`pairs` order isn't defined and `{...}` will always be ordered.
There is no reason to have random behavior where it can be
predicted at no additional cost.
Test that the placement function used in the default configuration
behaves as intended. This test was failing before the no_overlap and
no_offscreen fixes in two previous commits.
Signed-off-by: Sergey Vlasov <sigprof@gmail.com>
The awful.placement.no_offscreen function did not work properly when
composed with other placement functions; in particular, the default
configuration (awful.placement.no_overlap+awful.placement.no_offscreen)
was broken. The compose function sets args.pretend=true and puts the
result of the previous placement function into args.override_geometry
before calling the next placement function, but no_offscreen did not use
args.override_geometry, therefore the result of the previous placement
function was discarded.
All other placement functions use `geometry_common(c, args)` to get the
current client geometry; `area_common(c)` should be used only when
getting geometry of other clients.
This change also fixes the problem with margin handling (adding margins
should not affect the window size, only the window position should
change); the test output which was adjusted in commit 0275d3537d
is adjusted again to account for this change.
Signed-off-by: Sergey Vlasov <sigprof@gmail.com>
The awful.placement.no_overlap function was adding the window border
width to the client width and height (this is performed in
area_common(), which is called by geometry_common()), but did not
reverse this operation by calling remove_border() before returning the
final geometry; because of this, using no_overlap resulted in increasing
the window width and height by 2*border_width.
The bug was probably introduced in commit ebcc19844e (before
that commit no_overlap changed the window position directly instead of
relying on the new placement infrastructure), but was not noticed
because of other problems (e.g., in the default configuration the result
of no_overlap was overridden by the buggy no_offscreen).
Signed-off-by: Sergey Vlasov <sigprof@gmail.com>
The current taglist/tasklist allow filter function to remove elements
from the list. However they don't allow sorting or additional entries
to be listed.
This commit introduced such a concept. It will later be used by the
layoutlist where it becomes more relevant since layouts are used created
"objects".
Commit fec8d6aa8f fixes awful.placement.no_offscreen to behave
like other placement functions. This means that the margins=40 argument
that this test used and that was previously was just ignored, now
started working. Thus, there are now 40 pixels less on each side of the
client in this test.
Signed-off-by: Uli Schlachter <psychon@znc.in>
When poll() is interrupted because of a signal, it sets errno to EINTR.
GLib ignores this kind of failure.
However, a_glib_poll() calls a_xcb_check() at its end. This will call
xcb_poll_for_event() which internally might call recv(), which can fail
with EAGAIN if no new events are available. Thus, a_glib_poll() will
return an error and set errno to EAGAIN. This leads to the following
error message being printed by GLib:
GLib-WARNING: poll(2) failed due to: Resource temporarily unavailable.
Fix this by preserving the errno from g_poll() in a_glib_poll().
Signed-off-by: Uli Schlachter <psychon@znc.in>
GLib is very careful not to "fetch" any children that it was not asked
to fetch. Thus, if awesome inherits some children, it does not reap them
and starts collecting zombies.
Thus, this commit makes awesome handle SIGCHLD directly: On SIGCHLD, a
byte is written to a pipe. The main loop reads from this pipe and
collects children via waitpid(-1). Unknown children cause a warning to
be printed. We might want to remove this warning if it turns out to be
annoying.
This commit adds 79 lines and removes 89 lines. Thus, this is a net
reduction in lines of codes. This is because previously, we gave the
list of still-running children that we know about to the next awesome
instance we a --reap command line argument. This was added so that
awesome does not get zombie children. However, this commit fixes this
problem and makes all the code for this 'feature' unnecessary.
Fixes: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=886393
Signed-off-by: Uli Schlachter <psychon@znc.in>
When an icon is sent over dbus, we turn this into a cairo ImageSurface.
This is done by turning the actual icon data into a string and using
cairo.ImageSurface.create_for_data() to create a surface for this data.
However, this function only creates an ImageSurface that refers to this
data. It does not copy the data. Thus, when the Lua GC later frees the
string, we have a cairo surface that refers to already-freed data.
Fix this by duplicating the cairo surface, which makes cairo create a
copy of the data. Then, we finish the original surface. While doing
this, the string is kept alive in a local variable.
(Possibly) Fixes: https://github.com/awesomeWM/awesome/issues/2361
Signed-off-by: Uli Schlachter <psychon@znc.in>
Some window gravities require a division by two. Up to now,
test-gravity.c expected this division to always be rounded up. This
commit changes the code to also allow rounding down.
Signed-off-by: Uli Schlachter <psychon@znc.in>
Previously, this function overwrote the value of its argument with the
result. After this change, the function merely changes the given
variable by the calculated argument.
Thus, the old behaviour is achieved by setting the variable to zero
before the call, which all callers already did. However, for most
callers this change means that a temporary variable can be removed and
instead xwindow_translate_for_gravity() will directly change the target
variable.
No change in behaviour intended.
Signed-off-by: Uli Schlachter <psychon@znc.in>
When a tag is specified by name, awful.rules only searched for the tag
on the client's screen. This commit extends the search to all screens,
but only if no specific screen was specified for the new client by some
rule.
Signed-off-by: Uli Schlachter <psychon@znc.in>
There are two ways in which the input focus can change: Lua can request
a change and the X11 server can inform us that the input focus changed
(because some application changed it).
In the first case, we still have to inform the X11 server about the
desired change, in the second case we must not to avoid races due to
X11's asynchronous nature.
However, there was a case where we screwed up: When a focus change is
still pending, meaning that Lua assigned the focus elsewhere, but we
have not yet sent this focus change to the X11 server, we could get an
event from the X11 server telling us that the focus changed. To make
sure that the pending focus change is not lost, we sent the focus change
out in this case (call to client_focus_refresh() in
event_handle_focusin()). After sending out this pending call, we updated
the internal state to record that whatever the X11 server just told us
had the focus. The intention was that our just sent-out focus change
will cause the X11 server to send a new event and our to-be-focused
client then has the focus.
However, if the pending focus change was for a client which only showed
up in this event loop iteration, the client was still banned. This means
that client_focus_refresh() would call client_unban() to be able to give
the focus to this client. However, since awesome (partly) allows to
"focus" currently banned clients, client_unban() recorded that there is
a pending focus change. This caused confusion later on.
In this specific bug, a main window opened a dialog, and when this
dialog was closed, a new dialog window was opened immediately. When the
first dialog was closed, Lua (the focus history) gave the input focus to
the main window. Now, a new dialog showed up and Lua focused it. Next,
we received the event from the X11 server telling us that the main
window was focused. Because there was still a pending focus change to
the new dialog window, event_handle_focusin() called
client_focus_refresh() to send out this focus change. This set
globalconf.focus.need_update to false and continued. However, because
the new dialog only just now appeared, it was still banned, meaning that
client_focus_refresh() had to call client_unban(). This set
globalconf.focus.need_update to true. Thus, when client_focus_refresh()
returned, globalconf.focus.need_update was incorrectly true. Next,
event_handle_focusin() recorded that the main window had the focus.
Thus, it now appeared as if there was a pending focus change for the
main window. Next, we got the event from the X11 server telling us that
the dialog is now focused, and because focus.need_update was set,
awesome now send out a focus change request for the main window.
Fix this race by unsetting globalconf.focus.need_update at the end of
client_focus_refresh() and not at the beginning, thus making sure that
client_unban() cannot set this flag again.
Fixes: https://github.com/awesomeWM/awesome/issues/2220
Signed-off-by: Uli Schlachter <psychon@znc.in>
The typo was introduced in commit a1941efc9.
Its effect should be minimal: :item_enter() itself does not care about
the 'mouse' option, but it forwards to :exec(). Here, an action is
invoked either if it was not caused by the mouse, or if it was caused by
the mouse and either auto_expand is enabled (which is the default), or
the item-to-be-executed is actually the active item.
In other words, it is quite non-trivial to come up with a case where
this typo made a difference. But of course that's no reason to leave the
typo in.
Fixes: https://github.com/awesomeWM/awesome/issues/2347
Signed-off-by: Uli Schlachter <psychon@znc.in>
When a window has a WM_TRANSIENT_FOR property that is later unset,
awesome would still keep c.transient_for pointing to the previous
"parent client". This commit fixes that.
First, property_update_wm_transient_for() is fixed so that it unsets
c->transient_for_window if the WM_TRANSIENT_FOR property is deleted.
Additionally, this then calls client_find_transient_for() to update the
c->transient_for pointer.
Secondly (and a bit unrelated), this changes client_find_transient_for()
so that it always sets c->transient_for. Previously, if updating this
property would introduce a cycle in the transient_for relation, it would
just leave c->transient_for with its old value. After this change, it
gets explicitly set to NULL instead.
Signed-off-by: Uli Schlachter <psychon@znc.in>
Emmanuel Lepage Vallee
Yauhen Kirylau
mergify[bot]
Uli Schlachter
Sergey Vlasov
Peter Kasza
Mak Kolybabi