Memleak found by ASAN:
Direct leak of 21 byte(s) in 1 object(s) allocated from:
#0 0x7fe7a917fe8f in malloc (/usr/lib/x86_64-linux-gnu/libasan.so.6+0xa9e8f)
#1 0x7fe7a9129874 (/usr/lib/x86_64-linux-gnu/libasan.so.6+0x53874)
#2 0x7fe7a5a23469 in weston_wm_window_read_properties ../xwayland/window-manager.c:574
#3 0x7fe7a5a28d3b in weston_wm_handle_map_request ../xwayland/window-manager.c:1178
#4 0x7fe7a5a31660 in weston_wm_handle_event ../xwayland/window-manager.c:2291
#5 0x7fe7a8c261a1 in wl_event_loop_dispatch ../src/event-loop.c:1027
Signed-off-by: Marius Vlad <marius.vlad@collabora.com>
Memleak found by ASAN:
Direct leak of 40 byte(s) in 1 object(s) allocated from:
#0 0x7fe7a917fe8f in malloc (/usr/lib/x86_64-linux-gnu/libasan.so.6+0xa9e8f)
#1 0x7fe7a5a40736 in theme_create ../shared/cairo-util.c:419
#2 0x7fe7a5a3363c in weston_wm_create ../xwayland/window-manager.c:2619
#3 0x7fe7a5a2017e in weston_xwayland_xserver_loaded ../xwayland/launcher.c:313
#4 0x7fe7a90b4d14 in handle_sigusr1 ../compositor/xwayland.c:57
#5 0x7fe7a8c2585d in wl_event_source_signal_dispatch ../src/event-loop.c:685
#6 0x7ffcdb04ef6f ([stack]+0x1df6f)
Signed-off-by: Marius Vlad <marius.vlad@collabora.com>
There is more state than just the application window width and height that
affects whether calling weston_wm_window_configure() is necessary: everything
that affects the frame window, fullscreen state in particular. Therefore do not
skip the call by just width and height.
If send_configure() happens to be called "unnecessarily", this will now forward
some of those calls to the X11 clients. However, since it uses an idle task, it
should not result in a flood at least. And if send_configure() is spammed,
maybe that should then be fixed in its callers.
This patch should fix the misplacement of a fullscreen X11 window due to the
frame window being incorrectly sized and positioned, and the app window
incorrectly positioned inside the frame window.
The fullscreen window problems were observed in a case where the window does
not hit legacy_fullscreen() but first maps and then sets
_NET_WM_STATE_FULLSCREEN. Additionally the initial window size must match the
output size where it gets fullscreened. In that case the frame window was left
as if not fullscreened.
This practically reverts 3f53d9179b. I'm not sure
what problem that patch was fixing, but I couldn't make any resizing freeze.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>
Spotted this in debug log:
[xwm-wm-x11] XWM: configure window 4194324: x=32 y=32 width=1920 height=1080 border_width=0 stack_mode=0
[xwm-wm-x11] XWM: configure window 0: width=1984 height=1144
Trying to configure window 0 makes no sense. So do not try.
To avoid patching two different places with the same thing, refactor the code
into a common helper.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>
It would lead to use-after-free if there was a pending idle callback to
weston_wm_window_configure() when the weston_wm_window gets destroyed. Make
sure the callback will not fire.
Found by inspection.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>
This function is called also directly from weston_wm_window_set_toplevel(). If
configure_source is set at that point, simply resetting the pointer will "leak"
the source until it fires and calls this function again.
Let's keep the variable up-to-date by removing the source when called,
dispatched or not. This removes the second call. I only hope it doesn't cause
issues. This is also necessary if we intend to remove the source on window
destruction too.
Found by inspection.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>
It looks like commit ad0da4596d introduced a bug
for X11 windows that are initially fullscreen by adding code to the end of
xserver_map_shell_surface() while ignoring the 'return' that this patch
removes. That may have caused some annoying window state issues, but the
problem became more pronounced with 7ace831ca6
when used with an Xwayland version that honours _XWAYLAND_ALLOW_COMMITS.
In the latter case, there is a possiblity the window will never show up, as XWM
forgets to set allow_commits=true. However, the window may sometimes actually
show up due to an oversight in Xwayland: the Present code may be flipping the
window buffers and not checking _XWAYLAND_ALLOW_COMMITS if it is supposed
commit at all.
Since then, f568968f8a added more places where
allow_commits is set to true, masking the window-does-not-show-up issue. Window
pending state likely still remained an issue.
This patch properly fixes the "window never appears" issue by making sure
allow_commit=true is set. At the same time, it ensures the pending state
functions are called at the end of xserver_map_shell_surface(), which may fix
some window state issues like misplaced decorations and/or position of
initially-fullscreen windows. Unfortunately, it certainly does not fix all such
state problems.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>
The mask argument is uint16_t so declare the variable with the same.
Suggested-by: Daniel Stone <daniels@collabora.com>
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>
Knowing the kind of decoration drawn will help track down issues with
unexpected decorations.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>
Print the changes to the debug scope, helping to figure out why Xwayland is or
is not committing.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>
Add the missing newline to printing a property that is of type cardinal array.
Fixes messed up debug scope output.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>
Initially, `_XWAYLAND_ALLOW_COMMITS` was introduced in commit 7ace831ca
to avoid drawing the window content before it's ready to be shown.
But a repaint might also be triggered by the client damages before the
XWM has finished drawing its window decorations and drop shadows, which
previously was not too much of an issue since the XWM could still
finish updating the X11 window after the buffer was submitted.
However, with the addition of multiple window buffers in Xwayland [1]
which are aimed at preventing the X11 clients from updating the buffer
after it's been committed, this is no longer possible.
As a result, the use of multiple window buffers in Xwayland can cause
ugly repainting effects of the decorations if the buffer is submitted
before the XWM has finished painting its decorations.
Use the X11 property `_XWAYLAND_ALLOW_COMMITS` can be used to avoid
this, by controlling when Xwayland should commit changes to the Wayland
surface.
[1] https://gitlab.freedesktop.org/xorg/xserver/merge_requests/316
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
When all shared/ headers are included in the same way, we can drop unnecessary
include seach paths from the compiler.
This include style was chosen because it is prevalent in the code base. Doing
anything different would have been a bigger patch.
This also means that we need to keep the project root directory in the include
search path, which means that one could accidentally include private headers
with
#include "libweston/dbus.h"
or even
#include <libweston/dbus.h>
IMO such problem is smaller than the churn caused by any of the alternatives,
and we should be able to catch those in review. We might even be able to catch
those with grep in CI if necessary.
The "bad" include style was found with:
$ for h in shared/*.h; do git grep -F $(basename $h); done | grep -vF '"shared/'
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>
The main idea is to make libweston users use the form
#include <libweston/libweston.h>
instead of the plain
#include <compositor.h>
which is prone to name conflicts. This is reflected both in the installed
files, and the internal header search paths so that Weston would use the exact
same form as an external project using libweston would.
The public headers are moved under a new top-level directory include/ to make
them clearly stand out as special (public API).
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>
This crash was happening *during resizing* of an xwayland window that was destroyed.
Discovered by: John Good @archiesix
[@daniels: Moved tests below declarations.]
Instead of a compile time choice, offer the XWM debugging messages
through the weston-debug protocol and tool on demand. Users will not
need to recompile weston to get XWM debugging, and it won't flood the
weston log.
The debug scope needs to be initialized in launcher.c for it be
available from start, before the first X11 client tries to connect and
initializes XWM.
Signed-off-by: Pekka Paalanen <pq@iki.fi>
pass the wm_debug scope to weston_debug_scope_printf API to append
the scopename to the timestr
Signed-off-by: Maniraj Devadoss <Maniraj.Devadoss@in.bosch.com>
Reviewed-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
This is preparation for using the weston-debug infrastructure for
WM_DEBUG. dump_property() may be called from different debugging
contexts and often needs to be prefixed with more information.
An alternative to this patch would be to pass in the weston_debug_scope
as an argument to dump_property(), but then all callers would need to be
converted to weston-debug infra in a single commit.
Therefore require the callers to provide the FILE* to print to.
Signed-off-by: Pekka Paalanen <pq@iki.fi>
Signed-off-by: Maniraj Devadoss <Maniraj.Devadoss@in.bosch.com>
Reviewed-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Write the output of dump_property() out in one log call. When multiple
processes (weston and Xwayland) are writing to the same file, this will
keep the property dump uninterrupted by Xwayland debug prints.
This is also preparation for more development in the same direction.
Signed-off-by: Pekka Paalanen <pq@iki.fi>
Signed-off-by: Maniraj Devadoss <Maniraj.Devadoss@in.bosch.com>
Reviewed-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
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>
This (partially) reverts commit bef761796c.
This (partially) reverts commit 4d1cd36c9e.
This (partially) reverts commit 44fc1be913.
This (partially) reverts commit 6b58ea8c43.
The new xwm icon code has proven to be leaky and incomplete, and while
we have patches under consideration to fix the rest of its known problems
they still require changes and review cycles. Currently the known
leaks have been squashed, but it still picks wrong sized icons and
does no scaling, which can lead to very strange rendering. At window
close time the wrong sized icon appears above the window during fade out.
This patch reverts the mostly solid bits and keeps the unfinished
bits behind in favor of a simpler revert than removing the whole
thing.
Signed-off-by: Derek Foreman <derekf@osg.samsung.com>
Acked-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Quentin Glidic <sardemff7+git@sardemff7.net>
This reverts commit 332d1892bb.
And re-introduces the bug it was intended to fix, see:
https://lists.freedesktop.org/archives/wayland-devel/2017-December/036402.html
Reverting this because it causes harm to all xwayland clients - the
input region no longer gets adjusted when resizing windows.
start an xterm, resize it larger, you can no longer interact with the
new area of the window (including the server side decor).
Hopefully sort the last leaks introduced in commit 6b58ea8c
The window could be destroyed before it had a frame but after it had an icon
(I could trigger this with firefox), and the window could be assigned an icon
twice before it had a frame (I could trigger this with terminology).
The latter leak was
Reported-by: Scott Moreau <oreaus@gmail.com>
Signed-off-by: Derek Foreman <derekf@osg.samsung.com>
Reviewed-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
A memory leak introduced by 6b58ea8c led to me finding a bigger leak,
which is xwm was calling frame_create() without calling frame_destroy().
This meant that the associated icon_surface was not being destroyed,
leaving the destroy handler for it broken. Here we fix this by calling
frame_destroy() when the window is destroyed and free the reply in
the icon_surface destroy handler.
Reviewed-by: Derek Foreman <derekf@osg.samsung.com>
Acked-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
This reverts commit d2cb711d81.
I missed the call to cairo_image_surface_create_for_data() which assumes
the data will remain present until the cairo surface is destroyed. It
seems the existence of data depends on the reply not being freed.
This will need a more involved fix.
Sorry, I noticed this just seconds after I pushed the patch.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Fix memory leak introduced by 6b58ea8c. weston_wm_handle_icon() was
calling xcb_get_property_reply() without freeing the reply.
Reviewed-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
The cairo surface used for the icon must be completely given to the
frame as soon as said frame has been created. To prevent both the
window and the frame from sharing ownership of the icon, we set
window->icon_surface back to NULL right after creating or changing the
frame, only keeping it there when no frame has been created yet.
Fixes https://lists.freedesktop.org/archives/wayland-devel/2018-January/036655.html
Reported-by: Derek Foreman <derekf@osg.samsung.com>
Tested-by: Derek Foreman <derekf@osg.samsung.com>
Signed-off-by: Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
This fetches the _NET_WM_ICON property of the X11 window, and use the
first image found as the frame icon.
This has been tested with various X11 programs, and improves usability
and user-friendliness a bit.
Changes since v1:
- Changed frame_button_create() to use
frame_button_create_from_surface() internally.
- Removed a check that should never have been commited.
Changes since v2:
- Request UINT32_MAX items instead of 2048, to avoid cutting valid
icons.
- Strengthen checks against malformed input.
- Handle XCB_PROPERTY_DELETE to remove the icon.
- Schedule a repaint if the icon changed.
Changes since v3:
- Keep the previous Cairo surface until the new one has been
successfully loaded.
- Use uint32_t for cardinals. Unsigned is the same type except on
16-bit machines, but uint32_t is clearer.
- Declare length as uint32_t too, like in xcb_get_property_reply_t.
Signed-off-by: Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
Reviewed-by: Quentin Glidic <sardemff7+git@sardemff7.net>
The window frame was created with position and size which include
an offset for margins and shadow. Set the input region to ignore
shadow.
[daniels: Fixed type mismatch, removed unused variable.]
Signed-off-by: Ian Ray <ian.ray@ge.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Tested-by: Scott Moreau <oreaus@gmail.com>
If we configure a window with the same size and wait for the
sync alarm to go off, the resizing is gonna block. The event is
only handled is the size actually changed.
Signed-off-by: Louis-Francis Ratté-Boulianne <lfrb@collabora.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
A single client message can be used to modify two properties at once.
That's why when processing such messages we have to check both the second
and the third data entry for states that we must handle.
Signed-off-by: Ilia Bozhinov <iliyabo@hotmail.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
When we receive configure_notify we should update the surface's
position by calling xwayland_api->set_xwayland(). Otherwise some surfaces
like dnd surfaces from xwayland views are "stuck" at one place. When
setting XWAYLAND state though we should always call view_set_position(),
not just the first time we set this state.
Signed-off-by: Ilia Bozhinov <ammen99@gmail.com>
Reviewed-by: Quentin Glidic <sardemff7+git@sardemff7.net>
xwm would not let X clients change the focus behind the compositor's
back, reverting focus where it's supposed to be when this occurs.
However, X11 grab also issue focus in events, on which some clients
rely and reverting focus in this case braks the client logic (e.g.
combobox menu in gtk+ using the X11 backend).
Check if the focus event is actually coming from a grab or ungrab and
do not revert focus in this case, to avoid breaking clients logic.
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Tested-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Acked-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Acked-by: Daniel Stone <daniels@collabora.com>
Acked-by: Quentin Glidic <sardemff7+git@sardemff7.net>
This patch uses the new feature proposed for Xwayland in the patch
series https://patchwork.freedesktop.org/series/16610/ .
When the frame window is created, immediately forbid Xwayland commits on
it. This prevents commits before the decorations have been drawn and the
initial pending state has been set. Commits are enabled right after
drawing and setting.
This ensures that the decorations are fully drawn when a window is
mapped. This also solves the initial commit/pending race, but the race
is on again after mapping.
If Xwayland does not implement the needed support, we are just setting a
window property with no effect.
This patch is the final piece for solving T7622, excluding the
_NET_WM_SYNC_REQUEST handling.
Task: https://phabricator.freedesktop.org/T7622
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Louis-Francis Ratté-Boulianne <lfrb@collabora.com>
Acked-by: Daniel Stone <daniels@collabora.com>
Normal windows enter the MapRequest handler, which schedules drawing the
decorations. Then Xwayland realizes the window, which ends with a call
to xserver_map_shell_surface(). The decorations are already drawn, no
need to draw them a second time. However, MapRequest handler could not
set the pending state because the weston_surface did not exist at the
time, because it gets created only when Xwayland realizes the window,
which happens after XWM has forwarded the MapWindow in MapRequest
handler. Therefore set the pending state explicitly at the end.
Scheduling had it done much later anyway.
Now that the pending state is set much earlier, it seems to be more
likely that it gets set before Xwayland's first commit is handled. This
means that -geometry command line option of X11 apps already takes the
geometry (decorations) into account. I do not think it is reliable yet,
though.
There is still the race between Xwayland committing and XWM setting the
pending state assuming the very next commit latches it in appropriately.
The race exists not because of Wayland, but because WL_SURFACE_ID comes
via X11, and could be processed after wl_compositor.create_surface and
wl_surface.commit. That commit/pending race is solved by a following patch.
For override-redirect windows weston_wm_window_schedule_repaint()
reduced into a call to weston_wm_window_set_pending_state_OR(), so we
can just call that directly. It should not matter that the call is moved
to the end of the function.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Louis-Francis Ratté-Boulianne <lfrb@collabora.com>
Acked-by: Daniel Stone <daniels@collabora.com>
To me it was not obvious that this call is necessary, so provide some
rationale.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
When we as the WM tell the X server to map a window, it gets mapped. We
can start drawing into it immediately. There is no reason to wait for
any other events before drawing the decorations.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
This comes via Wayland, WL_SURFACE_ID comes via X11. They race. Nice to
get both printed.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Having it in a separate function makes it more clear what it is, and
allows it to be called from elsewhere.
This really is the set_pending_state() alternative for override-redirect
windows, because OR windows do not get a frame window created. Also OR
windows will never hit the normal set_pending_state() because
weston_wm_window_schedule_repaint() special-cased windows without a
frame window and returned early without scheduling.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Move the region fini just above the region init. They are a pair and
belong togeether. Split a long line.
Reads better this way. No functional changes.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Split the function into two:
- weston_wm_window_draw_decoration() that only draws the decorations
with Cairo, and
- weston_wm_window_set_pending_state() which sets up the surface state
to be latches into use on the next commit from Xwayland.
The new weston_wm_window_do_repaint() is the equivalent of the old
weston_wm_window_draw_decorations(), everything still happens the same
way as it was. Just some debug messages have been reworded.
weston_wm_window_read_properties() is moved into
weston_wm_window_do_repaint() because it is not strictly a part of
drawing decorations. The same with resetting repaint_source.
draw_decorations does not need the child position nor xwayland
interface. Also some convenience variables have been eliminated.
set_pending_state code has been un-indented by one level, so the change
is best viewed with whitespace changes ignored.
This patch makes the code more readable, and prepares for calling the
draw_decorations and set_pending_state from different places.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
X11 applications expect -geometry command line option to work for
setting the initial window position, but currently this does not work.
During map, detect X11 windows that set an explicit position. This works
by heuristics: if window position is not 0,0 then it is explicitly
positioned. Legacy fullscreen windows are also at 0,0 but these are
detected earlier.
Explicitly store the window position at map request time to detect
client-positioned windows, and use it as the suggested initial position.
weston_wm_window::x and y have been overwritten due to reparenting when
we eventually need the initial position.
This patch requires that the new set_toplevel_with_position() hook is
implemented in the shell.
Note that this patch is about positioning xwayland toplevels, not
override-redirect windows which are already handled.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Quentin Glidic <sardemff7+git@sardemff7.net>
Reviewed-by: Daniel Stone <daniels@collabora.com>
To reproduce the problem:
- start weston (x11 backend worked, glamor in Xwayland makes no
difference)
- start xterm
- very very slowly move the pointer in the xterm decorations onto or
away from a button
- the moment the decorations are updated, they will appear incomplete,
e.g. completely without buttons and title text
- if you cause just one more pointer motion event, the decorations will
update to completely drawn appearance
Another way to reproduce the problem is to have an xterm and change its
window title. This is easy if you use a shell prompt that updates the
terminal window title. When the title updates, decorations will be
half-drawn until something happens in XWM.
The fix: flush.
Apparently the drawing commands did not get flushed to the X server
until any other X11 action pushed them through.
xcb_flush() is the real fix here. cairo_surface_flush() is added just
for good measure, because documentation indicates it would be better
used, however it was not strictly necessary to fix the problem in my
experiments.
Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.co.uk>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Tested-by: Emmanuel Gil Peyrot <emmanuel.peyrot@collabora.com>
Reviewed-by: Emmanuel Gil Peyrot <emmanuel.peyrot@collabora.com>
Emmanuel Gil Peyrot
John Good
Marius Vlad
Jimmy Ohn
Pekka Paalanen
Link Mauve
Olivier Fourdan
Greg V
Pekka Paalanen
Pekka Paalanen
Derek Foreman
Scott Moreau
Ian Ray
Louis-Francis Ratté-Boulianne
Ilia Bozhinov
Ilia Bozhinov