this allows different display protocols to start their applications at
different times to ensure that any initialization has completed prior to
starting anything requiring a window
fix T3475
so i was profiling today .. leak hunting .. and i noticed. if you have
enough appss open - eg terminology, e uses a huge amount of memory...
for icons. terminology is 128x128 ... thats 64k per icon. open up a
lot of terminology windows and we duplicate that 64k per every window
on the wm sside because we get the data. it would apply for any app
that sets a netwm icon. this can be come rather silly if you have like
100 terminals. it's worse with larger icons (eg 256x256 - 256k per
icon).
this puts in a simply list for shared icons and a lookup on fetch to
de-duplicate and share icon data. this should drop memory usage
nicely.
@improvement
if multiple x11 clients receive focus during the same mainloop iteration,
an almost unbreakable cycle of window focus chaining will occur, resulting in
both windows being focused simultaneously--or so it appears--which results in
no window being able to receive input. to avoid this, ensure that only one x11
client can receive focus in a given loop iteration
due to event bursts, it's possible for multiple x11 clients to receive
mouse in events on during the same main loop iteration. in this scenario,
only the last client has received an actionable mouse in, and applying this
event after the dispatch has completed ensures that multiple clients do not
all receive mouse in+out events during the same loop
this greatly improves mouse-based focus reliability in a number of cases
This reverts commit 26a7ba3a58.
this can only occur if something forces an event flush during shutdown.
in this case, whatever is triggering the event flush is a bug, not the
dereferencing of a pointer which is guaranteed to exist for the normal
lifetime of the process
in the event of a wayland start, x11 comp init will fail, meaning that
cleanup must occur in order to avoid erroneous triggering of x11 handlers
#TooSoon
this removes the per desktop profile config and replaces it with a
per-screen one that is tied to a specific display so it is far more
logical than per desktop. this allows e to set up different scaling
per screen for apps that use elementary for example via this derived
profile.
this of course is slightly problematic for e itself since it now uses
elm - as this will cause e to go kind-of-crazy with differing profiles
as it fights with itself and elm if 2 screens have different profiles.
this requires elm to be fixed to allow custom profiles per window.
this also currently won't switch profile of a window when you
reconfigure screens.
@feature
xx
the values requested by the client will be based on its geometry and not
the geometry of the frame. using the frame geometry here results in windows
moving ↘ based on the top/left frame sizes
fix T2912
in the case that a mouse move event occurs, the compositor should validate
the event to ensure that the mouse cursor is actually over the window that
the event claims to be from
fix T2594
in order to continue rendering an iconic client without breaking icccm,
it's necessary to map the client's window again and unset iconic state
whenever rendering is needed, then re-set states when rendering
stops
ref T2788
E_Client->comp_data is null after a client has been deleted, so
attempting to handle events which require the dereferencing of this
pointer after a client has been deleted will result in a crash
these events should be rejected after delete regardless, since at this
time the compositor has stopped handling events for the client
ref f42c6aa187
i got a crash here and the bt was broken and i couldnt check if
_e_comp_x_client_data_get() returned null, but it's the only thing
that would make sense, so protect against this to avoid a crash. as
this was a one-off, i can't find out more,
@fix
in the case where a shaped window with many rects exists, there is a high
probability of the damage rect count being huge, leading to massive blocking for
each frame as the compositor attempts to fetch all of these rects from the xserver.
instead, the compositor can shortcut this by forcing a full-window damage any time
the rect count is sufficiently high, trading a blocking socket operation for some
amount of (potential) overdraw.
testing in affected scenarios has shown huge improvements: where previously the entire
compositor would lock up, things work as expected now
see https://bugzilla.mozilla.org/show_bug.cgi?id=1214746 for a sample case
it seems that damages for popup windows in some applications,
namely blink-based browsers, are reported incorrectly, resulting
in sometimes having a blank window
according to ICCCM 4.1.4:
Only the client can effect a transition into or out of the Withdrawn state
withdrawn windows cannot be shown under any circumstances. the best that can
be done is to try mapping the window and hope it decides to appear.
to prevent any inadvertent showing of the window before it leaves the
withdrawn state, we play games with the E_Client->ignored flag in order
to skip client evals until we get notified that maybe we want to stop
skipping those evals
ref T2745
gtk apps set an atom which provides information about the area
where non-window content (eg. shadows) may be drawn; this area
must not be used in placement calculations.
the easiest method for implementing this functionality was to add
a case to the compositor geometry interceptors which effectively
flip the client struct geometry values such that the E_Client->client
is outside of the more commonly used E_Client->x/y/w/h
fix T2744
failure to allow pixmaps/clients to be retrived by parent window will
result in api users being greatly inconvenienced after a reparenting has
occurred
it seems that since the first version of the enlightenment compositor
in e17, damage events in x11 have never been used correctly. using
the event struct members will only give the bounding box/area instead
of the damaged regions; the real regions must be explicitly fetched
from the server
this removes the need for a lot of hacks which were added over the years
to make override windows render correctly, and also probably reduces
rendering overhead slightly
in the case that the canvas window has just had focus set on it, apply this focus
and ensure that no client retains focus
this resolves a race condition where focusing the compositor canvas <-> client
extremely quickly would result in a client trying to steal focus when it was
not actually focused
a notable (but trivial) side effect is that now when flipping desks at high speed while using
mouse-based focus policies, the user is almost guaranteed to end on a desk which
has open windows on it
it seems that the reported damage events upon resizing an override window
are not accurate, and so we must force a full damage here while avoiding a
render queue in order to ensure that the full contents of the override will
be rendered in the next frame
fix T2045
these are triggered "in passing" when mouse in events occur and do
not necessarily indicate that the mouse has entered this specific window
failing to reject such events can cause mouse-based focus policies to
attempt to set focus onto windows which are not visible, resulting in
an infinite loop where no window is actually focused
in the event that these windows are different, event_window is the parent of window
which may or may not be explicitly tracked by an E_Client, so the wheel events here
should be sent to the parent as is done in mouse button events
fix T2604
it seems that some clients, eg. libreoffice, don't set the modal window
property on child dialogs. instead of fighting for focus, set up the child
as a modal on the parent and then avoid the whole issue
fix T2594
a client with this flag set here is unreliable to use as a stacking
reference since it has yet to be stacked and can be located anywhere
in the window stack.
fixes internal window stacking on startup
blocks execution of resizes until the surface commit arrives. reduces
the race condition between resize and render and eliminates frame drops
during slow resizes
xwayland compositing requires that we set up a root window cursor image
immediately since we'll be getting that cursor surface to display as soon
as the pointer goes out of an x11 client's window
this requires that both canvas cursors and window cursors be present for the same
E_Pointer object, even though only the canvas cursor is actually visible
#kansas
due to recent changes in ecore-input-evas, mouse events are propagated
differently; specifically, there are now "more" events than there previously were.
as a result, grabs on internal wins are no longer necessary, though they probably
never were necessary after the elm conversion
see 5cb6cdbc5e1a13ea0262e155983b494e6519abde in efl
I did an audit of this and it seemed that it no longer served the purpose
for which it was originally intended. specifically, this is for enforcing
click: raise/focus options, and so grabs must be in play on client windows
only when they are not focused to ensure that we get mouse events and can
then focus them. the grabs must then be removed once the window has focus
to avoid spurious mouse eventing
xwayland sets a wrong size on some (eg. menus) clients and wayland
cannot provide geometry or stacking information, so ensure that all
of this is copied over
also remove overrides from focus stack
this should fix the case of mouse-based focus policies trying to reapply
focus after another client has stolen it away without the pointer leaving
the window
in the case of different window <-> event_window, window is a child window
of event_window, and thus checking event_window here is valid (and necessary)
when you have click to focus we have a passive grab set up. somewhere
that window changed to the parent window instead of the client. this
leads to a side effect of a leave and enter event on clients for every
click. generally clients are ok with this, but some seem to have buggy
event handling. these enter/leave events are a side effect of the
passive grab even though we allow/replay the event.
this fixes that by placing passive grabs on the client window itself
instead of the parent.
@fix
these helper functions automatically account for "swapped" xwayland
clients and return the expected value from the wl client comp_data.
in this way, all of the current x11 compositor code can be reused with
minimal changes
in order to maximize the amount of reused code the following details the current
process for xwayland compositing:
* get map request from window
* force reparenting
* show window
* await WL_SURFACE_ID x11 message
* move x11 client data + pixmap onto corresponding wayland client
* business as usual with wayland compositing
this is pretty similar to the method of the reference code in weston,
except that there's no x11 compositor in weston
this follows 56cabf59c6 then
4e5521b4d8 where i have been trying to
fix a crash with e client and comp win references etc. i have gone
over all referencing with a fine tooth comb and found all the nigglies
i can., no leaks now, no crashes, no valgrind complaints etc. so i
call this fixed now. as best i know this is new in e20, so not a
backport fix
the refcoutning for e_comp and e comp clients seemed to be a bit off -
i read over every ref and unref carefully and fix it. this leads to
the com-_data being null (properly now), so now check for that too.