This makes --import-format=NV12 testable on e.g. intel
We only set nv12_format_found to true if we found that format and at
least one understood modifier. Store modifier verbatim instead of using
a boolean flag. Last advertised and supported modifier currently wins.
The NV12 DRM_FORMAT_LINEAR image should be green in the upper left
corner and white in the lower right.
Signed-off-by: Guido Günther <agx@sigxcpu.org>
Reviewed-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Rather than segfaulting by attempting to traverse an initially
null log handler pointer, explicitly print a message and abort.
Signed-off-by: Matt Hoosier <matt.hoosier@gmail.com>
Reviewed-by: Ian Ray <ian.ray@ge.com>
[Pekka: coding style fix]
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Only exit from main so control flow is in one place.
Signed-off-by: Guido Günther <agx@sigxcpu.org>
Reviewed-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
The SURFACE_BITS_COMMAND struct has changed and some members have been moved in the
bmp field.
Reviewed-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Tested-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Very useful for TRANSIENT_FOR property debugging.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Ian Ray <ian.ray@ge.com>
Add code to dump CARDINAL arrays from properties. Useful for debugging.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Ian Ray <ian.ray@ge.com>
Running 'weston-simple-egl -f -b' (fullscreen, unthrottled) caused a
crash in shell_ensure_fullscreen_black_view() due to
shsurf->fullscreen_output being NULL. Also shell_configure_fullscreen()
could crash on that condition.
Fix shell_configure_fullscreen() to bail out with minimal work if there
is no fullscreen_output.
It is unclear if anything will cause a reconfiguration when an output is
plugged in.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Ian Ray <ian.ray@ge.com>
In activate, do not call lower_fullscreen_layer() at all if the output
is NULL. It should not do anything in that case, per the existing
comment.
This is a tentative crash fix for a case where there are no enabled
weston_outputs at all.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Ian Ray <ian.ray@ge.com>
This is a tentative crash fix for a case where there are no enabled
weston_outputs at all.
If no output is given, just put the surface at 0,0. At least it should
become mostly visible if an output is plugged in, if not centered.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Ian Ray <ian.ray@ge.com>
This is a tentative crash fix for a case where there are no
enabled weston_outputs at all.
Let get_output_work_area() return a zero area if the given output is
NULL. If there is no output, there is no area. Unfortunately we cannot
return "no position" but have to use 0,0 instead.
In send_configure_for_surface(), this causes a maximized surface to
receive width=0 and height=0 in the configure event, which means the
client is free to choose the size. There is no correct size to send for
maximizing for no output.
In constrain_position(), this has no effect. The interactive move of a
surface is restricted to not go below the panel, so even if a user
managed to move a surface without an output, it just prevents the
surface moving beyond y=0.
In weston_view_set_initial_position(), get_output_work_area() will not
be called with NULL output anyway.
In set_maximized_position(), this makes it behave as if the output was
at 0,0 which is the default position of the first output.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Ian Ray <ian.ray@ge.com>
Allows to control the Pixman-renderer shadow framebuffer usage from
weston.ini. It defaults to enabled, and whether it is a good idea to
disable or not depends on the platform and the workload.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Ian Ray <ian.ray@ge.com>
Allow global control of the pixman shadow buffers. The compositor can
choose whether all output use or do not use a shadow buffer with the
pixman renderer.
The option is added to the end of struct weston_drm_backend_config to
avoid bumping WESTON_DRM_BACKEND_CONFIG_VERSION.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Ian Ray <ian.ray@ge.com>
Add a flag to pixman-renderer for initializing the output with a shadow
framebuffer. All backends were getting the shadow implcitly, so all
backends are modified to ask for the shadow explicitly.
Using a shadow buffer is usually beneficial, because read-modify-write
cycles (blending) into a scanout-capable buffer may be very slow. The
scanout framebuffer may also have reduced color depth, making blending
and read-back produce inferior results.
In some use cases though the shadow buffer might be just an extra copy
hurting more than it helps. Whether it helps or hurts depends on the
platform and the workload. Therefore let the backends control whether
pixman-renderer uses a shadow buffer for an output or not.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Ian Ray <ian.ray@ge.com>
Pixman-renderer uses a single internal shadow buffer. It is enough to
composite the current damage into shadow, but the copy to hw buffer
needs to include the previous damage because of double-buffering in
DRM-backend.
This patch lets pixman-renderer do exactly that without compositing also
the previous damage on DRM-renderer.
Arguably weston_output should not have field previous_damage to begin
with, because it implies double-buffering, which e.g. EGL does not
guarantee. It would be better for each backend explicitly always provide
any extra damage that should be copied to hw.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Ian Ray <ian.ray@ge.com>
This issue was likely introduced by "libweston: add
weston_view_set_output()" which forgot to ensure the output destroy
listener is removed when weston_view is destroyed, leading to freed
memory being left into the list.
This was quite easy to trigger by opening and closing an application
window a few times, leading various memory corruption symptoms.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Ian Ray <ian.ray@ge.com>
This issue was introduced by "desktop-shell: detect stale shell surface
outputs" which forgot to remove the output destroy listener when
shell_surface is destroyed, leading to memory corruption.
This was fairly easy to trigger by opening and closing an application
window a few times.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Ian Ray <ian.ray@ge.com>
When displays are hot (un)plugged, it may happen that
a shell surface is left with a stale pointer to an output
that has already been freed. Add an output destroy listener
to catch such situations and set the output pointer to NULL.
Signed-off-by: Semi Malinen <semi.malinen@ge.com>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Instead of desktop shell assigning view outputs directly,
use a new method, weston_view_set_output(). The method can
set up an output destroy listener to make sure that views
do not have stale output pointers.
Without this patch it is possible to end up in a scenario
where, e.g. configure_static_view() accesses memory that
has already been freed. The scenario can be provoked by
repeatedly plugging and unplugging a display. The faulty
memory accesses are reported by valgrind.
Signed-off-by: Semi Malinen <semi.malinen@ge.com>
Signed-off-by: Fabien Lahoudere <fabien.lahoudere@collabora.com>
Reviewed-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Rename user_data to wet, because it is called wet everywhere else.
Drop the local variable ec, because that is available as wet.compositor.
This models a little better that wet_compositor owns weston_compositor,
and not the other way around.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
This makes it easier to just pass wet_compositor around and take the
weston_compositor from it.
It feels weird to go from weston_compositor to wet_compositor all the
time in internal functions. It's necessary in callbacks that cannot
carry wet_compositor, but otherwise it is awkward.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Allow cloning up to 4 connectors from the same CRTC. All the
implementation bits support more than one head per output already.
Four is just an arbitary number, small but unlikely to ever be the
limiting factor in cloning since hardware is usually very restricted.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
For the attach on an enabled output to have an effect, we need to go
through drmModeSetCrtc or ATOMIC_ALLOW_MODESET.
v9:
- Add another XXX comment.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
When a head is detached from an enabled output, that output needs to go
through a modeset (drmModeSetCrtc() / ATOMIC_ALLOW_MODESET) so that the
connector is actually removed from the CRTC.
This has not yet been a problem, because an output could only have one
head at a time, and would be automatically disabled on detach. It would
be a problem with clone mode.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
If we are processing a connector that does not have an existing routing,
it is possible we pick a CRTC that was previously routed to a connector
we have not enabled yet. If that happens, the latter connector cannot
preserve its routing.
Check that no other connector we might enable later had this CRTC
before.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
To support shared-CRTC clone mode, the chosen CRTC needs to support
driving all the attached connectors. Replace the old algorithm with a
new one that takes into account all associated connectors.
Ideally it should use possible_clones mask to check which encoders (and
therefore connectors) actually can be in a cloned set. However, the DRM
documentation says about possible_clones and possible_crtcs masks both:
"In reality almost every driver gets this wrong."
- https://01.org/linuxgraphics/gfx-docs/drm/gpu/drm-kms.html#c.drm_encoder
Looking at a target device and its kernel where clone mode is desired,
possible_clones is indeed self-conflicting and would not allow cloning
at all. Therefore the implemented algorithm replaces the checking of
possible_clones with luck. It even goes out of its way to find any CRTC
for a configuration, even if not advertised by the kernel as not
supported.
Libweston would need infrastructure to allow trial-and-error CRTC
allocation: rather than picking one CRTC in advance and do or die, it
should try all available CRTCs one by one. Unfortunately that is not yet
possible, so this patch implements what it can. It is also the DRM
upstream opinion that trial-and-error with ATOMIC_TEST would be the way
to go.
Unlike the old algorithm, the new algorithm prefers routings that were
in place when Weston started instead of when enabling an output. When
you never temporarily disable an output, this makes no difference.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
compositor.c has 'touch', so use 'touch' here as well. It is not a
device to begin with.
Signed-off-by: Louis-Francis Ratté-Boulianne <lfrb@collabora.com>
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
v1 Tested-by: Matt Hoosier <matt.hoosier@gmail.com>
We have weston_seat_{init,release}_{pointer,keyboard,touch}() as the
backend-facing API. There is no need to expose the create/destroy
functions which have been for internal use only for quite a long time.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
v1 Tested-by: Matt Hoosier <matt.hoosier@gmail.com>
Add test_seat_release() as the counterpart of test_seat_init() instead
of open-coding it. This helps adding more code to test_seat_release()
later.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
Commit c81c4241d9 added this environment
variable. Document it.
It applies also to the fbdev-backend but that has no man page.
v2:
- Rewording by Peter Hutterer.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
If an input device is associated to an output that then gets disabled,
there is no case where associating to a different output would be
correct.
The output association is used for absolute positioned input devices,
and an input device like a touchscreen cannot ever be automatically
valid for more than one possible output - the touchscreen display
device.
Therefore do not automatically reassing implicitly associated input
devices to another output. This removes some log spam on shutdown.
In fact, if there can be more than one output at any time, absolute
input devices must be explicitly configured to associate with the
correct output, or the results are essentially undefined in any case.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
The use case driving this change is a clone mode setup, where the user
is hotplugging or unplugging a cloned touchscreen. Even if the output
and head are force-enabled, the touch device should still follow the
connector connection status. If there is no video signal for the
touchscreen (disconnected connector), then the touch input should be
ignored as well.
When the output is force-enabled, we need to trigger
output_heads_changed from connector status changes. If the head or
output are not force-enabled, the compositor will likely attach and
detach the head as appropriate. In clone mode, the attach or detach
needs to trigger output_heads_changed directly. In other cases, it may
be handled through the output getting enabled or disabled which are
different signals.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
Associating input devices with weston_outputs by the output name fails
when one output has several heads. We need to match against head names
instead of output names to be able to find all names.
This fixes touchscreen output association in shared-CRTC clone mode when
outputs or input devices appear or disappear.
Even though notify_output_create() is called only when new outputs
appear, the implementation is prepared to also remove output
associations. This will be handy in the future when this function will
handle also head detaching from an output that remains enabled.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
Removing the output association from an evdev_device is more than just
setting the pointer to NULL, one also needs to remove the destroy
listener and flag the destroy listener as unused (notify == NULL).
evdev_device_set_output() can already remove associations, so let it
also handle an assignment to NULL output.
Fix notify_output_destroy() to handle removing an association correctly.
Previously, the listener was left "used", which would mean the next call
to evdev_device_set_output() would have wl_list_remove()'d, accessing
freed memory. This could be triggered by having a touchscreen with a
specified output association, and unplugging then re-plugging the
corresponding output.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
In the future evdev_device_set_output() will start getting called more
often, redundantly. Short-circuit the setting if the chosen output is
already set for an input device. This reduces churn in the logs.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
Helps admins ensure the configuration is correct.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
The udev property WL_CALIBRATION is an old way of giving Weston a
touchscreen calibration matrix. It is Weston-specific.
The recommended way of setting up a calibration is to use the udev
property LIBINPUT_CALIBRATION_MATRIX, which libinput will load
automatically and therefore applies to all libinput using display
servers and applications.
The syntax of WL_CALIBRATION and LIBINPUT_CALIBRATION_MATRIX is
different as well: WL_CALIBRATION uses pixels as the translation part
units, which makes the values depend on the output resolution.
LIBINPUT_CALIBRATION_MATRIX on the other hand uses normalized units.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
Print a note that calibration got skipped if the input device supports a
calibration matrix but there is no associated output to compute it from.
Helps with debugging touchscreen calibration issues.
The code is reorganized and commented a bit, but this does not change
the behaviour.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
Struct 'evdev_device' has field 'devnode' which is initialized to NULL,
never assigned, and finally free()'d. Therefore it is useless.
Remove the dead field.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Peter Hutterer <peter.hutterer@who-t.net>
The head was just zalloc()'d, there is no need to memset it to zero.
If a function fails, it is preferable it leaves the output arguments
untouched.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Rename connector_get_current_mode() because it will be useful for
storing not just the current mode on creating a head.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Stop using a head for printing the mode list, because there could be
multiple heads. We already gather the mode list from all heads.
No need to print the connector id here, because it is logged with DRM
heads, and core prints the head names on output enable.
The "built-in" flag seemed dead, because it could only be printed if the
kernel provided no modes. If we want more detailed info on where modes
come from, we would need to inspect mode_info or add new flags to
drm_mode or weston_mode.
Add printing the pixel clock, because that is used by the video mode
duplicate removal code.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
If an output has multiple (cloned) heads, it should be enough for any
head to support backlight control for DRM-backend to expose it.
Inspect all attached heads for backlight control and improve the
logging.
Pick the initial backlight level from whatever happens to be the "first"
head, because it's simple.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
A single list of modes needs to be combined from the mode lists in each
attached head. We could just concatenate the lists, but that might
introduce duplicates. Try to avoid duplicates instead by using partially
fuzzy matching.
When a duplicate is found, try to figure out which is more suitable to
use in place of both. If one has the preferred flag and the other
doesn't, take the preferred one. Otherwise use the one already in the
list.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Previously the log contained one line for EDID data and another line for
the head, and you just had to know they belong together. Make it more
obvious to read by putting both head and EDID info on the same line.
We no longer print EDID data every time it is parsed (on every hotplug
event), but only if it changes. I did take a shortcut here and use
weston_head::device_changed as the print condition which relies on the
compositor clearing it, but a failure to do so just means we print stuff
even if it didn't change.
Head info updates also print the head info and not just the EDID data.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Fix this function to support more than one head per output.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Acked-by: Derek Foreman <derekf@osg.samsung.com>
Guido Günther
Matt Hoosier
David Fort
Pekka Paalanen
Semi Malinen
Louis-Francis Ratté-Boulianne