It was removed in 3.34 as part of 6ed5d2e2. And we thought that was the
only thread that might exist and use X11. But the top gnome-shell crasher
in 3.36 seems to suggest otherwise.
We don't know what or where the offending thread is, but since:
1. We used XInitThreads for years already prior to 3.34; and
2. Extensions or any change to mutter/gnome-shell could conceivably use
threads to make X calls, directly or indirectly,
it's probably a good idea to reintroduce XInitThreads. The failing assertion
in libx11 is also accompanied by a strong hint:
```
fprintf(stderr, "[xcb] Most likely this is a multi-threaded client " \
"and XInitThreads has not been called\n");
```
https://bugs.launchpad.net/bugs/1877075
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/1252https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1256
(cherry picked from commit 1d5f9b6917)
This changes the semantics a bit, e.g. we will never include the pointer
cursor sprite, as there is no way to know whether the caller wants to or
not.
We also change things a bit so that when we render to an offscreen paint
context, we don't emit the "paint" signal on actors, as doing so would
end up recursing in gnome-shell's screenshot and screencast code.
Closes: https://gitlab.gnome.org/GNOME/gnome-shell/-/issues/2567https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1222
Will be used by the stage to not paint the overlays. We skip all
overlays since overlays are only ever used for pointer cursors when the
hardware cursors cannot or should not be used.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1222
These phase callbacks are not intended to be inovked when something
secondary is painting the stage, such as a screen cast stream, or
similar. Thus, only invoke the callbacks when there is a view associated
with the paint context, which will not be the case for offscreen
painting.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1222
Either onto a framebuffer, or into a CPU memory buffer. The latter will
use an former API and then copy the result to CPU memory. The former
allocates an offscreen framebuffer, sets up the relevant framebuffer
matrices and paints part of the stage defined by the passed rectangle.
This will be used by a RecordArea screen cast API. The former to paint
directly onto PipeWire handled dma-buf framebuffers, and the latter for
PipeWire handled shared memory buffers.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1222
If there is a paint context available (i.e. for the phases that are
during the actual stage paint), pass it along the callbacks, so that
the callback implementations can change their operation depending on the
paint context state.
This also means we can get the current view from the paint context,
instead of the temporarily used field in the instance struct.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1222
A paint flag affects a paint operation in ways defined by the flags.
Currently no flags are defined, so no semantical changes are defined
yet. Eventually a flag aiming to avoid painting of cursors is going to
be added, so that screen cast streams can decide whether to include a
cursor or not.
Changes for gnome-3-36: Removed flag from offscreen context.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1222
The redraw clip that's painted together with the damage region has to be
copied earlier than we do right now. That's because if
PAINT_DAMAGE_REGION is enabled, buffer age is disabled and thus
use_clipped_redraw is FALSE. That means the redraw_clip is updated and
set to the full view-rect. If we copy the queued_redraw_clip after that,
it's also going to be set to the full view-rect. So copy the redraw clip
a bit earlier to make sure we're actually passing the real redraw clip
to paint_damage_region().
Also keep the queued_redraw_clip around a bit longer so it can actually
be used by paint_damage_region() and isn't freed before that.
While at it, move paint_damage_region() from swap_framebuffer() into
clutter_stage_cogl_redraw_view() so we don't have to pass things to
swap_framebuffer() only for debugging.
Fixes https://gitlab.gnome.org/GNOME/mutter/-/issues/1104https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1208
(cherry picked from commit 793a9d45e1)
When resizing an X11 window with client side decorations, the shadow is
clipped by the frame bounds so that we don't need to paint the shadow
under the opaque areas covered by the window and its frame.
When the X11 client uses the EMWH synchronization mechanism (like all
gtk-3 based clients), the actual window may not be updated so that the
actual window and it frame may be behind the expected window frame
bounds, which gives the impression of de-synchronized shadows.
To avoid the issue, keep a copy of the frame bounds as a cache and only
update it when the client is not frozen so that the clipping occurs on
the actual content.
Closes: https://gitlab.gnome.org/GNOME/mutter/issues/1178https://gitlab.gnome.org/GNOME/mutter/merge_requests/1214
(cherry picked from commit bd45a00fa3)
If drmModePageFlip() or custom_page_flip_func fails, process_page_flip() was
forgetting to undo the ref taken for that call. This would leak page_flip_data.
The reference counting works like this:
- when created, ref count is 1
- when calling drmModePageFlip, ref count is increased to 2
- new: if flip failed, ref count is decreased back to 1
- if calling schedule_retry_page_flip(), it takes a ref internally
- if calling mode_set_fallback(), it takes a ref internally
- all return FALSE paths have an explicit unref
- return TRUE path has an explicit unref
This issue was found by code inspection and while debugging an unrelated issue
with debug prints sprinkled around. I am not aware of any end-user visible
issues being fixed by this, as the leak is small and probably very rare.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1209
(cherry picked from commit 36111270aa)
When testing a laptop with intel and DisplayLink devices, attempting to set the
DL output as the only active output resulted in GNOME/Wayland freezing. The
main event loop was running fine, but nothing on screen would get updated once
the DL output become the only one. This patch fixes that issue.
DisplayLink USB 3 devices use an out-of-tree kernel DRM driver called EVDI.
EVDI can sometimes fail drmModePageFlip(). For me, the flip fails reliably when
hotplugging the DL dock and when changing display configuration to DL only.
Mutter has a workaround for failing flips, it just calls drmModeSetCrtc() and
that succeeds.
What does not work reliably in the fallback path is Mutter keeping track of the
pageflip. Since drmModePageFlip() failed, there will not be a pageflip event
coming and instead Mutter queues a callback in its stead. When you have more
than one output, some other output repainting will attempt to swap buffers and
calls wait_for_pending_flips() which has the side-effect of dispatching any
queued flip callbacks. With multiple outputs, you don't get stuck (unless they
all fail the exact same way at the same time?). When you have only one output,
it cannot proceed to repaint and buffer swap because the pageflip is not marked
complete yet. Nothing dispatches the flip callback, leading to the freeze.
The flip callback is intended to be an idle callback, implemented with a
GSource. It is supposed to be called as soon as execution returns to the main
event loop. The setup of the GSource is incomplete, so it will never dispatch.
Fix the GSource setup by setting its ready-time to be always in the past. That
gets it dispatched on the next cycle of the main event loop. This is now the
default behavior for all sources created by meta_kms_add_source_in_impl().
Sources that need a delay continue to do that by overriding the ready-time
explicitly.
An alternative solution could have been to implement GSource prepare and check
callbacks returning TRUE. However, since meta_kms_add_source_in_impl() is used
by flip retry code as well, and that code needs a delay through the ready-time,
I was afraid I might break the flip retry code. Hence I decided to use
ready-time instead.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1209
(cherry picked from commit 6e0cfd3e55)
According to the XSetSelectionOwner libX11 documentation:
[...] If the owner window it has specified in the request is later
destroyed, the owner of the selection automatically reverts to None,
but the last-change time is not affected.
This is indeed visible through the selection_timestamp field in
XFixesSelectionNotify events.
Use this to check whether the selection time is recent-ish (thus
likely coming from an explicit XSetSelectionOwner request) and honor
the client intent by setting a "NULL" owner. If the selection time
is too old, it's definitely an indication of the owner client being
closed, the scenario where we do want the clipboard manager to take
over.
This fixes two usecases:
- X11 LibreOffice / WPS clear the selection each time before copying
its own content. Mutter's clipboard manager would see each of those
as a hint to take over, competing with the client over selection
ownership. This would simply no longer happen
- Password managers may want to clear the selection, which would be
frustrated by our clipboard manager.
There's a slight window of opportunity for the heuristics to fail
though, if a X11 client sets the selection and closes within 50ms, we
would miss the clipboard manager taking over.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1206
(cherry picked from commit 5671f0a284)
The X11 selection source was being preserved after unsetting its
ownership. This is no leak as it would be eventually replaced by
another source, or destroyed on finalize. But it's pointless to
keep it.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1206
(cherry picked from commit 94b3c334e5)
Most visible with xwayland-on-demand, at the time of setting things up
for X11 selections, we don't forward the current state. This makes the
first started X11 app oblivious to eg. the current clipboard.
Syncing selections up at the time of initializing the X11 selection
stuff ensures that doesn't happen.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1186
(cherry picked from commit 167fd07e01)
We already have a signal callback that translates selection ownership changes to
data_device/primary .selection events. Given both will be run when a data source
is being replaced, and this event emission being deleted is kinda short sighted
in that in only knows about Wayland, rely entirely on MetaSelection::owner-changed
emission.
Fixes spurious .selection(null) events being sent when a compositor-local source
takes over the selection without the focus changing (eg. screenshot to clipboard).
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/1160https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1204
(cherry picked from commit 1363246d44)
We are meant to send a .cancelled event after the drop is performed
in certain situations, but only for version>3 clients. Since this is
all version 3 business, only set the drop_performed flag for v3
clients. This drops the need to perform version checks at the time
of cancelling (which is present for other usecases in v1).
Fixes emission of wl_data_source.cancelled for v1 clients.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/1177https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1203
(cherry picked from commit d4c3870286)
For the cases where we read a fixed size from the selection (eg. imposing
limits for the clipboard manager), g_input_stream_read_bytes_async() might
not read up to this given size if the other side is spoonfeeding it content.
Cater for multiple read/write cycles here, until (maximum) transfer size is
reached.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1198
(cherry picked from commit 4bdf9a1e70)
Flushing the X11 selection output stream may happen synchronously or
implicitly, in which case there is not a task to complete. Check there
is actually a task before returning errors. We additionally set the
pipe_error flag, so future operations will fail with an error, albeit
with a more generic message.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1198
(cherry picked from commit 1909977a67)
If a write_async() comes up while we are flushing on the background,
the task will be queued, but not deemed a reason on itself to keep
flushing (and finish the task) after a property delete event.
To fix this, do not ever queue up write_async tasks (this leaves
priv->pending_task only used for flush(), so the "flush to end"
behavior in the background is consistent). We only start a
background flush if there's reasons to do it, but the tasks are
immediately finished.
All data will still be ensured to be transfered on flush/close,
this makes the caller in this situation still able to reach to it.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1198
(cherry picked from commit 655a783891)
It does not make sense to check for the stream not being closed,
this might happen multiple times during the lifetime of the stream
for a single transfer. We want to notify the INCR transfer just
once.
Check for the explicit conditions that we want, that the remaining
data is bigger than we can transfer at once, and that we are not
yet within the INCR transfer.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1198
(cherry picked from commit a4596becc4)
The stream automatically flushes after data size exceeds the
size we deem for INCR chunks, but we still try to copy it all.
Actually limit the data we copy, and leave the rest for future
INCR chunks.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1198
(cherry picked from commit 7015bb3efd)
INCR transfers are mandated to finish with a final 0-size XChangeProperty
roundtrip after the final data chunk. Actually honor this and ensure we
iterate just once more for this.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1198
(cherry picked from commit 04d429b743)
This seemed to work under the assumption that a flush() call can
only result in one INCR roundtrip. This is evidently not true, so
we should hold things off until all pending data is actually flushed.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1198
(cherry picked from commit 8a2b82897d)
If say we want 32bit data, but have 2 bytes stored, we would simply
ignore flush requests. Allow (and don't clear) the needs_flush flag
if we have less than the element size accumulated.
Instead handle this in can_flush(), so it's triggered whenever we
have enough data to fill 1 element, or if the stream is closing
(seems a broken situation, but triggered by the caller).
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1198
(cherry picked from commit 967966cdee)
XMaxRequestSize/XMaxExtendedRequestSize are documented to return
the maximum size in 4-byte units, whereas we are comparing this
to byte lenghts. We can afford 4x the data here.
Since I don't know the payload size of the XChangeProperty request,
be generous and allot 400 bytes for it, we have some to spare.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1198
(cherry picked from commit 06d67b6abf)
When closing the lid of a laptop, we reconfigure all the monitors in order
to update the CRTCs and (if enabled) the global UI scaling factor.
To do this, we try first to reuse the current configuration for the usable
monitors, but if we have only monitor enabled and this one is on the laptop
lid we just end up creating a new configuration where the primary monitor is
the laptop one (as per find_primary_monitor() in MetaMonitorConfigManager),
but ignoring the user parameters.
In case the user selected a different resolution / scaling compared to the
default one, while the laptop lid is closed we might change the monitors
layout, causing applications to rescale or reposition.
To avoid this, when creating the monitors configuration from the current
current state, in case we have only one monitor available and that one is
the laptop panel, let's just reuse this configuration.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1200
(cherry picked from commit e48516679c)
The CRTC level transform (not necessarily the hw transform) must be
taken into account when calculating the position of the CRTC in the
stage coordinate space, when placing the hw cursor, otherwise we'll
place the cursor as if the monitor was not rotated.
This wasn't a problem in the past, as with rotation, we always used the
OpenGL cursor, so the issue newer showed.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1199
The port to per CRTC views was incomplete; we still used the logical
monitor layout as the stage view layout, while still using one view per
CRTC.
This worked fine for most cases, e.g. regular monitors, tiled or
non-tiled, transformed or non-transformed. Where it broke, however, was
when a monitor consists of multiple CRTCs. We already have the layout a
CRTC corresponds to on the stage kept with the CRTC metadata, so use
this directly.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/1170https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1199`
The CRTC level transform (i.e. not necessarily the one set on the
hardware) is what is relevant for calculating the layout the CRTC will
have on the stage, so only use the one that can be handled by the
hardware for the CRTC assignment.
This makes the CRTC layout valid for tiled monitors.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1199
Previously the tile coordinate was used to offset a CRTC scanout
coordinate within a larger framebuffer. Since 3.36 we're always
scanning out from (0, 0) as we always have one framebuffer per CRTC; we
instead use the tile coordinate to calculate the coordinate the tile has
in the stage view. Adapt calculation to fulfil this promise instead of
the old one.
This also corrects the tiled custom monitor test case.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1199
Dereference the loop variable rather than the original list head. This
fixes a regression introduced in 4413b86a3 ("backends: Replace
ClutterDeviceManager usage in favor of ClutterSeat", 2019-10-04) which
broke button scrolling with trackballs.
Closes:https://gitlab.gnome.org/GNOME/mutter/-/issues/1120
(cherry picked from commit 3e967d731a)
On VT switch, the devices are removed, which means for Wayland disabling
the keyboard.
When the keyboard is disabled, the associated `xkb_state` is freed and
recreated whenever the keyboard is re-enabled when switching back to the
compositor VT.
That means the `xkb_state` for Wayland is lost whereas the same for
clutter is kept, which causes to a discrepancy with locked modifiers on
VT switch.
To avoid that issue, preserve the XKB info only to dispose it when the
keyboard is eventually finalized.
Closes: https://gitlab.gnome.org/GNOME/mutter/issues/344https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1185
(cherry picked from commit 5b30a52bbd)
The motion events of tablets for example need to be mapped on the
selected screen area if the input device is configured to use only a
part of the active logical monitor.
To achieve this behavior each motion event is transformed using the
transformation matrix set for the input device.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/1118
At some point we crossed the streams... In a short timespan we had
1f00aba92c merged, pushing WacomDevice to a common parent object,
and dcaa45fc0c implementing device grouping for X11.
The latter did not rely on the former, and just happened to
merge/compile without issues, but would promptly trigger a crash
whenever the API would be used.
Drop all traces of the WacomDevice internal to MetaInputDeviceX11.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1183
(cherry picked from commit f0718c7d95)
