This change fixes modifier selection by delegating this to the allocators
and thus to libgbm, which can ask the driver for an appropriate modifier
for scanout, that results in the best possible performance.
We do not have this information, the order in which modifiers are returned
by EGL has no meaning and this is far better then testing modifiers
non-deterministically at random and choosing the first one, that does
not error out...
- renderer: Change the renderer api to not import different buffer types directly,
but import any supported wl_buffer
- renderer: Remove destroy_texture call and move responsibility into implementation
- gles2: Cache egl images as well as textures on wl_buffer userdata
- gles2: Implement delayed destruction of textures to avoid leaking or changing global state on drop
Tracking of Frames, so that only one unique one can exist at a time
(gles does not allow multiple frames being rendered in parallel)
lead to very unfriendly lifetime-heavy code. A renderer is already
*unique*, just move the code there and add an error variant to catch
misuses.
- Cursor is not necessary anymore as rendering hardware vs software
cursors just depends on "where" you render the cursor.
- Format is replaced by drm-fourcc and more related to buffers and
therefor the allocator module anyway.
- GL/Glium are replaced by the renderer module.
- Software was unused and very small anyway and can be implemented
with the new renderer api.
- SwapBuffersError is (for now) just moved into the top-level backend module.
This pulls a lot of code from anvil/src/glium_drawer and replaces glium
as glium has too many assumptions about the backend.
(Mostly that a fixed framebuffer exists, see fix bullet point.)
Depending on how picky glium is, we could try to re-introduce glium support
later with a some workarounds, but for now this is actually more
straight-forward to support and test.
- Add a new GL renderer, that only depends on egl as an initialization platform.
In particular do not depend on a surface being available.
Renderers share some basic drawing functions and may bind objects to render upon.
E.g. surfaces or buffers, textures, etc for the gles2 renderer.
- Be explicit about extensions we require and use. Use some very very common ones
to make our lives easier (e.g. BGRA and unpack to read in manditory shm formats).
- Enable GL debug output
- Allow more flexible rendering (e.g. 3D compositors) by allowing user-provided
projection matrices. Also provide helper functions to allow easy-ish handling
of surface and output transformations.
- Add API for renderers to tell the wayland-frontend about supported buffer-types.
- Also incoperate code from anvil/src/shm_load to handle buffer loading in the renderer
(as this is a renderer dependant operation).
- Remove EGLStream code, nvidia support will re-introduced,
when 470 hopefully lands with GBM support this summer.
- Greately simplify the native code by setting exclusively on
`EGL_EXT_platform_base` and its extensions for initialization.
- Remove generic parameters for the underlying objects on `EGLDisplay`
and `EGLSurface` by using trait objects instead.
- Allow creation of `EGLContext`s without a config for surface-less usage.
- Remove the `Device`, `Surface`, `RawDevice` and `RawSurface` traits.
We are not trying to merge the different responsibilities between...
- Drm-Surfaces: crtcs and planes for displaying buffers
- Gbm-Surfaces: allocating and managing front- and back-buffers
- EGL-Surfaces: Load drawing apis.
..anymore.
- Instead unify the legacy and atomic backends into one `DrmDevice` and
a `DrmSurface`, that under the hood dispatches onto the different modules.
- Do not keep a list of surfaces inside the devices. These are reference-counting
nightmares and a frequent issue due to the read-only nature of `Rc` and `Arc` and
thread-safety related problems.
- Surfaces are mostly doing state-tracking independently of each other
and contain some more helper functions.
- Simplify session code, because we do not need to pass the events
through three layers of devices and surfaces anymore.
- Each surface now represents a plane. Rendering to a plane is just like
any other rendering step, including the Cursor-plane.
(Legacy interfaces only have one to avoid dealing with short-comings of the api).
- Add module that deals with different kinds of buffers (memory and external),
their allocation and usage for rendering. Also try to properly support modifiers this time.
- Describe gbm functionality as an allocator (instead of a rendering device/surface).
- Also create a quick-and-dirty dumb buffer allocator for tesing / simpler tasks.
- Add a (current untested) wrapper for dmabufs and some code for converting from gbm for now.
- (also untested) Swapchain helper to manage front/backbuffers for surfaceless rendering.
It is not necessary to set the cursor position on winit since the
windowing system already updates the position on its own. Instead, doing
this makes the cursor (almost) stuck and unmovable. Thus, this commit
just removes that code from the winit backend.
Fixes: https://github.com/Smithay/smithay/issues/241
Signed-off-by: Uli Schlachter <psychon@znc.in>
* xwayland: Add LaunchHelper
Calling fork() in a multi-threaded process is a bad idea. I am getting
crashes at exit() and "man fork" says that only async-signal-safe
functions can be called after forking in a multi-threaded process.
exit() is not async-signal safe. Thus, forking needs to happen before
any threads are created.
This commit adds a LaunchHelper to the public API. This is a handle to a
forked child process. So far this does not do much, but the intention is
to use this later instead of fork()ing directly.
Signed-off-by: Uli Schlachter <psychon@znc.in>
* xwayland: Move the fork dance to LaunchHelper
No functional changes intended. This only moves the code over and does
not actually use the LaunchHelper for anything.
Signed-off-by: Uli Schlachter <psychon@znc.in>
* Move checking the launch status to LaunchHelper
This should again have no functional changes, but the error log output
might change slightly, since an IOError instead of a NixError is
returned.
Signed-off-by: Uli Schlachter <psychon@znc.in>
* xwayland: Set $DISPLAY earlier
This way, the WM already gets the correct value of $DISPLAY.
Signed-off-by: Uli Schlachter <psychon@znc.in>
* xwayland: Use the launch helper process
FD passing is now used to give the display number and the sockets to the
launch helper process. That process then forks of the actual process
that spawns Xwayland. Once Xwayland started, this is reported back with
another write on the pipe.
Signed-off-by: Uli Schlachter <psychon@znc.in>
* Add to the comment explaining Xwayland startup
Signed-off-by: Uli Schlachter <psychon@znc.in>
* Please rustfmt
Signed-off-by: Uli Schlachter <psychon@znc.in>
* Make clippy happy
Signed-off-by: Uli Schlachter <psychon@znc.in>
This fixes the following clippy warnings:
error: usage of `Rc<T>` when T is a buffer type
--> src/wayland/dmabuf/mod.rs:265:14
|
265 | formats: Rc<Vec<Format>>,
| ^^^^^^^^^^^^^^^ help: try: `Rc<[Format]>`
|
= note: `-D clippy::rc-buffer` implied by `-D warnings`
= help: for further information visit https://rust-lang.github.io/rust-clippy/master/index.html#rc_buffer
error: usage of `Rc<T>` when T is a buffer type
--> src/wayland/shm/mod.rs:91:14
|
91 | formats: Rc<Vec<wl_shm::Format>>,
| ^^^^^^^^^^^^^^^^^^^^^^^ help: try: `Rc<[wl_shm::Format]>`
|
= help: for further information visit https://rust-lang.github.io/rust-clippy/master/index.html#rc_buffer
Signed-off-by: Uli Schlachter <psychon@znc.in>
When starting xwayland with smithay, the first time I would get the
following output (do not ask what happens with displays 0, 1, 2. That's
not important right now):
Dec 29 14:13:31.031 DEBG Attempting to aquire an X11 display lock, D: 3, smithay_module: XWayland
Dec 29 14:13:31.032 DEBG X11 lock aquired, D: 3, smithay_module: XWayland
Dec 29 14:13:31.032 INFO Initialization completed, starting the main loop.
When killing the process with ctrl-c and starting it again, this
happened:
Dec 29 14:13:29.138 DEBG Attempting to aquire an X11 display lock, D: 3, smithay_module: XWayland
Dec 29 14:13:29.138 DEBG Failed to acquire lock, D: 3, smithay_module: XWayland
Dec 29 14:13:29.138 DEBG Lock was blocked by a defunct X11 server, trying again, D: 3, smithay_module: XWayland
Dec 29 14:13:29.139 DEBG Attempting to aquire an X11 display lock, D: 3, smithay_module: XWayland
Dec 29 14:13:29.139 DEBG X11 lock aquired, D: 3, smithay_module: XWayland
Dec 29 14:13:29.139 INFO Cleaning up X11 lock., smithay_module: XWayland
Dec 29 14:13:29.139 DEBG Attempting to aquire an X11 display lock, D: 4, smithay_module: XWayland
Dec 29 14:13:29.139 DEBG X11 lock aquired, D: 4, smithay_module: XWayland
Dec 29 14:13:29.139 INFO Initialization completed, starting the main loop.
The reason for the above behaviour is the
smithay::xwayland::x11_sockets::open_x11_sockets_for_display() failed.
The code successfully acquired the lock file, but then could not bind
the sockets. More specifically, the concrete socket in /tmp/.X11-unix/
already existed and thus bind() failed with EADDRINUSE.
(The code in X11Lock::grab() would then drop the already acquired
X11Lock and its Drop impl deleted the socket. Thus, when I started
things again, this time it successfully acquired display 3.)
Fix this removing the socket before trying to bind it.
This is also done by wlroots: In xwayland/sockets.x, the function
open_socket() has the same job as smithay's open_socket() function.
However, for non-abstract sockets, it would first unlink the target
before trying to bind:
if (addr->sun_path[0]) {
unlink(addr->sun_path);
}
Signed-off-by: Uli Schlachter <psychon@znc.in>
Here is a lock file not created by smithay:
$ xxd /tmp/.X0-lock
00000000: 2020 2020 2020 2035 3537 0a 557.
And this is a lock file created by smithay:
$ xxd /tmp/.X23-lock
00000000: 2020 2020 3130 3733 3332 107332
As you see, the difference is a trailing newline (0x0a).
Lock files 1 to 22 where also created by smithay, because I started it a
couple of times. Reading these lock files fails, since the code expects
a file with at least 11 bytes of content: It uses read_exact() with a
buffer of length 11 in the Err()-branch of the match that this commit
modifies.
Signed-off-by: Uli Schlachter <psychon@znc.in>
Handling SIGUSR1 requires setting a signal handler, which is a global
resource. This can also cause problems in the presence of threads, since
every thread needs to set up a signal handler. This commit instead
changes the code to use a pipe instead (actually a UnixStream). When the
child closes the pipe, the parent wakes up and knows "to do something".
Fixes: https://github.com/Smithay/smithay/issues/246
Signed-off-by: Uli Schlachter <psychon@znc.in>
libinput provides sub-pixel precision for pointer motion & touch events.
Keep this precision by switching all coordinates values from input
events to f64 (rather than i32 or u32). Otherwise, values are rounded
and part of the movment is lost.
Potentially fixes#224
slog-stdlog has a significant dependency tree and is basically unsued if
the downstream crate of Smithay always provides a logger (like anvil),
so it is not really needed.
Change the session backend to rely on Signaler to propagate its
signals. Also introduce the Linkable trait to allow generic composition
of objects needing to listen for signals.
Introduce the `backend_session_elogind` cargo feature which pulls
`backend_session_logind` and makes the logind session backend seek
`libelogind.so` instead of `libsystemd.so`.
Fixes#127
Rework the Session Notifiers so that they are calloop event sources
by themselves, allowing them to be inserted by the user without the
`bind_session` dance.
Also update the logind backend to use the current dbus-rs API, rather
than the deprecated one.
Victor Berger
Victor Brekenfeld
Christian Meissl
Victor Brekenfeld
Benjamin Lee
Poly
Gary Guo
Uli Schlachter
Uli Schlachter
Eric Anholt
Jonas Platte
Mateus Carmo M de F Barbosa
Kirill Chibisov