If a transient failure occurs in wlr_output_commit, re-render until it
doesn't happen. This could possibly be removed if we decide to
implement damage tracking in the future.
Unlike with X window managers, the display socket in Wayland isn't set
up prior to starting the compositor. Because of this, you can't pipe
the compositor's output directly into a program which needs access to
$WAYLAND_DISPLAY, which is a typical setup for this purpose. Existing
scripts have been forced to create a pipe/FIFO or a temporary file as an
intermediary.
Instead, send the status info directly to stdin of the -s command, which
*does* have access to $WAYLAND_DISPLAY.
Fixes#103.
Fixes#102. The "ideal" behavior might be to ignore buttons other than
the one being used for the action, but this is super-simple and still
seems reasonable.
The wlr-output-management protocol requires that either all of the
changes from an apply request be applied successfully, in which case a
"succeeded" event is sent, or all of the changes are reverted and a
"failed" event is sent. As written, this could partially commit
changes, then fail.
Test the changes first (even for an "apply" event), then commit or
rollback as appropriate.
Additionally, variables xcursor and xcursor_mgr are only used
when xwayland is defined, so I make the variables declaration
contingent on whether xwayland is being used
Replaces the outputOrder patch.
This avoids recalculating positions and allows to arrange monitors in
any order, not just from left to right.
The order in which monitors are defined in config.h still matters but
it's just the order in the list, not the actual position.
This is the order of *_destroy calls which resulted in the fewest
errors/leaks detected by Valgrind. Most of the errors come from the
gbm_allocator code - will have to figure out which destroy call is still
missing.
Similar to Linux kernel approach, encapsulate some of the uglier
conditional compilation into inline functions in header files.
The goal is to make dwl.c more attractive to people who embrace the
suckless philosophy - simple, short, hackable, and easy to understand.
We want dwm users to feel comfortable here, not scare them off. Plus,
if we do this right, the main dwl.c code should require only minimal
changes once XWayland is no longer a necessary evil.
According to `cloc`, this also brings dwl.c down below 2000 lines of
non-blank, non-comment code.
When a new client is spawned, fullscreen isn't disabled for all clients
in that monitor any more.
Instead, all fullscreen clients are kept fullscreen, while other clients
spawn in the background.
When fullscreen is disabled, all clients are rearranged.
This is made to make dwl more flexible allowing multiple fullscreen
clients at the same time, have floating clients on top of a fullscreen
one and let stuff happen without quitting fullscreen, like many other
WMs and DEs.
Disable fullscreen on all visible clients in that monitor also before
enabling it on another client.
quitallfullscreen() is reintroduced becouse is now more useful
set c->isfullscreen later to avoid making quitallfullscreen() disable
fullscreen on the current client
...in internal calls to restore pointer focus. Necessary for the
unclutter patch, and there's no harm in avoiding this call even in
mainline; might prevents issues in same edge cases.
Don't let internal calls to motionnotify(0) meant to update the pointer
focus from maplayersurfacenotify and destroylayersurfacenotify also
shift the keyboard focus to the surface under the cursor with
sloppyfocus.
wtf is the point of this crap? It makes the code harder to follow,
increases the line count and made me fail compilation a million times.
We shouldn't blindy follow everything about suckless's style.
Scaling a wlr_box without rounding can sometimes make the borders not
connected and nice. I noticed this when setting scale in monrules to 1.2
So I've went and copied what Sway did in desktop/output.c but without
having a second function and now using a random rounding macro I found
on the internet so as to not use round from math.h.
Distribute it as a patch like in dwm since graphical applications
usually provide their own keybinding; I guess it's only for terminals.
Note that even though these commits don't let you open multiple windows
in fullscreen and cycle between them like in dwm, with just
fullscreennotify spawning new windows or changing tag would still exit
fullscreen automatically, but you would have to toggle fullscreen twice
when switching back to the fullscreen window to enter fullscreen again,
so this is better since it avoids that.
quitfullscreen() was replicating the functionalities of setfullscreen(c,
0)
Reusing setfullscreen() in quitfullscreen() leads to a 3 line function,
which is useless since quitfullscreen() is used once anyway
This fixes the bug that happens when changing workspace (or any time
arrange() is called) where there are fullscreen windows, which are still
fullscreen but leave the space for layer surfaces like waybar (which
should be hidden when going fullscreen)
Also as soon one fullscreen window is found hte function returns to
improve efficiency
Store position and size of windows before going fullscreen. This is more
efficient than arrange() and also works with floating windows
All the clients keep their original position because arrange() isn't
used after quitting fullscreen
Because maprequest immediately calls wl_list_insert(&fstack, &c->flink),
in the following call to setmon(), the selclient() which is passed to
focusclient() as the old client is actually the newly mapped client, and
the real old one is never deactivated. You can see this by, for example,
opening Chromium's devtools, then spawning a terminal. The background of
the focused line in the devtools doesn't change from light blue to grey.
We can't just remove wl_list_insert(&fstack, &c->flink) from maprequest,
because calling wl_list_remove in focusclient() with a client that has
not been added to the list causes a segmentation fault.
Therefore we fix the focusclient call by not passing it the old client
every time, but instead using the wlroots function that gets the focused
surface and deactivate that, like in TinyWL.
This also avoids getting the selected client and passing it to
focusclient() on every call unnecessarily, and will allow removing
shouldfocusclients in a future commit by checking if old is a layer
surface instead.
It makes wl-clipboard work properly in neovim, without having to create
a transparent surface that steals focus and causes flickering. It's also
required for clipman.
The code in this else completely freezes my system when I run the
swayidle command to replicate xset dpms force off. No idea if it works
on multiple monitors, but for now avoid running when there's 1 monitor.
Also remove the comment with the function name in sway.
Since wlr_output_enable doesn't have any effect before finishing all the
procedure, a little hack allows to make use of focusmon(), which must
know the latest in about which output is currently disabled
Also improve performance in focusmon() and cleaner code in
outputmgrapplyortest()
With the recent changes in output-management, the extra argument in
closemon() would be needed only when unplugging the monitor, so it isn't
worth it anymore. Also now is more efficient.
m->link.next leads to errors if the monitor to disable doesn't have a
"next" (right) monitor and is currently focused. dirtmon() does more
checks.
In some previous commits m->link.next is told to be left monitor, which
is wrong
Also focusclient() explicitly checks for disabled monitors (this fixes
in case of more than one disabled monitor)
Focus the top client on newmon, which we know for sure that it isn't
going to be unplugged or disabled and actually set that as the focused
monitor to move the focus. This is necessary to prevent crash when
disabling monitors with the output-management patch.
This allows to fix output-management: move clients to the monitor on the
left of the disabled one, instead of the leftmost (which might happen to
be the disabled one)
Also using wl_list_foreach() and then brake after the first iteration is
ugly and inefficient
When using wlr-randr every monitor's configuration is reevaluated, so it
must check which monitors are actually being disabled. The only way to
correctly do that is to compare the names.
When a monitor is disabled with wlr_randr, all clients on that monitor
aren't lost but they are moved to the leftmost monitor with the same
method that handles monitor hot unplug
There is no need to repeat this. This needs to be reculalculated in my
output-management implementation too, and since I'm already calling
updatemons, this patch avoids having to repeat the assignment again.
quitfullscreen() was replicating the functionalities of setfullscreen(c,
0)
Reusing setfullscreen() in quitfullscreen() leads to a 3 line function,
which is useless since quitfullscreen() is used once anyway
This fixes the bug that happens when changing workspace (or any time
arrange() is called) where there are fullscreen windows, which are still
fullscreen but leave the space for layer surfaces like waybar (which
should be hidden when going fullscreen)
Also as soon one fullscreen window is found hte function returns to
improve efficiency
Floating widndows with "x < removed monitor's width" aren't resized
(they used to disappear in negative coordinates).
Actually delete monitors when they are unplugged, recalculate sgeom and
give a new monitor to clients that were on the removed one with setmon()
arrangefloat() funcion has been exploded to save iterations in
cleanupmon().
Also if a monitor that supports auto suspension is turned off, dwl will
count it as unplugged (it will become unreachable and all clients will
be moved to the leftmost monitor). However, if at least one monitor
isn't plugged in, dwl will still crash the same as before.
Unlike sway, when the output configuration is changed and restored,
(unplug + plug the same monitor for example) previous application
positions aren't kept. This is due to the fact that on sway every
workspace is unique among all monitors.
Compensate the coordinate changes when adding a new monitor.
Every test so far confirms that monitors are always added to the left,
on top of the list, so every floating window's x coordinate has to be
incremented by the width of the new monitor.
When a monitor is created or removed, the geometries of the old ones
must be updated. This is also more efficient than before since we
calculate the monitor geometries only when creating and destroying
monitors. arrangelayers() is needed to recalculate m->w. arrange() is so
clients don't move to the left monitor when plugging or unplugging
monitors (clients keep the same coordinates but the field below them
changes).
The bug was caused by usable_area's x and y not being set in
arrangelayers. For example if on a 2nd HD monitor, x should be 1920
while the first one ends at 1919. So I don't see why m->m should be
recalculated after creating the monitor.
If you don't recalculate the monitor's geometry before arranging,
clients get arranged in the first monitor. I don't understand why this
fixes the bug since tile() uses m->w rather than m->m, nor why it needs
to be recalculated after creating the monitor but sway does it too.
Although not necessary to fix the bug I also made arrangelayer() do like
sway again and recalculate usable_area instead of reusing m->m, since
m->m seems to be incorrect until it gets recalculated shortly after in
arrange(), so I suspect that leaving usable_area = m->m will cause
issues under certain circumstances.
Someone with a multi-monitor setup or better knowledge of Wayland may be
able to figure out the cause of the bug. For now, this makes layer shell
work.
When a layer surface is destroyed focus should be returned to the last
client. Luckily if there are multiple overlays the previous overlay
still gets focused.
Store position and size of windows before going fullscreen. This is more
efficient than arrange() and also works with floating windows
All the clients keep their original position because arrange() isn't
used after quitting fullscreen
rename Layer to LayerSurface; layer should refer to overlay, top, bottom
or background
LayerSurface variables are always called layersurface
wlr_layer_surface_v1 variables are always called wlr_layer_surface
Honestly not sure why a specific surface is focused rather than the
client figuring that out. Seems to work in a quick test, but we can
remember this commit if something breaks for, I dunno, mouse people.
This arranges the function into some logical tasks: deactivate the old
client, update wlroots' keyboard focus, update our data structures, and
activate the new client. The last two only need to be done when
focusing something new, so an early return saves some horizontal space.
The getatom function returns the atom variable, which is only
initialized in case of a success. This results in a maybe-uninitialized
warning/error. After this commit, now a zero value is returned in case
of error.
attach_render tells the output that a "new" buffer has been prepared
(even if we haven't changed it). We need to call that and then commit
it to keep the render loop going.
Software cursors will freeze momentarily during layout updates, but I
suspect that this is not as easily fixed as it sounds. You can force
software cursors by running:
WLR_NO_HARDWARE_CURSORS=1 ./dwl
* xwayland: add server and basic window functionality
* xwayland: add server and basic window functionality
* xwayland: add server and basic window functionality
* xwayland: add server and basic window functionality
It was just exiting with code 1 for me. The problem turned out to be
that you *need* to set XDG_RUNTIME_DIR or it won't work (I think e.g.
systemd does that for you, but on Void it's not set by default), so
mention that in the README.
