Hardware planes are going to make E_Comp_Wl_Buffer lifetimes harder to
manage, so we need to let the E_Comp_Wl_Buffer object outlive the
resource attached to it.
We already track a busy count, so we just have to use it to prevent
deleting a busy buffer.
the frame deletion callback would trigger far too late, allowing
crashes to occur if the selection target client exited between the
time of the drop event and the selection send event
==13853== Invalid read of size 8
==13853== at 0x5C7C56: _e_comp_wl_surface_destroy (e_comp_wl.c:1804)
==13853== by 0xA999971: destroy_resource (wayland-server.c:611)
==13853== by 0xA9A06F4: for_each_helper (wayland-util.c:374)
==13853== by 0xA9A073F: wl_map_for_each (wayland-util.c:387)
==13853== by 0xA999C87: wl_client_destroy (wayland-server.c:763)
==13853== by 0xA999216: wl_client_connection_data (wayland-server.c:283)
==13853== by 0xA99C2B0: wl_event_source_fd_dispatch (event-loop.c:90)
==13853== by 0xA99CC11: wl_event_loop_dispatch (event-loop.c:423)
==13853== by 0xA787AC0: _cb_create_data (ecore_wl2_display.c:272)
==13853== by 0xDBE984D: _ecore_call_fd_cb (ecore_private.h:333)
==13853== by 0xDBEC01B: _ecore_main_fd_handlers_call (ecore_main.c:1992)
==13853== by 0xDBEC8A9: _ecore_main_loop_iterate_internal (ecore_main.c:2379)
==13853== by 0xDBEA672: ecore_main_loop_begin (ecore_main.c:1292)
==13853== by 0x441DA9: main (e_main.c:1089)
==13853== Address 0x30ba5d90 is 176 bytes inside a block of size 1,424 free'd
==13853== at 0x4C2ED4A: free (vg_replace_malloc.c:530)
==13853== by 0x4603D6: _e_client_free (e_client.c:588)
==13853== by 0x5475A8: e_object_free (e_object.c:119)
==13853== by 0x5477C4: e_object_unref (e_object.c:152)
==13853== by 0x5473D7: e_object_del (e_object.c:60)
==13853== by 0x5C7C51: _e_comp_wl_surface_destroy (e_comp_wl.c:1803)
==13853== by 0xA999971: destroy_resource (wayland-server.c:611)
==13853== by 0xA9A06F4: for_each_helper (wayland-util.c:374)
==13853== by 0xA9A073F: wl_map_for_each (wayland-util.c:387)
==13853== by 0xA999C87: wl_client_destroy (wayland-server.c:763)
==13853== by 0xA999216: wl_client_connection_data (wayland-server.c:283)
==13853== by 0xA99C2B0: wl_event_source_fd_dispatch (event-loop.c:90)
==13853== by 0xA99CC11: wl_event_loop_dispatch (event-loop.c:423)
==13853== by 0xA787AC0: _cb_create_data (ecore_wl2_display.c:272)
==13853== by 0xDBE984D: _ecore_call_fd_cb (ecore_private.h:333)
==13853== by 0xDBEC01B: _ecore_main_fd_handlers_call (ecore_main.c:1992)
==13853== by 0xDBEC8A9: _ecore_main_loop_iterate_internal (ecore_main.c:2379)
==13853== by 0xDBEA672: ecore_main_loop_begin (ecore_main.c:1292)
==13853== by 0x441DA9: main (e_main.c:1089)
==13853== Block was alloc'd at
==13853== at 0x4C2FA50: calloc (vg_replace_malloc.c:711)
==13853== by 0x5471A4: e_object_alloc (e_object.c:20)
==13853== by 0x467AD5: e_client_new (e_client.c:2596)
==13853== by 0x5C7F11: _e_comp_wl_compositor_cb_surface_create (e_comp_wl.c:1858)
==13853== by 0xADBDC57: ffi_call_unix64 (in /usr/lib64/libffi.so.6.0.2)
==13853== by 0xADBD6B9: ffi_call (in /usr/lib64/libffi.so.6.0.2)
==13853== by 0xA99EEED: wl_closure_invoke (connection.c:935)
==13853== by 0xA999581: wl_client_connection_data (wayland-server.c:371)
==13853== by 0xA99C2B0: wl_event_source_fd_dispatch (event-loop.c:90)
==13853== by 0xA99CC11: wl_event_loop_dispatch (event-loop.c:423)
==13853== by 0xA787AC0: _cb_create_data (ecore_wl2_display.c:272)
==13853== by 0xDBE984D: _ecore_call_fd_cb (ecore_private.h:333)
==13853== by 0xDBEC01B: _ecore_main_fd_handlers_call (ecore_main.c:1992)
==13853== by 0xDBEC8A9: _ecore_main_loop_iterate_internal (ecore_main.c:2379)
==13853== by 0xDBEA672: ecore_main_loop_begin (ecore_main.c:1292)
==13853== by 0x441DA9: main (e_main.c:1089)
this used to be set on show since the assumption was that show was the
first time the client would be seen, but this turns out to be incorrect
and results in focus being set too early (breaking policy)
this is a remnant from xdg5-only code where new_client meant "first buffer". with
xdg6, this is no longer the case since a surface can receive infinite commits
without ever having a buffer attached
ref 9a82f7bcb0
xdg6 allows for clients without buffers to be configured such that the
first buffer will match the configured state
if a client is sized before this point, the changes.size flag will be set
This reverts commit e1c3120689.
this commit revealed a number of issues with the xdg6 implementation related
to unconfigured buffer management: see subsequent patches for a less
sledgehammer-y solution
ref 5497fadce4
let's say you sue tiling or some module and it wants a window by
default to maximize or fill the screen or be size XxY ... this stops
the client first having a buffer smaller (or larger) and then sizing
down rendering 2 times (one of the renders is pointless). this makes
initial buffer render/show seamless as it should be in wayland.
So yeah, I've literally used sed to replace every occurrence of
ecore_time_add() with ecore_timer_loop_add() because I'm reasonably
confident that no part of E has a legitimate need for timer based on the
exact current time.
It would be really nice if I'm not wrong. :)
The reason for this is the incredible spew of clock_gettime() calls I'm
seeing on an ARM system (that should have a vdso for gettime, but...)
This can amount to thousands of system calls per second.
#YOLO
Reverting this as it ends up causing multiple events being handled
(touch and pointer) inside various clients if you have both touch and
pointer enabled. Will need a different fix here....
This reverts commit 7906537c02.
Small patch to enable sending wl_touch down/up events when pointer
mouse button events are handled. This is needed in the case where we
do Not have any mouse pointer at all, but we do have touch support.
ref T5094
NB: This allows weston-simple-touch client to operate in Enlightenment
now. There is still something strange happening with EFL clients in E
wrt touch events tho...
Signed-off-by: Chris Michael <cp.michael@samsung.com>
Eina_List *l is already previously defined at the top of this
function. Since we are just using it for list iteration, there is no
need to define it again.
Signed-off-by: Chris Michael <cp.michael@samsung.com>
this adds core basic handling for window stacks where windows behave
correctly as a single unified stack something like what naviframe does
but out-of-window so you can including multiple processes. only on x11
right now as it's being supported/worked on.
as we dont plan to kepe naviframe in future, this is the way to go.
naviframe "pages" will be windows in a stack. the wm should do the
nice thing. in e this will be very nice. for now elsewhere we use
transient_for so a wm would treat this like a bunch of dialogs with a
single parent window. i guess in a desktop thats probably what you
might expect. e will be a little more "finesse" filled.
need to make ibar, tasks,m win menu and winlist (alt-tab) respect this
and only show the top member of a stack.
need to send messages to clients when they are "top" or "middle" or
"bottom" or "alone" in the stack or something so decorations can change.
should add soem new border signals in theme (for both SSD and CSD) to
make this look nice. will need some config additions for that and
ability for e comp to do the right thing
but this is a solid start
We shouldn't be doing this, but there's a collective memory that
this was put in place to fix stuck modifier bugs.
If we run into stuck modifiers again because of this patch, then we
should be fixing them in a different way.
If anyone bisects to this point, I apologize - assign me a ticket.
The focus in timer has been firing for deleted clients, this causes a
NULL pointer dereference.
Then again, maybe the timer should've been disabled by now...
_parent_client_contains_pointer() shouldn't return true if there is no
parent client. This could result in leaving stale resources in the
keyboard focus list and crash the compositor.
Similar set of changes to the recent input region changes. They may not
have been quite so broken to begin with, but it's probably less confusing
to treat both types of region the same way.
It was impossible to set an empty input region.
Input regions were being clipped to the size of whatever buffer was
previously attached when they should really be clipped at commit
time to whatever buffer was attached in the commit.
The intersection operation created round off errors and
e_comp_object_input_area_set is clipping anyway, so we can simplify this
and retain precision.
We don't have any idea at this point what the region will be used for,
and tilers clip to their area on ingress. If we're going to use tilers
for wayland regions we need to make them huge and clip them later.
This fixes a crash where mousing back into internal windows would
cause a crash (from efl wayland mouse pointers).
Signed-off-by: Chris Michael <cp.michael@samsung.com>
This is needed because if we shutdown E then try to run EFL apps in
Weston, the Elm Softcursor mode would have been saved as 'off' so we
should be sure to reset this.
Signed-off-by: Chris Michael <cp.michael@samsung.com>
This commit disables elementary softcursor mode when running as a
wayland compositor. This stops any EFL Wayland Client apps from
creating their own mouse pointer and thus we do not have 2 mouse
pointers on EFL apps anymore.
Signed-off-by: Chris Michael <cp.michael@samsung.com>
Replaces any checked-in wayland protocol files with auto-generation.
In some cases this means renaming include files that didn't use "standard"
names, or adding missing xml files. Any source edits are simple search and
replace, there should be no functional changes.
This fixes a problem when focus moves between two xwayland clients.
My testing of gtk applications shows no regressions, bug if xdg_shell
popups start behaving oddly again, look here first.
Summary:
kbd.keys are only added from e_comp_wl_key_down().
And in that function, duplicated keys are not add in the list.
So only one key is going to be in the list.
Test Plan: Watch kbd.keys list during key press / release.
Reviewers: raster, devilhorns, zmike!, ManMower
Reviewed By: ManMower
Subscribers: duna.oh, input.hacker, cedric
Differential Revision: https://phab.enlightenment.org/D4230
We kept an extra reference for wayland clients and dropped it on their final
render to stop crashes. However, if they're not on screen at the time of
their deletion this reference still needed to be dropped.
Fixes clients remaining on the deskmirror after dying on another desktop.
Moves and resizes tripped up the new xdg focus logic, so test if the
focus is leaving the client entirely before trying to determine if it's
leaving into a parent surface.
There are actually toolkits that create surfaces, do nothing with them,
and destroy them. Sending keyboard leave events for this causes problems.
Fixes a bug in handling of some GTK popups.
in a choice between fixing a corner case popup behavior and breaking dnd
or having functional dnd and adding hacks to fix corner case popup behavior,
adding more hacks was the obvious correct solution
ref 03a4ecbdb0
Mike Blumenkrantz
Derek Foreman
Carsten Haitzler
Chris Michael
JengHyun Kang