GitLab

Mutter recently added headless tests, and when running those tests the
Wayland compositor runs for a very short time.

Xwayland is spawned by the Wayland compositor and upon startup will
query the various Wayland protocol supported by the compositor.

To do so, it will do a roundtrip to the Wayland server waiting for
events it expects.

If the Wayland compositor terminates before Xwayland has got the replies
it expects, it will loop indefinitely calling `wl_display_roundtrip()`
continuously.

To avoid that issue, add a new `xwl_screen_roundtrip()` that checks for
the returned value from `wl_display_roundtrip()` and fails if it is
negative.
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Reviewed-by: Roman Gilg <subdiff@gmail.com>
Reviewed-by: Jonas Ådahl <jadahl@gmail.com>

Since we now only work with UTF-8 (or ISO8859-1) text in the clipboard,
we don't need to setlocale().

This avoids including Xdefs.h, which means we avoid all the issues with
_XSERVER64 effecting how types are defined by that.

All helper client code now uses xcb, so calling XSetAuthorization() is
no longer needed.

This is the last reference to libX11 from helper clients, so linking
with x11-xcb and libX11 is no longer required.

Also drop (unneeded?) linking with libXau.

Also drop installing these prerequistes on AppvVeyor.

Also move prototypes for functions in winauth.c from win.h into a new
header, winauth.h, and include that where needed.

Convert clipboard integration code from libX11 to xcb

This drops support for COMPOUND_TEXT.  Presumably some ancient
(pre-2000) clients exist which support that, but not UTF8_STRING, but we
don't have an example to test with. (Given the nature of the thing, the
users of those clients probably work in CJK languages)

Supporting COMPOUND_TEXT would also involve writing (or extracting from
Xlib) support for the ISO 2022 encoding.

v2:
Fix the length of text property set by a SelectionRequest

The length of the text property is not neccessarily the same as the
length of the clipboard text before it is d2u converted (specifically,
if that contains any '\r\n' sequences, it will be shorter as they are
now just '\n')

Always use CF_UNICODETEXT clipboard format.  Windows will automatically
down-convert to CF_TEXT for clients which request that.

This is subtly different in one way: if CF_TEXT is requested, we now
post CF_UNICODETEXT and it is converted to CF_TEXT *in the locale of the
requesting process*.  Previously, we would convert to CF_TEXT *in our
locale* and post that.

It looks like the code in the !X_HAVE_UTF8_STRING case didn't actually
work correctly, but fortunately that has never been true...

The original Win32 clipboard API is widely regarded as terrible, since
it relies on clients co-operatively managing the clipboard viewer chain,
and a single buggy client can break it for all other clients.

The last Windows version only supporting that API was Windows XP (5.1),
EOLed in 2014.

(This requires MinGW-w64 w32api 6.0.0 or later for
Add/RemoveClipboardListener correctly exported by the x86_64 user32
implib)

XChangeProperty() requests larger than the ~16MB permitted even with
BigReq will fail BadLength

Also, relax the timeout mechanism so it allows 1 second between events,
rather than 1 second for the entire transfer, as transfers of large
pastes can take more than 1 second.

Also, prefer UTF8_STRING encoding to COMPOUND_TEXT encoding

This option was implemented before the drivers were split in ≈2006,
and e.g. XWin still supports it.

With this commit, Xorg regains support, so that the following configuration can
be used to set the repeat rate for all keyboard devices without having to modify
Xorg command-line flags or having to automate xset(1):

Section "InputClass"
        Identifier "system-keyboard"
        MatchIsKeyboard "on"
        Option "XkbLayout" "de"
        Option "XkbVariant" "neo"
	Option "AutoRepeat" "250 30"
EndSection
Signed-off-by: Michael Stapelberg <stapelberg@google.com>

Initialized to a constant value, never modified, never varied by device.
Signed-off-by: Adam Jackson <ajax@redhat.com>

Initialized to a constant value, never modified, never varied by device.
Signed-off-by: Adam Jackson <ajax@redhat.com>

The GLX_ARB_create_context path (with which this should all get unified,
someday, sigh) already enforces this, but the classic path does not.
It's effectively assumed by the implementation anyway, so let's enforce
it rather than do crashy things.
Reviewed-by: Michel Dänzer <mdaenzer@redhat.com>
Signed-off-by: Adam Jackson <ajax@redhat.com>

glxProbeDriver() concatenates __DRI_DRIVER_GET_EXTENSIONS with driver name
to get symbol name for get_extension function. Unfortunately that doesn't
work for drivers that have hyphen in their name, e.g. sun4i-drm --
get_extensions() for these uses underscore instead.

As result dlsym() doesn't find get_extension() function and AIGLX
initialization fails resulting in following message in Xorg.0.log:

(EE) AIGLX error: sun4i-drm does not export required DRI extension

Replace all non-alpha-numeric characters with underscore to fix the issue.
Signed-off-by: Vasily Khoruzhick <anarsoul@gmail.com>

We were only calling xwl_present_unrealize_window for the toplevel
window, but the list can contain entries from child windows as well,
in which case we were leaving dangling pointers to freed memory.

Closes: #1000
Fixes: c5067fea "xwayland: Use single frame callback for Present
                     flips and normal updates"
Reviewed-by: Olivier Fourdan <ofourdan@redhat.com>
Tested-by: Olivier Fourdan <ofourdan@redhat.com>

Reviewed-by: Michel Dänzer <mdaenzer@redhat.com>
Signed-off-by: Yuriy Vasilev <uuvasiliev@yandex.ru>

This allow x-server to run with -depth 16.
Reviewed-by: Michel Dänzer <mdaenzer@redhat.com>
Signed-off-by: Yuriy Vasilev <uuvasiliev@yandex.ru>

Recently, rooted Xwayland crashes on wlroots-based compositors, because
wlroots removed the deprecated wl_shell protocol.
This MR fixes this by changing the code in question to the xdg-shell
protocol. My motivation do this: on etnaviv-based embedded platforms,
rooted Xwayland is much faster and doesn't cause UI rendering bugs
compared to rootless Xwayland.
Signed-off-by: Lukas F. Hartmann <lukas@mntre.com>

This code block was moved from a function that returns 0 for failure to a
function that returns 0 for Success in commit
649293f6. Change the return value to
BadValue to match the other checks in _XkbSetMapChecks.

Set nTypes to xkb->map->num_types when XkbKeyTypesMask is not set, to
allow requests with the XkbKeyTypesMask flag unset in stuff->present to
succeed.

Fixes a potential heap smash when client->swapped is true, because the
remainder of the request will not be swapped after "return 0", but
_XkbSetMap will be called anyway (because 0 is Success).
Signed-off-by: Peter Harris <pharris@opentext.com>

The server swaps part of the request in _XkbSetMapChecks instead of
SProcXkbSetMap (presumably because walking the XkbSetMap request is hard,
and we don't want to maintain another copy of that code).

Swap the first time _XkbSetMapChecks is called, not the second time.
Signed-off-by: Peter Harris <pharris@opentext.com>

xwayland: Remove unnecessary xwl_window_is_toplevel() check from xwl_output_set_window_randr_emu_props()

Since the recent fix to call xwl_output_set_window_randr_emu_props() from
ensure_surface_for_window(), it is now only called on a toplevel window,
so the is-toplevel check is not necessary for the
xwl_output_set_window_randr_emu_props() case.

This commit moves the check to xwl_output_set_randr_emu_prop_callback()
so that we only do it when we are walking over all Windows of a client
to update the property on a change of the emulated resolution.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>

For window-manager managed windows, xwl_realize_window is only called for
the window-manager's decoration window and not for the actual client window
on which we should set the _XWAYLAND_RANDR_EMU_MONITOR_RECTS prop.

Usualy this is not a problem since we walk all client windows to update
the property when the resolution is changed through a randr call.

But for apps which first do the randr change and only then create their
window this does not work, and our xwl_output_set_window_randr_emu_props
call in xwl_realize_window is a no-op as that is only called for the wm
decoration window and not for the actual client's window.

This commit fixes this by making ensure_surface_for_window() call
xwl_output_set_window_randr_emu_props on the first and only child of
window-manager managed windows.

Note this also removes the non-functional xwl_output_set_window_randr_emu_props
call from xwl_realize_window, which was intended to do this, but does not
work.

This fixes apps using the ogre3d library always running at the
monitors native resolution.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>

Some clients, which use vidmode to change the resolution when going fullscreen,
create an override-redirect window and never trigger the screen->ResizeWindow
callback we rely on to do the xwl_window_check_resolution_change_emulation().

This causes us to not apply a viewport to them, causing the fullscreen window
to not fill the entire monitor.

This commit adds a call to xwl_window_check_resolution_change_emulation()
at the end of ensure_surface_for_window() to fix this. Note that
ensure_surface_for_window() exits early without creating an xwl_window
for new windows which will not be backed by a wayland surface and which
thus will not have an xwl_window.

This fixes ClanLib-0.6.x and alleggl-4.4.x using apps not properly
fullscreening.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>

The code building the mode-list does the following to deal with screen
rotation:

    if (need_rotate || xwl_output->rotation & (RR_Rotate_0 | RR_Rotate_180)) {
        mode_width = xwl_output->width;
        mode_height = xwl_output->height;
    } else {
        mode_width = xwl_output->height;
        mode_height = xwl_output->width;
    }

This means we need to do something similar in xwl_output_set_emulated_mode()
to determine if the mode being set is the actual (not-emulated) output mode
and we this should remove any emulated modes set by the client.

All callers of xwl_output_set_emulated_mode always pass a mode pointer
to a member of xwl_output->randr_output->modes, so we do not need to
duplicate this code, instead we can simply check that the passed in mode
is modes[0] which always is the actual output mode.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>

Not only hook the ResizeWindow method of the screen (which really is
MoveAndResize) but also hook the MoveWindow method for checking if we
need to setup a viewport for resolution change emulation.

Our resolution change emulation check if the windows origin matches
the monitors origin and the windows origin can also be changed by just
a move without being resized.

Also checking on a move becomes esp. important when we move to checking
on changes to the top-level non-window-manager client (X11)Window instead
of on changes to the xwl_window later on in this patch series.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>

The recent change to use the top-level non-window-manager Window drawable
coordinates from xwl_window_check_resolution_change_emulation() in
combination with only calling it on a resize when the top-level window
is moved breaks things with mutter/gnome-shell.

When fullscreening a X11 window, mutter moves its window-decoration Window
wrapping the top-level Window to the monitor's origin coordinates (e.g. 0x0)
last. This updates the top-level's drawable coordinates, but as the
actual MoveWindow is called on the wrapper Window and not on the toplevel
we do not call xwl_window_check_resolution_change_emulation() and we never
enable the viewport.

This commit fixes this by also calling
xwl_window_check_resolution_change_emulation() if the Window being moved
is an xwl_window itself.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>

When a reparented window is resized directly check the emulation instead of
doing this only when the window manager parent window is resized, what might
never happen.

For that to work we need to make sure that we compare the current size of the
client toplevel when looking for an emulated mode.

Changes by Hans de Goede:
- Remove xwl_window x, y, width and height members as those are no longer used.
- Add check for xwl_window_from_window() returning NULL.
Signed-off-by: Roman Gilg <subdiff@gmail.com>
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>

Make window_get_none_wm_owner return the first non-wm-window instead of the
owner (client) of the first non-wm-window and rename it to
window_get_client_toplevel to match its new behavior.

This is a preparation patch for switching to using the drawable coordinates
in xwl_window_should_enable_viewport()

Changes by Hans de Goede:
- Split this change out into a separate patch for easier reviewing
- Rename window_get_none_wm_owner to window_get_client_toplevel to match
  its new behavior
Signed-off-by: Roman Gilg <subdiff@gmail.com>
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>

An X11 window manager might add a chain of parent windows when reparenting to a
decoration window.

That is for example the case for KWin, which reparents client windows to one
decoration and another wrapper parent window.

Account for that by a recursion into the tree. For now assume as before that
all X11 window managers reparent with one child only for these parent windows.

Changes by Hans de Goede:
- Move the xwl_window_is_toplevel() from a later patch in this series here
  as it really belongs together with these changes
- Drop no longer necessary xwl_window argument from window_get_none_wm_owner
  parameters
Signed-off-by: Roman Gilg <subdiff@gmail.com>
Reviewed-by: Hans de Goede <hdegoede@redhat.com>
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>

When a viewport is already created we can reuse this object instead of
destroying it and getting a new one for updating the source rectangle and
destination size.
Signed-off-by: Roman Gilg <subdiff@gmail.com>
Reviewed-by: Hans de Goede <hdegoede@redhat.com>
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>

Instead of iterating over all clients which are listening for events on the
root window and checking if the client we are dealing with is the one
listening for SubstructureRedirectMask | ResizeRedirectMask events and thus
is the window-manager, cache the client-id of the window-manager in
xwl_screen and use that when checking if a client is the window-manager.

Note that we cache and compare the client-id rather then the ClienPtr,
this saves reading the ClientPtr from the global clients array when doing
the comparison.
Suggested-by: Olivier Fourdan <ofourdan@redhat.com>
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>

The value is not the current msc of the window, but the target value
the client sets independently of the window speicific msc offset. Make
this clearer.
Signed-off-by: Roman Gilg <subdiff@gmail.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

Move the code portion down. That way it is at a similar position as in
the window mode file.
Signed-off-by: Roman Gilg <subdiff@gmail.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

Make the code more readable by going through some logical abort
conditions. Also make the function only about updating the crtc
msc value and not about also returning the next target msc.
Signed-off-by: Roman Gilg <subdiff@gmail.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

Unfold and extensively annotate the target-msc adjustment function, to make
it easier to understand what's happening and why.
Signed-off-by: Roman Gilg <subdiff@gmail.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

We can use value arguments instead of pointers when adjusting the timings
by returning the adjusted value. This improves the readability.
Signed-off-by: Roman Gilg <subdiff@gmail.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

Xorg supports the '-version' command line option, add something similar
to Xwayland.

Closes: xorg/xserver#976Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Reviewed-by: Jonas Ådahl <jadahl@gmail.com>
Reviewed-by: Michel Dänzer <mdaenzer@redhat.com>

xf86platformProbeDev didn't check the device path, fix it.

This is a problem when trying to set up a non-PCI device via
explicit xorg.conf.d configuration.

An USB DisplayLink device, being non-PCI was always set up
as a GPU device assigned to screen 0 instead of a regular
framebuffer, potentially having its own dedicated screen,
despite such configuration as below. Only the relevant parts
of the configuration are quoted, it's part of a larger context
with an Intel chip that has 3 outputs:
* DP1 connected to an LCD panel,
* VGA1 connected to an external monitor,
* HDMI1 unconnected and having no user visible connector

Section "ServerFlags"
        Option          "AutoBindGPU" "false"
EndSection

...

Section "Device"
        Identifier      "Intel2"
        Driver          "intel"
        BusID           "PCI:0:2:0"
        Screen          2
        Option          "Monitor-HDMI1" "HDMI1"
        Option          "ZaphodHeads" "HDMI1"
EndSection

Section "Device"
        Identifier      "UDL"
        Driver          "modesetting"
        Option          "kmsdev" "/dev/dri/card0"
        #BusID          "usb:0:1.2:1.0"
        Option          "Monitor-DVI-I-1" "DVI-I-1"
        Option          "ShadowFB" "on"
        Option          "DoubleShadow" "on"
EndSection

...

Section "Screen"
        Identifier      "SCREEN2"
        Option          "AutoServerLayout" "on"
        Device          "UDL"
        GPUDevice       "Intel2"
        Monitor         "Monitor-DVI-I-1"
        SubSection      "Display"
                Modes   "1024x768"
                Depth   24
        EndSubSection
EndSection

Section "ServerLayout"
        Identifier      "LAYOUT"
        Option          "AutoServerLayout" "on"
        Screen          0 "SCREEN"
        Screen          1 "SCREEN1" RightOf "SCREEN"
        Screen          2 "SCREEN2" RightOf "SCREEN1"
EndSection

On the particular machine I was trying to set up an UDL device,
I found the following structure was being used to match
the device to a platform device while I was debugging the issue:

xf86_platform_devices[0] == Intel, /dev/dri/card1, primary platform device
xf86_platform_devices[1] == UDL, /dev/dri/card0

devList[0] == "Intel0", ZaphodHeads: DP1
devList[1] == "Intel1", ZaphodHeads: VGA1
devList[2] == "UDL"
devList[3] == "Intel2", ZaphodHeads: HDMI1 (intended GPU device to UDL)

When xf86platformProbeDev() matched the UDL device, the BusID
check failed in both cases of:
* BusID "usb:0:1.2:1.0" was specified
* Option "kmsdev" "/dev/dri/card0" was specified

As a result, xf86platformProbeDev() went on to call probeSingleDevice()
with xf86_platform_devices[0] and devList[2], resulting in the
UDL device being set up as a GPU device assigned to the first screen
instead of as a framebuffer on the third screen as the configuration
specified.

Checking Option "kmsdev" in code code may be a layering violation.
But the modesetting driver is actually part of the Xorg sources
instead of being an external driver, so he "kmsdev" path knowledge
may be used here.
Signed-off-by: Böszörményi Zoltán <zboszor@pr.hu>