../src/backends/native/meta-kms-device.c:267
#4 meta_kms_update_states_in_impl at ../src/backends/native/meta-kms.c:604
#5 update_states_in_impl at ../src/backends/native/meta-kms.c:620
#6 meta_kms_run_impl_task_sync at ../src/backends/native/meta-kms.c:435
#7 meta_kms_update_states_sync at ../src/backends/native/meta-kms.c:641
#8 handle_hotplug_event at ../src/backends/native/meta-kms.c:651
#9 on_udev_hotplug at ../src/backends/native/meta-kms.c:668
Related: https://bugzilla.redhat.com/show_bug.cgi?id=2131269
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2709>
There is no need to use the 'bypass-*' method of event processing in the
changed function since in all cases the 'bypass-*' variable was set, any
following event processing functions would ignore the event anyway.
Simplify things a bit by just returning TRUE if the event is consumed.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2714>
It is generally assumed here and there that the pointer at all point in
time is within some logical monitor, if there is any logical monitor to
be within.
With the input thread, this was for a short amount of time not reliable,
resulting in crashes in combination with hotplugging or suspend/resume,
where monitors come and go quickly.
What happens is that the pointer at first is within a logical monitor,
but when that logical monitor is removed, while the new monitor
viewports are handed to the input thread, the constraining happens
asynchronously, meaning there is a time between between the new
viewports are sent, and before clutter_seat_query_state() starts
reporting the constrained position.
If a new client mapped a maximized window during this short time frame,
we'd crash with
#0 meta_window_place at ../src/core/place.c:883
#1 place_window_if_needed at ../src/core/constraints.c:562
#2 meta_window_constrain at ../src/core/constraints.c:310
#3 meta_window_move_resize_internal at ../src/core/window.c:3869
#4 meta_window_force_placement at ../src/core/window.c:2120
#5 xdg_toplevel_set_maximized at ../src/wayland/meta-wayland-xdg-shell.c:429
#6 ffi_call_unix64 at ../src/x86/unix64.S:105
#7 ffi_call_int at ../src/x86/ffi64.c:672
#8 wl_closure_invoke at ../src/connection.c:1025
#9 wl_client_connection_data at ../src/wayland-server.c:437
The fix for this is to make sure that the viewports are updated and
pointers constrained synchronously, i.e. the main thread will wait until
after the input thread is done constraining before continuing.
Related: https://bugzilla.redhat.com/show_bug.cgi?id=2147502
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2711>
The CRTC cursor sprite scale was incorrectly assumed to be always 1.0
when using the default not-scale-monitor-framebuffer mode. This is
harmless in most cases, as most clients provide HiDPI capable cursors,
but for the ones that didn't, we'd end up drawing their cursors
unscaled, when using the cursor planes.
Fix this by using the "texture scale" which is what is intended for
this.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/2477
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2698>
Cursor planes tend to be ARGB8888 and support no other format (ideally
we should not hard code this, but un-hard-coding that is for another
day), and if we put e.g. a XRGB8888 buffer in there, it'll either result
in the gbm_bo allocation failing (it doesn't allow USE_CURSOR with any
other format) or mode setting failing if using dumb buffers directly.
In the former case, we'll fall back to OpenGL indefinitely, and in the
latter, we'll have failed mode sets as long as we try to set the invalid
cursor buffer as the cursor plane.
Change things to process all buffers that are not ARGB8888 using the
scale/rotate machinery we already have, turning XRGB8888 into ARGB8888.
Related: https://gitlab.gnome.org/GNOME/mutter/-/issues/2477
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2698>
