Inspired by #5384, but instead of just using the border width, this PR
reports the actual frame extents for the window, which may also include
the title bar (for floating windows and tiled windows in plain split
containers, but not for tiled windows in stacked/tabbed containers).
The existing `con_border_style_rect()` function should already handle
all configuration options which can affect the decoration sizes (if it
does not, that would also show up in other places); its result just
needs to be converted into the format used by the `_NET_FRAME_EXTENTS`
property.
This PR fixes#4292 probably in the best way possible (the reported
`_NET_FRAME_EXTENTS` values should always match the actual sizes of
window frame elements which are actually drawn into the X11 frame window
into which the client window is reparented). The only really problematic
case is with the stacked/tabbed containers, for which the title bar is
actually drawn into a completely separate window, therefore the title
bar size cannot be reported in `_NET_FRAME_EXTENTS` (actually I tried to
calculate the size of those decorations and add it to the top decoration
size, but that did not change the behavior of `picom`).
<details><summary>Large screenshots here (3840×2160)</summary>
Example of configuration with `hide_edge_borders smart` — a single
window does not have borders, so only the top frame size is non-zero:
but multiple windows have borders:
Changing border width works too (although with `border normal 8` you can
see that the top border overlaps the title text, because on the i3 side
that border does not really exists, and `picom` just draws it over; also
the pixel sizes reported by `xprop` and `xwininfo` are not identical to
what is specified in i3, because I use 168 dpi on this system, therefore
4 px in the i3 config = 7 dpx):
Handling of tabbed containers is less perfect though. Here is a single
tabbed container with `hide_edge_borders smart`, so it does not really
have a border — note that all frame extents are zero, and the titlebar
is rounded separately (although it could easily be excluded from
rounding, that does not really help much):
Once the border actually appears, you may notice that the top part of
the `picom` border actually gets drawn over the top part of the window,
partially obscuring the top line in this terminal (`picom` does not mind
that the top frame size is reported as 0):
Some examples of floating windows (no major problems there):
Options like `hide_edge_borders both` work too when gaps are removed
(although the resulting behavior with `picom` is probably not very
useful — the rounded border gets drawn only if all of the left, bottom
and right borders are present):
The same with `border pixel 8` (note that windows with only the top
border hidden still get the rounded border treatment by `picom`, but the
border overlaps the top part of the window):
</details>
---------
Signed-off-by: Sergey Vlasov <sigprof@gmail.com>
Co-authored-by: Orestis Floros <orestisflo@gmail.com>
This was originally mentioned in #3085 but left for a future PR.
One of the noticeable limitations is that pressing the modifier while
the drag is already initiated, will not swap the containers but instead
cancel the drag. This is because of how `drag_pointer()` is written and
would be quite an involved case to handle it.
The crash was brought up in a comment in
https://github.com/i3/i3/discussions/6076#discussioncomment-9536969
The cause is that the command criteria are matching a window in the
scratchpad. In that case, the assertion in get_output_for_con() fails.
That happens because there is no `Output` for the `Con` output of a
scratchpad window.
I've decided to *not* remove the offending assertion but rather rely on
the caller not using the function with internal containers. My reasoning
is:
1. If get_output_for_con can return NULL then the caller will either
segfault (which is worse) or needs to check the return for NULL.
2. The case where the return can be NULL is already known, it happens
for internal containers.
3. Therefore, the caller should already prevent the situation with a
call to con_is_internal(). Thus, the `assert`ion can remain.
There is also the potential fix of con_get_workspace returning some
arbitrary output (e.g. first in the list or currently focused one)
instead of NULL. This can lead to more tricky to catch bugs.
This fixes the following warnings on 32 bit systems
```
[60/108] Compiling C object i3.p/src_regex.c.o
In file included from ../include/all.h:40,
from ../src/regex.c:10:
../src/regex.c: In function ‘regex_new’:
../include/log.h:29:33: warning: format ‘%lu’ expects argument of type ‘long unsigned int’, but argument 2 has type ‘size_t’ {aka ‘unsigned int’} [-Wformat=]
29 | #define ELOG(fmt, ...) errorlog("ERROR: " fmt, ##__VA_ARGS__)
| ^~~~~~~~~
../src/regex.c:35:9: note: in expansion of macro ‘ELOG’
35 | ELOG("PCRE regular expression compilation failed at %lu: %s\n",
| ^~~~
[93/108] Compiling C object i3-input.p/i3-input_main.c.o
../i3-input/main.c: In function ‘finish_input’:
../i3-input/main.c:173:29: warning: format ‘%ld’ expects argument of type ‘long int’, but argument 2 has type ‘size_t’ {aka ‘unsigned int’} [-Wformat=]
173 | printf("occurrences = %ld\n", cnt);
| ~~^ ~~~
| | |
| | size_t {aka unsigned int}
| long int
| %d
```
One case when this might be useful is when i3 is restarted and there are
children that terminate after the previous i3 instance shut down but
before the new one set things up.
Fixes#5756
While this initially worked fine, at some point these patches broke
because libcairo started calling shmget(2) - a syscall not covered by
any pledge promise - and a common pitfall when using pledge with
graphics-oriented applications.
Various attempts were made to fix them, but at some time they were
simply disabled in the OpenBSD port:
a4a9f41dd75a03c386ba
This seems pointless and creates needless friction both for the i3 team
who was willing to carry ugly code and for the OpenBSD ports maintainers
who had to disable that code again.
Let's abandon this experiment.
Since there is no separate error handling the `SIGUSR2` signal is
registered to get the write return code after exiting the program.
Fixes#5958
---------
Signed-off-by: Andre Werner <andre.werner@systec-electronic.com>
eg if you have workspaces: { 1, 2:a, 2:b, 3 } and are on workspace 1,
then 'workspace next' should traverse 1 -> 2:a -> 2:b -> 3 -> 1 instead
of 1 -> 2:a -> 3 -> 1.
Fixes#4452
This had pretty much identical behaviour to hide_edge_borders which made
it confusing. The `hide_edge_borders smart_no_gaps` implementation has an extra check
which fixes#5406.
Enforces a rule that we have followed for years now. Yes, the diff is
quite big but we get it over with once and we prevent having to nit-pick
future PRs.
If a window occupies the entirety of its workspace vertically and/or horizontally, pass it the _NET_WM_STATE_MAXIMIZED_{HORZ, VERT} atoms. This helps applications like Google Chrome draw the tab bar correctly and handle tab clicks correctly (see https://crbug.com/1495853).
This change is based on work from @yshui in #2380.
Grabing the pointer produces a `GrabFrozen` error in applications that
are run from key bindings. Since we don't need the pointer in such
cases, we can change the call to use ASYNC. This seems to be a
historical leftover.
I've tested locally that these still work:
- bindsym $mod+x ...
- bindsym --release $mod+x ...
- bindsym $mod+button1 ...
- bindsym --release $mod+button1 ...
- bindsym --release $mod+x exec program that grabs the keyboard
now works (see original issue)
Even in the main branch, I actually couldn't get `import` and `xdotool`
to fail with the pointer being frozen, potentially because these
programs wait a bit for the pointer to be unfrozen like i3lock does.
This patch came up in
https://github.com/i3/i3/issues/5735#issuecomment-1781321011
I wonder why the pointer is actually grabbed.
The argument I change in `xcb_grab_key` there, is `pointer_mode`, from
https://www.x.org/releases/X11R7.7/doc/man/man3/xcb_grab_key.3.xhtml:
```
pointer_mode
One of the following values:
XCB_GRAB_MODE_SYNC
The state of the keyboard appears to freeze: No further keyboard events are generated by the server until the grabbing client issues a releasing AllowEvents request or until the keyboard grab is released.
XCB_GRAB_MODE_ASYNC
Keyboard event processing continues normally.
```
I traced via `git blame` the usage of `xcb_grab_key` throughout 14 years
of i3 development and it seems that `pointer_mode` was always set to
`XCB_GRAB_MODE_SYNC`, going all the way back to
b664456706.
Fixes#5735
Instead of creating a graphics context for every surface_t, this commit
adds a cache that allows to "remember" up to two GCs. Thus, the code
uses less GCs. When a GC from the cache can be used, this also gets rid
of a round-trip to the X11 server. Both of these are tiny, insignificant
savings, but so what?
Since GCs are per-depth, this code needs access to get_visual_depth().
To avoid a code duplication, this function is moved to libi3.
Fixes: https://github.com/i3/i3/issues/3478
Signed-off-by: Uli Schlachter <psychon@znc.in>
Unfortunately, grabbing / ungrabbing doesn't seem to work correctly in
xvfb so we can't really test this.
I also fixed the deduplication code in bindings_get_buttons_to_grab().
Having just a single fork is beneficial, as it preserves the approprate
parent information for the children of i3, which is useful in some
scenarious e.g. when a child wants to do something on the wm's exit
(possible via `prctl(PR_SET_PDEATHSIG, ...)`).
Moreover, this is a zero-cost benefit: i3 is already using libev with
the default loop, which automatically reaps all the zombie children even
when there is no corresponding event set.
Resolves#5506
