Don't switch the internal tracking of focus to the swaylock surface,
to allow for switching back to the previously active window (or the
currently active window, if some new process changed).
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
When destroying lock surfaces, we really should only unlock a
desktop_shell if the set of lock surfaces has dropped to zero (since
callers need to do a set_lock_surface for every output).
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
Exherbo installs architecture dependent data in a different place than architecture
independent data. More concretely: binaries go in /usr/$chost/{bin,lib},
data goes in /usr/share and configs in /etc, /etc is already configurable
through CMAKE_INSTALL_FULL_SYSCONFDIR but the datadir was not. This
patch fixes it so that things can be pushed in the right places.
- Make sure CMake always finds absolute paths for Cairo, Pango and GdkPixbuf
- Add forgotten json-c include path to swaymsg/CMakeLists.txt
- Disable -Werror because of assert warnings
- Add correct /proc/pid/file path for FreeBSD
- Use libepoll-shim on FreeBSD
- Only use Linux capabilities on, well, Linux
For workspace containers, swayc_change_layout also changes ->layout alongside
->workspace_layout when it's a sensible thing to do. There is an additional test
for 'layout toggle' command which ensures that containers will be tiled
horizontally after toggling from tabbed or stacked.
If workspace layout is set to tabbed or stacked, its C_VIEW children
should get wrapped in a container. Alongside that, move_container was
modified to retain previous functionality.
LD_PRELOAD enables keyloggers to easily be made. This solution isn't
perfect - really a secure system wouldn't have LD_PRELOAD at all. It was
a stupid idea in the first place.
Some users may want to switch buttons on their input devices, turns out
libinput already supports it. Let's add a support for it in our config.
Signed-off-by: Michał Winiarski <knr@hardline.pl>
This does two important things when using alpha:
1. At corners, borders don't double in opacity
2. Foreground elements (text) can be made transparent and you can see
fully through to the wallpaper
Add swayc_change_layout function, which changes either layout or
workspace_layout, depending on the container type. Workspace being
always L_HORIZ makes this much more i3-compatible.
This is necessary because commands in the config file (mode for
instance) emit ipc events, and if ipc_init has not been called the
ipc_clients_list is not initialized, and we segfault. This fixes that
bug.
This implements commands such as:
workspace number 9: test
If a workspace with the given number exists then it will be focused,
otherwise a new workspace with the given name will be created.
When using a bar on a named output, load_swaybars() requires the
output to be active (ie. in the root container), but this is not the case if
the bar is added to the last output. To fix this, load_swaybars() is now
called after the output has been added to the root container.
After fixing that, swaybar would segfault due to using the wrong index
variable when loading outputs and config.
Indicator border color is now used only when a container is the only
child. Reason for the change? i3 does it this way. Sway container
borders are now a bit more similar to i3 ones.
This patch makes it possible to move views between outputs using the
`move left|right|up|down` commands.
It behaves similar to i3 with one important difference. The focus will
always follow the view being moved, unlike i3 where the focus doesn't
always follow the view to a new output (I assume that's a bug in i3).
`vertical` and `horizontal` should be swapped.
If border_left and border_right were set to 0, the vertical borders
instead of the horizontal borders would be hidden.
i3 handles this command equally.
Fixes#819. If workspace is focused and command 'move container to
workspace/output' is issued, workspace child containers are wrapped in a
new container and moved according to command.
This is a feature that can't work the same as i3, because there is no
real window classes in wayland. This way, we avoid null titles as much
as possible.
When workspace_auto_back_and_forth is enabled, workspaces get switched
twice with previously mentioned command, which is not the expected
behavior.
Removes one redundant creation of previous workspace.
Focus the container on the output (not the output itself) when an output is
focused.
This is intended to fix a bug where borders are not updated correctly when
switching the vt away/back to sway.
When switching back to a workspace after new window creation, it is now
necessary to descend the focus stack into the focused container of the
workspace to determine which container should get the focus. This is because
the `set_focused_container()` function no longer automatically descends into
the focus stack to find the correct view to focus.
Replace `update_view_border()` with `update_container_border()`. The latter
should handle both the case where the container is a view or if the container
has children.
The previous implementation of focus handling assumed that only views can be
focused. Containers can also be focused with a command like `focus parent` or
`focus child`.
Change `set_focused_container()` to handle the case of the given container
being a container with children and update borders accordingly.
Deeply nested containers which had their layouts changed didn't update
their actual_geometry, this messed up their child containers. Those got
width and height of 0, which was then decreased for stacked/tabbed
containers by title height. Underflow ensued, these containers suddenly
had height 4294967273. In short, not updating actual_geometry didn't
play nicely with nested containers.
When titlebar is hidden, top border of the topmost view inside
tabbed/stacked container will not be drawn. This is changed in layout.c
On the other hand, top border should be drawn sometimes, for example
when titlebar is hidden on a view that is not the topmost inside
tabbed/stacked container. This is changed in border.c
Previously, cmd_kill only closed a focused view, while containers were
not affected. Now it closes all views that are children of the focused
container.
Whenever a stacked or tabbed container has focused, paint the titlebars of all
its child windows to be focused as well to indicate the parent stack/tab
container has the focus.
In `move_focus()`, when given an output, set the focus to the workspace of that
output instead of the output itself.
This fixes a bug that did not allow users to switch between outputs introduced
in afc6ad6.
It also fixes other issues before that commit when a workspace with children
was selected and the user tried to switch focus in the direction of another
output.
In the `focus parent` command, do not set focus above the workspace level.
These containers are not meant to be focused.
This prevents a crash on repeated `focus parent` commands.
Panels were explicitly rendered by calling wlc_surface_render in
handle_output_pre_render. Calling wlc_surface_render does not set the
surface's geometry (like wlc_view_set_geometry does). Sway did not call
wlc_view_set_geometry for panels, so wlc defaulted their geometry to be at
the origin. This is not correct for bars unless their location is top.
Furthermore, for a surface to receive pointer events, its mask has to be
set to visible. This causes wlc to render these surfaces, causing panels
and backgrounds to be rendered twice.
This commit makes panels and surfaces visible, sets the correct geometries
and removes the code that explicitly rendered them.
Tabbed/stacked containers are now created only if a view is present on
the workspace. If a view is created on previously empty tabbed/stacked
workspace, it gets wrapped in a container.
Prior to this commit all windows (e.g. shell surfaces) were handled the same
way in handle_view_created. Since backgrounds and panels have to be treated
differently, they could not be shell surfaces. This changes checks whether
a client is a background or a panel in handle_view_created and exists to
let them be dealt with elsewhere.
This fixes a segfault, when trying to get parent of the workspace/root container/(?), as it is not assuered that the view's parent node is not null in the loop
Escape line return when reading from a file with the '\' character.
Similar to shell scripts.
Signed-off-by: Roosembert Palacios <roosembert.palacios@epfl.ch>
Escape line return in configuration file with the '\' character. Similar
to shell scripts.
Signed-off-by: Roosembert Palacios <roosembert.palacios@epfl.ch>
Values cannot be negative or 0; if so uses the default 75x50.
Uses the same syntax as i3: floating_minimum_size <width> x <height>, although the x can be anything.
when creating a new output, move to that output all extant workspaces
that are assigned to that output.
(unrelated) remove comment that was no longer applicable, fix spacing in
an assignment
When headers were installed in more sofisticated places (but package
config knows it right), it revealed missing paths in CMake
configuration. Lets fix it.
This patch makes sure to clear the border buffer of fullscreen view so
the border doesn't get drawn behind a fullscreen view, which would be
visible if the view was transparent.
When creating a new view, wlc usually returns an initial geometry with
size 1x1. Setting those values as desired width/height causes a problem
for some windows (QT5) because they don't request a new geometry for
instance when made floating, so the floating window becomes 1x1.
To fix this problem we can just omit setting the desired width/height on
new_view and instead let the clients request a certain size if they feel
like it. e.i. gnome-calculator.
Fix#578
I made this configurable but I didn't make the command for it. That's
left as an exercise to an eager contributor.
mod_scroll_behavior [gaps inner|gaps outer]
Would merge implementions of more behaviors for mod+scroll, if anyone
has some neato ideas.
The borders are implemented as a surface/buffer attached to each view
which is sent to and rendered by wlc in the view_pre_render callback.
All the drawing logic is handled in sway/border.c and all the logic for
calculating the geometry of the border/view is handled in
`update_geometry` in sway/layout.c (same place as gaps are calculated).
The include command (`include <path>`) makes it possible to include sub
config files from the main config file (or from within other sub config
files).
The include command uses the following rules for including config files:
* the `path` can be either a full path or a path that is relative to the
parent config. Shell expansion is supported, so it's possible to do
`include ~/.config/sway.d/*`.
* The same config file can only be included once (to prevent include
cycles). If a config is included multiple times it will just be
ignored after it has been included once.
* Including a sub config file is the same as inserting the content of
that file into the parent config, thus rules about overwriting
bindsyms etc. works the same as for a single config.
Implement #542
Sway has been very strict when it comes to key handling. Only on an
exact match would a bindsym be triggered.
This patch makes it less strict by for instance allowing the key combo
`$mod+1+2` to act as `$mod+2` if 2 was the last pressed key and `$mod+1`
if 1 was the last pressed key.
The new key handling uses the following algorithm:
1. List of bindings sorted by number of keys in binding (already the
default)
2. Find all bindings covered by the current keyboard state and list them
by same order as in 1.
3. Select the first binding from the list where the last pressed key is
part of the binding.
Addresses #452
Before passing a command to a command handler the quotes are stripped
from each argument in the command. This is usually the wanted behavior
but causes a problem in the case of `exec` where quoted arguments can be
required when passing the exec command to `/bin/sh -c`.
This patch makes `exec` a special case and doesn't strip quotes from the
arguments. It will just pass the exec command verbatim to the exec
command handler.
Fix#518
desktop_shell.panel_size was only used to determine if sway should
rearrange the output when rendering the panel in the output_pre_render
hook. This is not needed since the output will have been arranged at
that point.
It also caused sway to rearrange all the time when running with two
or more different monitors/resolutions because panel_size kept changing
with every output_pre_render callback.
Should fix#514
Swaylock spawns and focuses a view for each output in sway. This can
sometimes move the focus to a new output after locking and unlocking the
screens.
This patch makes sure that the output which had focus when swaylock
was invoked, will regain focus once swaylock is closed/unlocked.
Fix#499
Documents most of the bar commands in sway-bar(5) manpage.
The following command has not been document because they haven't been
fully implemented yet:
* mode
* hidden_state
* modifier
* tray_output
* tray_padding
Close#375
This should be a real fix for #509
This schedules a render when a background or panel is added to sway
through the desktop shell interface, that makes sure the render isn't
scheduled before the bg or panel is ready and you don't end up with a
black screen until the cursor is moved.
This makes sure that the outputs are rendered when sway is launched, so
the user doesn't have to move the cursor before the background and bar
gets rendered on screen.
Fixes#509
It's possible to assign workspaces to certain outputs using the command:
workspace <name> output <output>
However, this did not work in some cases where the workspace was
assigned before the given output was made available to sway.
This patch fixes those cases.
In anticipation for #375, reorganized and augmented slightly sway(5)
so it makes a difference between commands intended for configuration,
commands intended for control, and those that can serve as both.
This patch aims to correctly handle moving focus <left|right|up|down>
between outputs.
For instance, if moving from one output to a new output at the left of
the current one, it should focus the right-most view/container on the
new output, and the opposite if moving from right to left. This should
happen regardless of the previously stored focus of the new output.
This also handles moving to a new output above or below the current one.
Calling `exit` in sway_terminate prevents sway from correctly shutting
down (freeing data, cleanly terminating the ipc server, etc.).
A better way is to exit straight away if the failure occurs before
`wlc_run` and use sway_abort as usual if it occur when wlc is running.
Sway used to attempt sending an IPC command composed of every argument
after the first non-option argument encountered.
Now, raises an error if an option is encountered before the intended command.
Some options such as -h or -v take effect when parsing, so they still
apply.
The get-socketpath long option had an undocumented short alternative
as `p`. It has been removed.
However, the code in the options array is still the 'p' char.
This makes sure that a named output config is applied before the
general wildcard config when a new output is created. This ensures that
the config:
output * ...
output NAME ...
behaves the same way as:
output NAME ...
output * ...
This makes IPC GET_PIXELS use the new `wlc_pixels_read` call instead of
the deprecated `wlc_output_get_pixels`.
The old version worked by passing a callback function to wlc which would
grab the pixels and send them to the IPC client.
The new version works by maintaining a list of clients who have
requested the pixels of some output and then grap and send the pixels in
the output_post_render hook of the `wlc_interface`.
This adds quotes around multiword arguments before they are passed to
`/bin/sh -c` in an exec command.
Example:
I connect to irc like this:
exec termite -e "mosh server tmux a"
Without this patch the arguments are passed to sh as:
termite -e mosh server tmux a
When it should be:
termite -e "mosh server tmux a"
For the command to work.
This is a possible fix for #384.
To be honest I don't fully understand why this bug is happening, but I
have narrowed it down to the view stack in wlc and how sway orders views
in very specific situations (those described in #384).
Anyway this should fix the problem by eliminating the call to
`wlc_view_bring_to_front` which isn't really needed anyway since sending
all invisible views to the back is the same as bringing all visible
views to the front (rotating the view stack).
CMAKE_INSTALL_FULL_SYSCONFIG is not actually passed to
the C preprocessor. I remember it working, so I must have
messed up somewhere last time I touched this.
This is fixed by manually passing its value to the C preprocessor
through the SYSCONFDIR definition
Fix#444
This is a temporary fix, the real fix is to store the commands as a
formatted argv array, so they don't have to be reformatted all over the
place.
This implements the IPC binding event for keyboard bindings. It is
slightly different from the i3 implementation [1] since sway supports
more than one non-modifier key in a binding. Thus the json interface has
been changed from:
{
...
"symbol": "t",
...
}
to:
{
...
"symbols": [ "t" ],
...
}
[1] http://i3wm.org/docs/ipc.html#_binding_event
This is a "simple" version of --release (same as i3) that only supports
a binding that contain one normal key. e.g.:
bindsym --release $mod+x exec somthing-fun
I didn't bother implementing it for a combination like `$mod+x+z` since
it is a bit tricky to get right and also a bit weird to actually do on a
keyboard.
This makes sure that a modifier event is only sent for active bar
modifiers, and that it is only sent once for each of those modifiers.
An active bar modifier is a modifier defined for a bar with `mode hide`
and `hidden_state hide`.
The i3wm config locations are visited _before_ using the fallback
configs. The man page was confusing - it talked about the fallback
configs first, but also said they are looked at "at last". By changing
the order of the sentences, this should be clearer.
Detects when a bar modifier key is pressed/released and sends a modifier
IPC event to any listeners (usually swaybars).
This way a swaybar can listen on the modifier event and hide/show the
bar accordingly (not implemented yet)
The modifier event looks like this:
{
"change": "pressed", // or released
"modifier": "Mod4"
}
This fixes https://github.com/SirCmpwn/sway/issues/431 by returning
focus to the fullscreen view. Also it fixes the issue with the
fullscreen view pointer not being set which did my head in
This fixes a bug where the key at index 0 in the `key_state_array` would
be overwritten by the next pressed key. This broke any bindings
consisting of multiple non-mod keys like: `$mod+a+b`.
If the width or height of a container can't be evenly distributed to its
children, then the layout algorithm still thought it got it right (due
to using decimals) which caused a gap of one or more pixels for some
window arrangements.
This is fixed by this patch by first rounding off the width and height
(so that decimals are never introduced) and then adjusting the last
view in a container to fill the remaining pixels (which now is counted
correctly due to the decimals being removed).
Also, due to the way gaps are implemented, an odd sized gap can never be
aligned properly, so just adjust to closest even number.
I've tried to make as few changes, as possible.
Usually the reason for using qsort_r is, that you can pass an extra userdata pointer to the
compare function. However, in sway list_sort wrapped qsort_r and always called a wrapper
function for comparing, the wrapper function then had the real compare function as argument.
The only thing, that the wrapper function does, is dereferencing the 'left' and 'right' function
arguments before passing them to the real compare function.
I have renamed list_sort to list_qsort to avoid confusion (so nobody tries to use list_qsort like
list_sort) and removed the wrapper functionality. Now the dereferencing must be done in the
compare function, that gets passed.
Some compare functions were used in both list_sort and list_seq_find. To make the difference
clear, I've added a '_qsort' suffix to the compare functions, that are intended to be used with
the new list_qsort. (In other words: list_qsort is not compatible anymore with list_seq_find).
- Changed and renamed function (it isn't used anywhere but in commands.c, and only for sorting):
compare_set -> compare_set_qsort
- New wrapper functions:
sway_binding_cmp_qsort (for sway_binding_cmp)
sway_mouse_binding_cmp_qsort (for sway_mouse_binding_cmp)
Our initial implementation of `bar { }` assumed that the commands could
only be used in the config. This is not true for two commands:
* bar mode
* bar hidden_state
This patch makes it possible to issue these commands outside a bar
block, for instance through swaymsg
$ swaymsg bar mode hide bar-0
This does not implement the `barconfig_update` IPC event which should be
trigged from these commands. I have added TODO's where this should be
added once implemented.
Track each panel separately via its wl_resource. `set_panel_position`
might be called before `set_panel`, so reuse panel config.
Place the position in panel_config so that each panel has its own
position.
This works by tracking the pids of the child processes in the related
output container and terminating the processes and spawning new ones on
a config reload.
Should solve: #347
If the output is not at the correct size then that info must be queried
from wlc. The output size is used by e.g. seamless mouse to detect
output edges.
With this patch the output size is now correct and the workspace size is
adjusted according to any panels.
Without this patch seamless mouse would fail to detect outputs
above/below each other if there was a panel in between because the
output would offically end where the panel started, not at the actual
screen edge.
fixes#308
Ordered by number ascending, with insert before same numbers.
Workspaces without numbers are appended at the end of the list.
Example order:
1 2:named 3:the_second 3:the_first 9 FIRST_NAME SECOND_NAME ...
If the id is defined by another bar it will just use the default id for
the bar. Typically `bar-x`.
If the id command is used multiple times within a bar block, the last
one will 'win'.
If focus would move in an output with a fullscreen view, it always
leaves the output. If focus would enter an output with a fullscreen
view, it always focuses the fullscreened view.
CMake takes a `FALLBACK_CONFIG_DIR` flag which is the directory where the
standard configuration file `config` is copied at installation.
If loading from typical configuration directories fails, sway loads
FALLBACK_CONFIG_DIR/config (/etc/sway/config by default).
When turning a float to a non-float, `get_focused_container` might
return another floating view, causing the active view to be inserted
into the floating list on its workspace instead of the normal child list
which it should. (Since it has `is_floating` as false the resulting
discrepency triggered other bad behaviour eventually leading sway to
crash.)
This patch fixes that by simply checking floating status before making
it a sibling.
Using 'flag' results in duplicate code paths for short and long options.
This broke the -q short option in swaymsg, because there was:
{"quiet", no_argument, &quiet, 'q'}
Which will set quiet to 'q' and return 0, not 'q'.
First of all because it's not needed that early, and second of all
because there's a bug where calling `sway --get-socketpath` via `popen`
causes the child sway process to spin/hang instead of returning EOF.
(Specifically `(unset SWAYSOCK; swaymsg)` hangs.) This patch fixes that.
(Also note that this patch moves the "detailed review" comment, so I
guess this patch requires extra detailed review?)
This function looks for bound commands that start with `workspace` (ie.
the commands that change to a static workspace) and fetches the
workspace name.
However, if it's actually a list of commands, then the parsing will pick
up the delimiter ("," or ";") and also fail to recognize keywords
("next" etc).
This patch fixes that by properly separating with delimiters.
This makes escaping the arguments obsolete.
Also avoid dynamic memory allocation for the output id. It only supported ids up
to 99. Now we support up to 999, and take 4 bytes off the stack instead.
A criteria is a string in the form of `[class="regex.*" title="str"]`.
It is stored in a struct with a list of *tokens* which is a
attribute/value pair (stored as a `crit_token` struct). Most tokens will
also have a precompiled regex stored that will be used during criteria
matching.
for_window command: When a new view is created its metadata is tested
against all stored criteria, and if a match is found the associated
command list is executed.
Unfortunately some metadata is not available in sway at the moment
(specifically `instance`, `window_role` and `urgent`). Any criteria
string that tries to match an unsupported attribute will fail.
(Note that while the criteria code can be used to parse any criteria
string it is currently only used by the `for_window` command.)
This fixes a compiler warning:
../sway/extensions.c: In function ‘set_background’:
../sway/extensions.c:16:37: warning: implicit declaration of function ‘malloc’ [-Wimplicit-function-declaration]
struct background_config *config = malloc(sizeof(struct background_config));
^
../sway/extensions.c:16:37: warning: incompatible implicit declaration of built-in function ‘malloc’
../sway/extensions.c:16:37: note: include ‘<stdlib.h>’ or provide a declaration of ‘malloc’
We can't handle them currently (the criteria needs to e.g. be passed to
each command handler which then needs to do the right thing), so it's
better to just do nothing than to create unexpected results (because the
command was executed on the wrong view).
(Before this patch any command list with a criteria string would simply
fail to parse, so this is at least a step in the right direction.)
This also fixes a bug where issuing a new "workspace a output b" command
for an already assigned workspace would not work (the old config would
be found first and used instead).
This does not work as expected. I think the problem is on the wlc side.
Please review, @Cloudef. To reproduce the issues:
1. Run sway
2. Open terminal in sway
3. Run swaybg
swaybg will create a surface and ask to have it set as the background,
but wlc_handle_from_wl_surface_resource will return 0. If the swaybg
surface is a shell surface, then it works - but wlc complains about the
pointer type and segfaults as soon as the pre-render hook tries to draw
the background.
When querying for an adjacent output we now need an absolute position in
order to know which adjacent output that matches. (The position is
either the current mouse position or the center of the currently focused
container, depending on context.)
If two outputs have one edge each that at least partially align with
each other they now count as adjacent.
Seamless mouse is affected by this and now properly moves and positions
itself between outputs with "uneven" placement (as long as they have at
least some part of the edge adjacent to each other).
When focusing or moving a container in a specified direction the center
of the current focused container decides where to look for an adjacent
output. So if e.g. an output has two adjacent outputs to the right and a
"focus right" command is issued then it's the placement of the currently
focused container that decides which output actually gets focused.
Also, if an output has at least one output adjacent in some direction
but the entire edge is not covered (ie. it has "holes" with no outputs),
then the algorithm will choose the output that is closest to the
currently focused container (this does not apply to seamless mouse, the
pointer will just stop at the edge in that case).
After adding pid to the socket path the `--get-socketpath` command broke
because it doesn't know the pid of the running instance. Fix this by
setting and querying `SWAYSOCK`.
Also ignore `SWAYSOCK` upon normal startup if a socket exists at that
location (ie. from another sway instance), and don't overwrite `I3SOCK`
if it exists either.
Socket now includes pid in the filename (fixes nested sway sessions or
old sockets causing problems).
Fixed warnings on strict aliasing and cleaned up relevant code in
general.
When yes, the old behaviour of adding half the inner gap around each
view is used.
When no, don't add any gap when an edge of the view aligns with the
workspace. The result is inner gap only between views, not against the
workspace edge.
The algorithm is not perfect because it means the extra space is
distributed amongst edge-aligned views only, but it's simple, looks good
and it works.
Place mouse at center of focused view when changing to a workspace on a
different output, if option is enabled. (This replicates existing i3
option.)
This can be triggered in multiple ways:
A) via `workspace <name>` which changes output
B) via `focus <direction>` which changes output
C) via `focus output <name>` which (obviously) changes output
Replicates i3 option. Verbosity level given as command line argument
becomes default log level, and using 'debuglog toggle' switches back and
forth between default and debug (or L_ERROR and debug if default is also
L_DEBUG).
This is an undocumented feature (the word "number" is just ignored
anyway), but it exists to be compatible with i3 config syntax.
Plus some code cleanup at the same time.
If e.g. a window has a popup open then that will lock the current focus,
making a workspace switch denied.
So don't move the mouse pointer in such cases.
In i3 the ipc reply will contain a human readable error message, and
this patch replicates that behaviour.
However, that error message is also useful for logging, which this
patch takes advantage of.
E.g. instead of logging errors directly in commands.c/checkargs, it is
fed back to the caller which eventually ends up logging everything with
maximum context available (config.c/read_config).
So instead of logging e.g. "Error on line 'exit'" it will now log:
"Error on line 'exit': Can't execute from config."
Before this patch sway would proceed into find_handler which would cause
a crash trying to parse a NULL string.
This could be triggered via e.g. `i3-msg -s $(sway --get-socketpath)`.
When using an i3 config verbatim in sway this switch to exec or
exec_always might appear.
Before this patch the switch would be passed to /bin/sh, causing
an error, and the command would not be run.
