Summary:
When a CSD button interaction under wayland leads to a compositor action
like move or resize, we essentially "give back" that button press to the
compositor, and it never sends us a mouse up for it.
We need to internally fire a mouse up event to fix up state so the client
doesn't think the mouse is still down. Until now we've been doing this
by setting a flag when we start a move/resize and checking it at next
pointer enter for the window.
This leads to unsolvable races and wacky bookkeeping, and runs afoul of
the fact that we're not actually guaranteed a pointer enter immediately
after a move completes. There is absolutely no way at all on wayland to
know if a move or resize operation has completed.
So, let's just fire the mouse up immediately on start of interaction,
which is raceless.
This fixes a years old bug where dragging a window might leave a stuck
mouse up, and allow hilighting text without drag after the window drag
completes. (elementary-test -to "text editor" with multiple windows open
exhibits this bug)
Depends on D6127
Reviewers: zmike, cedric
Reviewed By: zmike
Tags: #efl
Differential Revision: https://phab.enlightenment.org/D6128
Summary:
These should never be called, so they shouldn't be API.
This kind of internal state fiddling should happen implicitly in the
API that need it.
@beta_break
Depends on D6126
Reviewers: zmike, cedric
Reviewed By: zmike
Tags: #efl
Differential Revision: https://phab.enlightenment.org/D6127
Summary:
This is only fired to trigger a cursor set under wayland, but that cursor
set should be done unconditionally on mouse in.
However, mouse in was being discarded because mouse out was being deferred
when the window was "grabbed" for moving.
If instead we just let the mouse out occur as it should, the cursor
is properly updated on mouse in.
Depends on D6118
Reviewers: zmike, cedric
Reviewed By: zmike
Tags: #efl
Differential Revision: https://phab.enlightenment.org/D6119
Clean up various places where we do flushes that we don't need to
because some immediately following action is going to cause a flush.
Also fix places where we flush without actually doing anything.
False commit when a commit is already pending is an error, but for safety
it should be a nop.
Currently it would overwrite the existing frame callback which could
cause problems on window destruction.
We should be using dmabuf sync ioctls instead of mmap/munmap every draw,
this makes that happen. The surface code continues to do what its always
done, and map/unlock.
This is now done in ecore_evas where it should be. alpha_set now does
only what its name claims it does - sets whether a surface has an alpha
channel or not.
When alpha changes we need to blow away all our buffers in reconfigure,
even if they're still the right dimensions. Otherwise we can end up
rendering to an XRGB buffer when we wanted to use ARGB and we won't have
proper transparency.
Window geometry x, y are the offset from the top left corner of the
buffer, and not screen co-ordinates, so has nothing to do with output
geometry and can't be used to determine which window we're on.
Now that we track surface enter/leave events we can just give the first
output in the list of outputs we know we're on.
Under wayland we can set minimized but not unset it, nor can we tell
if it's been unset. This means we can't cache the value, we need to
make the protocol request any time ecore_wl2_window_iconified_set is
called.
ref T6834
We actually can't ever query this, it's clearly defined that way in the
protocol. There is absolutely no way to ever know if we're iconified.
ref T6834
Apparently when we initiate a client side move in ecore_wl2 we flag that
and send a mouse-up immediately on the next pointer enter.
Do the same for resize.
At some point this might need to be revisited, we should probably be
sending a "cancel" at the start of client initiated move/resize instead
of an up at the end?
Fix T6422
ecore_wl2_window_commit() must be called during window size negotiation,
but this currently trips a warning when no frame callback has been
received for the first commit. We can't even have frame callbacks at
that point because no buffer is attached.
Don't set up the commit_pending logic until after we have a buffer.
As strlen() cannot accept NULL (segfaults), we should check for valid
key, keyname, and compose strings here before passing to strlen().
@fix
Signed-off-by: Chris Michael <cp.michael@samsung.com>
Small patch to reduce calls to strlen when sending key events. This
patch is loosely based on Phab D5567
@fix
Signed-off-by: Chris Michael <cp.michael@samsung.com>
We had a "clever" optimization that would keep a buffer on resize
if it was resizing up horizontal and fit within the previously
allocated stride.
Unfortunately, there still needs to be a buffer reconfigure between
client and compositor that wasn't taking place. Remove this for now.
Since surface handling is now done via modules, we need to ensure
the library can't be shutdown while a surface exists. Otherwise,
we get a segfault trying to call a function we've unmapped.
Fixes a bug on shutdown for some wayland clients using software
rendering.
If we have more buffers than we need for 100 frames then drop the oldest.
This can happen if we're on a hardware plane and then removed from it, or
really whenever the compositor feels like holding onto a few frames.
Trimming the queue too soon could result in having to do a costly full
frame redraw, so we wait a while to make sure we don't need one again.
Having more frames than we need costs us a little every draw since we
always use the oldest available. It also wastes memory.
When a surface leaves all outputs we can discard its buffers to save
memory.
Currently most compositors don't send leave events for iconify, so this
pretty much just saves us a cursor buffer under weston for now, but in
the future it could be used for freeing resources of offscreen (fully
occluded or iconified) windows.
We used to abandon all buffers even if they were locked. This can't
actually free them until they're all released by the compositor.
Instead just free any buffer the compositor doesn't have locked, so we can
still re-use them when they're released without needing a full redraw.
There's probably room for additional cleverness here. If we get a new
frame event before the buffer release we may want to keep it, and if we
get the release first we may want to abandon it.
It was currently only used internally and had the side effect of
creating a new buffer instead of just returning the existing one.
Now it's useful to external callers, as it only returns the existing
wl_buffer and has no freaky side effects.
It was originally thought that this could be common code for multiple
back-ends, but that doesn't really make sense now, so rename it to match
the other dmabuf functions.
The backend should allocate its own private data and return it instead
of a bool.
This assumes all back-ends will need some manner of private data, which
is certanly true for the one back-end we provide.
The specific surface code only needs these generic surface bits to pass
to buffer_create, so make a helper function for that instead of queries
for w, h, and alpha.
This tries to resize the buffer's useable area to fit the specified size -
this is possible if the stride of the buffer is larger than the current
width.
Clear out the window callback when doing session recovery, and
make sure we have a valid on if we get a double commit.
This should stop a session recovery crash, and fix a small leak per
recovery.
b48781aa6c fixed multiple bugs where the
display wasn't flushed correctly, however it was a little overzealous.
Some of the flushes were added after calls that only updated internal
state, some in internal functions in which the caller was already going
to flush, and some were after wayland protocol calls that are double
buffered anyway and won't do anything until a following commit.
Also, I've removes at least one long standing flush where the recently
added flush is in a better location than the original.
in the case where a connection was not actively rendering, there was nothing
which would trigger a display flush, leading to applications potentially
deadlocking
@fix
This fixes cycles of init/shutdown/init where ecore event types would
become invalid, since they are now stored in a dynamic array rather than
a statically stored array.
The risk here is that if a module of EFL tends to init/shutdown in a
"normal" scenario then the event type array will grow in a leaking
manner. This could be fixed by resetting those event ID's only when the
loop actually exits (EFL_EVENT_DEL on the main loop). I'm not using
EFL_EVENT_DEL in this patch as this would add too many event callbacks
to the main loop object, which may result in slightly slower event calls
to it, affecting the overall performance.
This allows something that only has the Ecore_Wl2_Window (ie: something
that isn't engine code) to force dropping of all the buffers.
This should be safe to call at any time as the buffer handling logic
will properly cleanup the buffers when async render is done with them
or the compositor releases them.
This will eventually be used when a wayland client receives a
wl_output.leave events to indicate it isn't displayed on any outputs.
Coverity detected a resource leak here because we were not freeing the
malloc'd 'obo' variable.
Fixes Coverity CID1382907
Signed-off-by: Chris Michael <cp.michael@samsung.com>
commit 0cf806005e correctly fixed a
leaked buffer. However, other code was already accounting for the
leaked reference to the buffer manager, so an extra deref happened
and broke the universe - but only on hardware that no developer
has access to for testing.
Small patch to destroy our test buffer before we exit the
_ecore_wl2_buffer_test function so that we do not leak here.
@fix
Signed-off-by: Chris Michael <cp.michael@samsung.com>
We no longer allocate 3 buffers at startup, we now allocate only as needed.
Trimming the queue will come later, as there are some situations where we
might need 3 buffers and later drop down to 2 (when on a hardware plane)
Most clients will only ever need 2 buffers, so this is a reasonable RAM
savings.
We should only open this when actually testing dmabuf. Otherwise we're
just wasting time and adding an opportunity to fail shm init over
unrelated issues.
Calling this multiple times even after it fails the first time is a legit
thing now. We'll be doing that when we want to test dmabuf at connection
start.
We use immediate mode dmabuf creation at runtime, but this can result in
clients being killed with no option to fallback if the buffers can't be
consumed by the compositor.
This test should catch when a system can allocate a dmabuf buffer and the
compositor claims to accept dmabuf, but the buffer can't actually be used
for whatever reason. We'll then use wl_shm at runtime instead of dmabuf.
It does us no good to be able to allocate dmabuf capable memory if the
compositor can't handle it. This should fix failures on systems where
allocation is possible but the compositor doesn't advertise dmabuf.
There are some binds at startup that result in additional information
being sent, so we may need to call wl_display_sync() multiple times, and
only send the client a SYNC_DONE event when the final one completes.
This moves all the platform specific buffer allocation into ecore_wl2
instead of the engine.
Note that this makes an internal struct available in the header. This
will be removed shortly.