This is more reliable since it uses absolute coordinates instead of a
relative movement which could cause the cursor position to get out of
sync with the view.
This is the same approach used for resize.
- Move the decision whether a configure should be tracked or not into
the xdg toplevel/xwayland code.
- Only track configures for xdg toplevels with the transaction system
if the dimensions of the view are affected.
Currently we may resize fullscreen views when they become visible/not
visible when switching tags even if their fullscreen state remains
constant. This is suboptimal, and as it turns out also much more complex
to implement.
We need to initialize the geometry on map to ensure the first commit is
handled correctly.
Also we don't care about the x/y of the geometry, only the width/height.
In commitTransaction() we currently the current view state to determine
whether or not to enable the view's scene tree. However we don't update
the view's current state until after that check.
Moving fullscreen views between outputs now works properly.
A case in which we did not inform the client that it is no longer
fullscreen has been fixed as well.
The race is as follows:
1. Output A commits and sets render state to pending_lock_surface
2. Output B commits and sets render state to pending_lock_surface
3. Output A presents and sets render state to lock_surface
4. maybeLock() does not lock because waiting on output B
5. Output A commits and sets render state to pending_lock_surface
6. Output B presents and sets render state to lock_surface
4. maybeLock() does not lock because waiting on output A
The scene_layer_surface may be destroyed before handleDestroy is called,
which means we can't rely on it to access the wlr_layer_surface in
destroyPopups().
This implementation as it stands is incomplete/buggy and will make
updating to wlr_scene more complex.
It will be reimplemented after updating to wlr_scene is complete.
It looks like having the empty error capture |_| on the else branch of
the if statement causes the else branch to be ignored by the compiler.
This should be a compile error, as the condition of the if statement is
a bool, not an error union.
These new functions allow testing commits without messing up the
pending state of the output and needing to rollback. The new apply()
function also makes the code considerably more concise.
A user reported a crash that only reproduces when preloading a hardened
malloc implementation. From the stack trace, this use-after-free seems
to be the most likely cause. Yay hardened malloc!
Outputs that are part of the layout but currently disabled (e.g. due
to use of wlr-output-power-management) are not correctly handled as
river currently waits for them to present a new locked frame before
sending the locked event.
This new frame never comes however since the output is disabled. Fix
this by maintaining the correct Output.lock_render_state as outputs
are enabled/disabled.
Additionally add missing maybeLock() calls to handle the case that all
outputs in the layout are disabled.
Curently since the struct is semantically passed by value not reference,
there is no guarantee that the `seat_node.data.focused.view == &self`
comparison will work as intended. Since updating to Zig 0.10.0, it seems
this latent bug has now manifested and the focused view title is no
longer sent to the client when it changes.
Fix this by taking the view argument by constant pointer instead.
Instead of stashing the active view and setting Seat.focused to the
Xwayland OR surface when a child OR surface of a currently focused
Xwayland view is given keyboard focus, keep Seat.focused set to the
Xwayland view.
Such Override Redirect surfaces are commonly used for drop down menus
and the like, and river should behave as if the parent Xwayland view
still has focus.
This ensures that the riverctl focus-view next/prev commands continue to
work as expected while a popup is open, the correct focused view title
will be sent over river status, etc.
It's also cleaner to centralize this logic in XwaylandOverrideRedirect
and keep it out of Seat.zig.
Xwayland OR menus may disappear if their parent view is deactivated. The
heuristic and ICCCM input model implemented prior, used to determine whether an
OR surface may take focus, does not cover all menus, so retaining parent view
activation works as a catch-all solution for handling unwanted OR menu focus.
We currently scale the width/height of rectangles based on the scaled
x/y instead of the unscaled x/y. This however leads to inconsistent
width/height due to rounding. Fix this bug by basing width/height
scaling off of the original x/y.
Furthermore, fix a typo where we scaled the height off of the x
coordinate instead of the y coordinate.
If the client commits a protocol error or otherwise crashes before the
session has been fully locked, we currently try to send the locked event
without checking if the client has been destroyed.
This commit adds the necessary if statement.
There's currently a potential race in the implementation that can be hit
during unlocking. This is not a security vulnerability, but it does
cause the compositor to crash due to a failed assertion.
This commit simplifies the code and fixes the race as well as tightening
up the assertions around this state/control flow even further.
Currently we send the locked event after rendering and commit of blank
or lock surfaces buffers on all outputs. However, this is technically
not enough to ensure that the buffers have been presented.
Instead, listen to the wlr_output present event to ensure that no
normal, "unlocked" content is possibly visible.
The wlroots rendering API expects colors to be provided with
premultipled alpha but we currently do not parse them as such. This
causes blending with e.g. a transparent border color to be very broken.
The calculation of view.pending.box.x for snap right should be based on
output_width (not output_height).
The inverse applies to view.pending.box.y for snap down.
Currently the spawn-tagmask applies to the currently focused output.
This however means that it is lost if the monitor is unplugged and makes
it hard to set for all outputs.
Change this to make the command apply to all outputs.
This is a breaking change.
When focus-follows-cursor is used with cursor-warp, some windows will
get focus before the cursor properly "enters" the window since they have
a larger input-region than their window geometry, this causes the cursor
to be yanked to the middle unexpectedly.
This fix makes it so the focus is only given when the cursor enters the
window geometry.
It was forgotten to destroy the callback server side object when sending
the destructor event. With the new zig-wayland version, this cannot be
forgotten.
This is nice simplification and allows us to abort startup if the
default xkb configuration (perhaps influenced by XKB_DEFAULT_*
environment variables) is invalid.
Currently the session lock client has no 100% safe way to know when it
is safe to suspend after requesting that the session be locked.
For a suspend to be safe the compositor must have either blanked or
rendered a lock surface on all outputs before suspending. This is
because the current framebuffer on suspend appears to be saved and
displayed again after suspend, at least on my Linux system.
If a new "locked" frame for all outputs is not rendered before suspend,
an "unlocked" frame or frames will likely be briefly displayed on resume
before the lock surfaces are rendered or the screen is blanked.
To fix this, wait until a lock surface has been rendered on all outputs,
or if that times out until all outputs have been blanked, before sending
the locked event to the client.
Resolving this race on the compositor side without protocol changes
is the most effective way to avoid this potential information leak,
regardless of which session lock client is used.
This also cleans up the code by using @Type(), eliminating the need
for the argFlag() and boolFlag() functions.
Allowing [:0]const u8 arguments makes this parser useful for
river-control commands as well.
Warp the cursor to the center of the focused view if the cursor is not
in the bounding box of that view already. This helps the user to keep
track of their cursor when they mostly use the keyboard and the cursor
becomes hidden most of the time, also helps trackpad/trackpoint users.
This reduces the impact of keyboard groups on the Keyboard.zig
implementation and otherwise improves consistency with patterns used
elsewhere in rivers code.
There are also two small changes to the riverctl interface:
- keyboard-group-add-keyboard is renamed to keyboard-group-add
- keyboard-group-remove is added to support removing keyboards from a
group.
The fact that this call is missing is a bug, as the changes made by
arranging the output layers as well as changes to the focus will not be
fully applied.
It is not guaranteed that the next layout_demand event after a user_command
event has the same active tags (for example when there are no views visible).
As an example, a user could trigger a user_command while no views are visible,
then switch to a different tag set which has active views. The active tags of
the previous layout_demand may also be different.
Therefore it is impossible to correctly implement a layout generator which has
user commands apply only to the currently active tag set, which is solved by
this patch.
When river or wlroots write to a closed socket it could generate SIGPIPE
causing the whole desktop to seemingly "crash" with no error log of any
kind. So we ignore the SIGPIPE and just let the write fail with EPIPE to
be handled normally.
Currently we don't send an enter event when a new keyboard device is
created which causes issues when switching ttys. On switching away the
keyboard device is destroyed and leave is sent. Currently on switching
back the keyboard device is re-created but no enter event is sent before
we start sending key events, which is a violation of the protocol.
In cases like multiple hi-res monitors connected through a USB dock, the
preferred mode can fail to work. Such an output was then ignored by river,
which made it impossible to even set another mode manually.
Sway has been reported to solve this issue, so let's employ their solution
and fall back to another mode if possible.
This fixes the return type of Foo.fromWlrSurface() functions which can
in fact return null in the edge case that the role matches but the
corresponding object has already been destroyed.
Still TODO are:
- Touch support for drags
- Mapping input devices to outputs (necessary for good multi-monitor
touch support)
Co-authored-by: Daan Vanoverloop <daan@vanoverloop.xyz>
Some popup menus are not covered by the `overrideRedirectWantsFocus()`
heuristic (e.g. in IntelliJ IDEA), so before focusing an OR window,
its input model should also be checked to ensure that it is able to
take input focus. This appears to fix the popup menus in IntelliJ IDEA,
which would otherwise disappear immediately due to unwanted focus.
- The lifetimes of the Keyboard and Switch structs are now directly
tied to the corresponding InputDevice, which has become a field of
those structs.
- Seat capabilities are now properly updated on removing a keyboard.
These changes align with input device refactoring in upstream wlroots
which will make updating to easier 0.16.0.
Currently we use "switch_device" because that's what the enum variant
happens to be named in zig-wlroots so that it doesn't conflict with the
switch keyword.
This however wasn't really thought through and "switch" makes more sense
to expose to the user.
This was removed a while back because it was buggy and I didn't know
of anyone using it. Since refactoring it is now trivial to implement
and I know of at least one person using it, so I don't mind reviving it.
Until now, Seat.setFocusRaw sent all pressed keys to the client,
including ones that should be eaten. (Try e.g. changing focus to a
nested wlroots compositor with a terminal open to easily see it.)
However, only filtering out the eaten keys is not enough; they were
eaten only once all mappings had been executed. Therefore, the original
function had to be split into one looking up mappings and another
executing them.
Until now, only the event (press/release) for which a mapping was
present got eaten, and the other was passed to the client. From this
commit, a press mapping eats both events and a release mapping eats
nothing (and a press+release combo eats both).
This fixes behavior of some clients that do not make a difference
between press and release (e.g. Firefox with a fullscreen video
exiting fullscreen even on an Esc release event).
Adding the set_override_redirect listener on creation and removing it
on destruction was missed when implementing override redirect state
change handling.
When checking keys for matching mappings, previously we did two checks:
1. Keysyms translated by xkb.
2. Raw keysyms
This commit removes the first check, so only the second is checked.
We're doing this because of strange behavior that xkb shows for some
layouts and keys. When pressing `Shift Space` on some layouts (Swedish
among others), xkb reports `Shift` as consumed. This leads to the case
that we cannot distinguish between `Space` and `Shift Space` presses
when doing a correct translation with xkb.
e.g. `riverctl map -layout 0 normal Super Y spawn foot`
When this mapping is checked against a pressed key, layout 0 will be used to translate the pressed key instead of the currently active layout.
The number denotes to an index of the layouts set with
`XKB_DEFAULT_LAYOUT`.
Currently when the surface under the hidden cursor changes, we
passthrough() which results in the cursor being made visible and giving
pointer focus to the surface under the cursor if any. Obviously this is
not desirable as the cursor is supposed to remain hidden until moved.
This added check prevents this.
From the riverctl.1 man page:
*hide-cursor* *timeout* _timeout_
Hide the cursor if it wasn't moved in the last _timeout_
milliseconds until it is moved again.
The default value is 0, which disables automatically hiding the
cursor. Show the cursor again on any movement.
*hide-cursor* *when-typing* *enabled*|*disabled*
Hide the cursor when pressing any non-modifier key. Show the cursor
again on any movement.
The motivation for this change is to simplify the implementation
and remove a massive footgun that is currently present and causing
UB/crashes. If a layer surface is destroyed in arrangeLayers() then
the code in LayerSurface.handleCommit() after the arrangeLayers()
call accesses free'd memory. This is of course possible to fix,
but it's far simpler to loosen up the protocol implementation a bit.
The new implementation is also in line with what sway and the new
wlroots layer shell helper do and is perfectly valid according to
the protocol.
I have encountered a crash (failing assert) if a view specified a fixed
size less than this minimum, and according to ifreund this behavior was
planned to be removed, anyway.
In particular, this fixes a crash that can happen if a client is made
non-fullscreen and then, while that transaction is ongoing, made
fullscreen again.
Notably, we no longer call both wlr_output_test and wlr_output_commit
when applying an output config, which seems to fix or workaround an
occasional crash since updating to wlroots 0.15.0.
Currently river will place the surface at the top or left edge if
opposing anchors are set without a 0 width/height. Instead, center
the surface between the anchors.
We currently don't properly handle xdg surface geometry of the parent,
which causes popups to render partially off-screen in some cases.
GTK4 clients such as easyeffects seem to trigger this issue reliably.
I thought this should be fine as river won't yield to the event loop
when Seat.deinit() is called before the wlroots seat is destroyed, but
a segfault on exit has been reported with a stack trace mentioning
wlr_seat_destroy(). Let's hope this clears that up.
The new code to dedup XcursorManager.setCursorImage() calls for
efficiency currently doesn't handle clients setting the cursor properly.
This commit corrects this oversight.
This patch allows to focus outputs by clicking on the empty background and by
clicking on layer surfaces without keyboard interactivity. This makes it
possible to use the cursor to focus outputs with no visible views.
This also fixes problems with pointer interactive layer surfaces (for example
launchers and docks) on non-focused outputs.
Currently wlroots sends use the drag destroy event before sending the
wl_data_device.leave event to the client, which makes things a bit
awkward. My patch fixing this has been merged to wlroots master so we
can remove this when upgrading to wlroots 0.15, but until then this
workaround will fix the issue.
This doesn't really matter that much as unrecognized options will still
trigger a help message to be printed, but -h is much more standard so
lets make the predictable choice here while sticking to only single '-'
flags.
mpv for example has key bindings to set the window size to a multiple of
the video resolution. This is a valid use case for client-size resizing
of the view and river should respect this if the view is floating.
This greatly improves the UX of this feature, as views moving under a
stationary cursor (as happens during the zoom command for example) will
no longer trigger focus change.
Currently the implementation treats the x/y coordinates of
View.State.box as layout coordinates instead of output-relative. This
causes issues when using an output not at 0,0.
Layout generators are generally pretty fast. The timeout is only reached when
the generator is faulty / stuck. In that case, freezing for 1 second is simply
bad UX.
Currently the view destruction sequence is started as soon as a view
is unmapped. However, this is incorrect as a client may map the view
again instead of destroying it.
Instead, only start the view destruction sequence when the underlying
xdg toplevel or xwayland surface is destroyed.
Currently if destroy() is called while a subsurface is mapped a dangling
commit listener is left behind. This is obivously a problem, so check if
the subsurface is mapped in destroy() and remove the listener if needed.
Currently if another configure is in flight after the one we are
tracking the serial of and the client acks the second configure as well
(or only the second configure) before committing, we will never realize
the configure we are tracking has been acked.
Instead, listen for the ack_configure signal and set a bool that we can
check on surface commit.
This probably isn't an issue that would actually be hit by well behaved
clients as river doesn't send redundant configure events. However,
having correct code is always better even if it's slightly more complex.
This extends the `csd-filter-add` command to allow matching on window
titles as well, using a `csd-filter-add kind pattern` syntax. The
following kinds are supported:
* `title`, which matches window titles
* `app-id`, which matches app ids
Only exact matches are considered.
As an example following configuration applies client-side decorations to
all windows with the title 'asdf with spaces'.
riverctl csd-filter-add title 'asdf with spaces'
It turns out that wlroots requires us to do a bit more than just create
the wlr_viewporter. Docs are being added to the wlroots header in
https://github.com/swaywm/wlroots/pull/3171
This extends the `float-filter-add` command to allow matching on window
titles as well, using a `float-filter-add kind pattern` syntax. The
following kinds are supported:
* `title`, which matches window titles
* `app-id`, which matches app ids
Only exact matches are considered.
As an example following configuration floats all windows with the title
'asdf with spaces'.
riverctl float-filter-add title 'asdf with spaces'
Repeating mappings are created using the -repeat option to the map
command:
% riverctl map normal $mod+Mod1 K -repeat move up 10
- repeating is only supported for key press (not -release) mappings
- unlike -release, -repeat does not create distinct mappings: mapping a
key with -repeat will replace an existing bare mapping and vice-versa
Resolves#306
Currently if a view is mapped while some other view is fullscreen, it
will not be added to the focus stack, which means that if the fullscreen
view is then closed the view which was not added to the focus stack will
not be focused.
To fix this, always add views to the focus stack on map.
The view.unmap() call may synchronously destroy the view, which makes
the the code removing listeners which is currently run after
view.unmap() access free'd memory.
To fix this, simply change the order of the calls to match that of
XdgToplevel.handleUnmap().
Menus, tooltips, etc. can extend beyond a view's borders. Render views
after their borders so floating content appears on top.
Unfocused floating content can still be obscured by views higher in the
stack and the focused view.
If using the on-output-change cursor warp option river currently crashes
when the last real output is disabled as the noop output used as a
fallback is not present in the output layout.
This is pretty much unusable after recent improvements to the cursor
code, and was totally broken causing a stack overflow as soon as the
cursor was moved over any surface until the previous commit.
Furthermore, none of the core contributors or people active on IRC seem
to use it.
Currently we hit a stack overflow as we do not check if the target view
already has keyboard focus before calling Seat.focus() in
Cursor.passthrough(). To fix this, simply add this check.
When an xdg toplevel, layer surface, etc is destroyed, it is not
guaranteed that all the children in the surface tree have already been
destroyed. If there are still children around, destroying the root of
the tree would leave dangling pointers.
To fix this, destroy all children when destroying any node in the tree.
The current format of #RRGGBBAA is problematic as # starts a comment
in POSIX compliant shells, requiring escaping/quoting and increasing
complexity.
This is a breaking change.
This arrange is actually required because the post_fullscreen box might
not hold the correct dimensions if the view was made fullscreen while
a transaction was already in progress.
Currently if a view is moved from layout to fullscreen while a
transaction involving that view is in progress the saved buffers are not
dropped, which causes stale state to be rendered.
This is guaranteed to already be set to the layout being committed. It
is set either when a client binds a new layout object or when the user
changes the layout namespace in use.
This is currently unused and I don't like the approach anymore
regardless. If/when we need positional arguments (probably when
implementing the upcoming river-control protocol in rivertile)
they should be handled separately from flags.
This commit also improves the CLI error reporting to always print the
usage string if invalid arguments were passed.
The previous commit re-introduced a bug fixed by a3c65713 which caused
the pointer enter event not to be sent until moving the pointer when
switching tag focus or otherwise manipulating the window manager caused
the cursor to end up over a new surface.
If the current Cursor.maybeResetState() function is called while in
passthrough mode, it will send a pointer motion event. This is
unnecessary as we have already sent the same pointer motion event at
least once.
Also refactor the code slightly and improve naming.
Handling output destroy now requires the wlr_output_layout to still be
around, as we need it to properly handle cursor state. In order to make
sure that all outputs are destroyed before the wlr_output_layout is,
simply destroy the backend before calling Root.deinit().
Now that we properly handle state changes during cursor operations,
blocking these commands if the target view is the target of a cursor
operation is unnecessary complexity. It is also inconsistent as we
don't block changing the tags of the view.
This was slightly out of sync with Cursor.surfaceAt() which did not
fullscreen or xwayland unmanaged views properly. Also simplify things
and improve correctness by always rendering all xdg popups. A view
losing focus does not always mean that all popups will be destroyed.
A transaction may move the current target of a cursor action to a
non-visible tag, make it fullscreen, or otherwise change things such
that the current cursor state no longer makes sense.
To handle this, check if we should reset cursor state every time a
transaction is committed.
Currently the spawn command takes any number of arguments and naively
joins them together with spaces before passing them as the single
argument of `/bin/sh -c`. This however produces unexpected results as
soon as shell quoting gets involved in the arguments passed to spawn.
For example, running
riverctl spawn foo "bar baz"
will execute `/bin/sh -c "foo bar baz"`, unexpectedly splitting bar and
baz into separate arguments. To avoid this confusion, make the spawn
command take only a single argument, forcing the user to quote properly
to spawn multi-argument commands.
- Remove advertise_view and advertise_done events. Using the information
provided by these for any purpose would make the layout far less
predictable. Futhermore, in the months this has been available for use,
to my knowledge nobody has actually used it for anything useful.
- Replace the set/mod layout value events with a single user_command
event. This simplifies the protocol and is more flexible for clients.
- Add a layout_name argument to the commit request. This name is an
arbitrary, user-facing string that might, for example, be displayed by a
status bar. This was present in early drafts of the protocol, but was
removed in favor of river-options. Since river-options itself has since
been removed and this feature is nice to have, re-add it.
- Rename main factor to main ratio in rivertile. The "factor" name was
just legacy from dwm, "ratio" is much more accurate.
This code is complex and increases maintenance burden but doesn't
add any functionality, only eye-candy.
Futhermore, neither I nor any of the core contributors use it.
There may be a place in river for such eye-candy down the line, in which
case this code could be revived. Currently river is early enough in its
development that our focus should be on core functionality instead.
A true "default" config doesn't make sense for river. Everyone who uses
river seriously will customize their init script. Futhermore, the
current behavior of embedding the install path of the default system
config in the river binary is complex and prone to breaking.
On output change, if the cursor is not already on the newly focused
output, it will now be warped to its center. The check is necessary,
since focusing outputs with the pointer will be implemented in
the future.
A client is free to change its mind and request a different
size/anchor/etc after recieving the initial configure but before
attaching and committing the first buffer. This means that we should
respond to such a situation with a new configure.
mako has been observed doing this in the wild for example.
Currently in handleUnmap() we call View.unmap() before removing
listeners. However View.unmap() may destroy the view before returning
if the transaction started doesn't have to wait on any configures.
To ensure that we don't try to remove listeners which have already been
free'd, do this before calling View.unmap().
The Layout struct holds a pointer to the Output which becomes invalid
when the Output is destroyed so we must ensure all the layouts of an
Output are destroyed first.
The transaction system exists to coordinate size changes of all views
in a layout in order to achieve frame perfection. Since many clients
do not need to commit a new buffer in response to a activated state
change alone, this breaks things when such a configure event is tracked
by the transaction system. Instead, simply send activated and fullscreen
configures right away but still track this state in a double-buffered
way so that e.g. border color changes based on focus are frame-perfect.
This also fixes a related issue with the transaction system where views
that did not need to commit in response to our first configure were not
rendered until their next frame.
This protocol involves far too much accidental complexity. The original
motivating use-case was to provide a convenient way to send arbitrary
data to layout clients at runtime in order to avoid layout clients
needing to implement their own IPC and do this over a side-channel.
Instead of implementing a quite complex but still rigid options protocol
and storing this state in the compositor, instead we will simply add
events to the layout protocol to support this use case.
Consider the status quo event sequence:
1. send get_option_handle request (riverctl)
2. roundtrip waiting for first event (riverctl)
3. send set_foo_value request (riverctl)
4. receive set_foo_value request (river)
5. send foo_value event to all current handles (river)
6. receive foo_value event (rivertile)
7. send parameters_changed request (rivertile)
8. receive parameters_changed request (river)
9. send layout_demand (river)
And compare with the event sequence after the proposed change:
1. send set_foo_value request (riverctl)
2. receive set_foo_value request (river)
3. send set_foo_value event (river)
4. send layout_demand (river)
This requires *much* less back and forth between the server and clients
and is clearly much simpler.
Options are now all global but may be overridden per-output. If an
output local value is requested but none has been set, the global value
is provided instead. This makes for much better ergonomics when
configuring layout related options in particular.
Run the init command in a new process group and send SIGTERM to the
entire group on exit. Without doing this, only the sh invocation used
for the `sh -c` would receive SIGTERM.
This is particularly useful when starting a per-session server manager
as the init command.
Replace the current layout mechanism based on passing args to a child
process and parsing it's stdout with a new wayland protocol. This much
more robust and allows for more featureful layout generators.
Co-authored-by: Isaac Freund <ifreund@ifreund.xyz>
