For some reason I made the same api misdesign with ply_key_file_t
that I made when writing GKeyFile...it returns duplicated memory for
no good reason.
This commit sprinkles frees around.
If a device gets added when we're already deactivated, plymouth shouldn't
process the device, since processing it effectively activates plymouth.
This commit pulls the udev monitor fd out of the event loop while
plymouth is deactivated so new events are deferred until reactivation.
Modified by Iain Lane <iain.lane@canonical.com>: Also deactivate the
timer that finds all devices known to udev after an interval, when
paused.
If plymouth doesn't get explicitly "activated" then when
GDM tries to deactivate it, the deactivation request is
a noop.
One aspect of being active, though is having ownership and
control of the terminal. This happens immediately, even
before a splash is shown.
The `deactivate` request needs to relinguish such control,
unconditionally, since some display server is about to use
the tty.
This commit fixes that.
When running in a container with plymouth installed, plymouth is started
unnecessarily and systemd prints warnings:
[ OK ] Reached target Shutdown.
Sending SIGTERM to remaining processes...
Sending SIGKILL to remaining processes...
Process 253 (plymouthd) has been been marked to be excluded from killing. It is running from the root file system, and thus likely to block re-mounting of the root file system to read-only. Please consider moving it into an initrd file system instead.
It makes little sense to start plymouth in contains, so add
'ConditionVirtualization=!container' everywhere where
ConditionKernelCommandLine=!plymouth.enable=0 appears to disable plymouth
in containers.
BugLink: https://bugzilla.redhat.com/show_bug.cgi?id=1337611
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Closes: https://gitlab.freedesktop.org/plymouth/plymouth/issues/27
Fix all compiler warnings (and one type) except for the deprectation warnings
for the gdk_screen_get_monitor_* functions used in the x11 renderer.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
plymouth uses function type casts for callbacks in quite a few places, fixing
these needlessly complicates the code, so lets pass -Wno-cast-function-type.
This fixes 218 warnings like this one:
ply-command-parser.c: In function ‘ply_command_parser_stop_parsing_arguments’:
ply-command-parser.c:680:48: warning: cast between incompatible function types from ‘void (*)(ply_command_parser_t *)’ {aka ‘void (*)(struct _ply_command_parser *)’} to ‘void (*)(void *, int, ply_event_loop_t *)’ {aka ‘void (*)(void *, int, struct _ply_event_loop *)’} [-Wcast-function-type]
(ply_event_loop_exit_handler_t)
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
This are 3 issues with the detailed logs handling:
1) plymouth attaches to the session directly on a show-splash command
(in on_show_splash()), but it does not tell systemd to start printing
details until the splash is actually shown after the splash_delay.
2) If the splash is actually shown during the initrd (e.g. a diskcript
password is necessary) then we tell the initrd systemd instance to
print details, but we don't tell the regular initrd instance which takes
over as pid 1 after the switch-root to print details.
This leads to rather inconsistent logging/printing behavior, e.g.:
* If a diskcrypt password is asked for, we only log details from
the initrd phase.
* If the boot is shorter then splash_delay no details are logged
* If the user presses ESC during boot during the initrd, only initrd
messages are printed
* If the user presses ESC during boot after the initrd, only normal
messages are printed
This commit fixes both these issues by:
1) Telling systemd to print details as soon as we have attached to the session;
and to stop printing details when we detach from the session (*)
2) Telling systemd to print details after the rootfs has been remounted rw
*) This is necessary to have a smooth transition to e.g. gdm if the splash
has not shown because the boot is shorter then splash_delay
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Start listening for keypresses on the first show_splash() call, so that
pressing ESC while we're delaying show the non-details splash will show
the details splash.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
ply_renderer_head_map() gets only called from map_to_device() which
calls activate() directly afterwards which calls
ply_renderer_head_set_scan_out_buffer(), so there is no need for the
reset_scan_out_buffer_if_needed() call.
Not only is it not needed, but it is actually harmful, there are 2 problems
woth it:
1) Normally the drm plugin gets instantiated by ply-renderer.c with
rendered->is_active=true, backend->is_active=false. The
rendered->is_active=true causes the first ply_renderer_activate call
to be a no-op without calling backend->activate(). So when the first
map_to_device() calls happen activate() has not been called yet and we've
not yet claimed master rights, so ply_renderer_head_set_scan_out_buffer()
calls will always fail, resulting in this in a ply-trace:
Mapping buffer for 1920x1080 renderer head
Redrawing 1920x1080 renderer head
Setting scan out buffer of 1920x1080 head to our buffer
Couldn't set scan out buffer for head with controller id 41
This is harmless, but also shows that the reset_scan_out_buffer_if_needed()
is really not needed.
2. If deactivate_renderer() gets called before the first show-splash then
rendered->is_active will become false, so renderer_activate() done before
map_to_device() will now actually call backend->activate() claiming
drm master rights and setting backend->is_active=true.
The map_to_device() -> ply_renderer_head_map() call done after this, calls
ply_renderer_head_redraw() -> flush_head() which under 1. was a no-op
as it exits directly when backend->is_active=false. But now it actually
flushes the buffers by calling reset_scan_out_buffer_if_needed(). This
itself is fine.
But since reset_scan_out_buffer_if_needed() has already happened in
ply_renderer_head_redraw() the reset_scan_out_buffer_if_needed() call this
commit removes would always return false (no reset necessary) causing
ply_renderer_head_map() to destroy the buffer and return an error.
This results in the splash briefly showing, followed by the core soon after
trying another map_to_device(), which again briefly shows the splash, etc.
With the end result being a badly flickering display.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Since commit eb147e52b1 ("renderer: support reactivating renderer without
closing it first"), the show_theme() call done by
toggle_between_splash_and_details() will reactivate the renderers after
switching to details mode, causing the drm renderer to switch the screen
from text to graphics mode hiding the details being logged on the console.
This commit fixes this by only calling ply_device_manager_activate_renderers()
and ply_device_manager_deactivate_renderers if the splash uses pixel-displays.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
https://bugs.freedesktop.org/show_bug.cgi?id=107047
hide_splash() should be the counter-part of show_splash(). show_splash()
calls ply_device_manager_activate_renderers() (through show_theme()).
2 of the 3 callers of hide_splash() are already calling
ply_device_manager_deactivate_renderers() directly before calling
hide_splash(). This commit moves the deactivate call into hide_splash()
so that it also gets called from the 3th code-path, which is when
the user hits the escape to key to toggle from the splash to details.
It's important that plymouth deactivates its renderers before going
to details, because those renderers can block the kernel from
initializing fbcon, which the kernel will start doing lazily in the
future:
https://lkml.org/lkml/2018/6/26/489.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
https://bugs.freedesktop.org/show_bug.cgi?id=107047
At the moment, ply_renderer_activate() doesn't work immediately following
ply_renderer_deactivate(). This is because the renderer isn't marked
inactive until it's closed.
This commit marks the renderer inactive when it's deactivated.
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
https://bugs.freedesktop.org/show_bug.cgi?id=107047
If neither "rhgb" nor "splash" is on the kernel cmdline, then
plymouth forces the "details" splash. This splash is merely
a passthrough plugin, where it makes boot looks like plymouth
isn't even running.
In this case, the code sets PLY_DEVICE_MANAGER_FLAGS_IGNORE_UDEV.
The idea is to not bother waiting for udev events notifying
plymouth when graphics devices show up, since it doesn't need
to use the grpahics devices directly anyway.
Unfortunately, it does still erroneously try to setup graphical
renderers in this case, including the /dev/fb renderer.
Before commit e4f86e3c, these graphical renderers failed because
they were given the wrong device name, but since that fix, they're
suceeding. We definitely don't want the /dev/fb renderer to
load if we're ignoring udev on efi systems, since during very
early boot /dev/fb is backed by efifb, something we never want to
use. efifb is supposed to get replaced during the boot process
by other fb implementations like say radeondrmfb, virtiodrmfb or
bochsdrmfb, and some of those implementations can't handle the
transition if /dev/fb is open at switchover time.
This commit adds a new flag to tell the device manager to
not bother trying to setup graphical renderers when details are
forced.
http://bugzilla.redhat.com/1518464
The used name for the feature "drm" results in the configure option
"--enable-drm" and not "--enable-drm-renderer".
Signed-off-by: Ulrich Ölmann <u.oelmann@pengutronix.de>
In order to hand off the drm fd to another program we need
to be able to run that other program.
This commit adds an API for running a program with no arguments,
in the background.
Right now, anytime we update the root fs in the daemon we treat
it like we switched from initramfs to the main root filesystem.
We shoudn't do that when shutting down since we're going to go
the other direction.
This commit changes on_newroot to look at the current mode to
decide what to do.
For the moment it doesn't do anything in the shutdown case, but
that will change shortly.
plymouthd currently sticks around on the main filesystem during
shutdown, while shutdown proceeds on the initramfs.
This commit adds a unit to make it jump into the initramfs too.
This is important, so we can run the drm escrow binary from
the initramfs.
Right now plymouthd forces itself to stay alive for the duration of
shutdown, so the splash screen can stay up until the power is killed.
This causes unclean mounts in some situations following system updates.
Rather than keep plymouthd around, all we really need to do is keep
the drm fd around. That can be down with a separate program executed
just in time from the initramfs.
As a first step toward acheiving that goal, this commit creates a
simple new program "plymouth-drm-escrow" that doesn't do anything
at all but try to stay alive.
A future commit will change plymouthd to run plymouth-drm-escrow,
and pass it the drm fd.
On some devices the LCD panel is mounted in the casing in such a way
that the up/top side of the panel does not match with the top side of
the device.
Examples of this are:
-Tablets where the LCD panel is mounted upside-down (various models)
-Clamshell design devices use portrait tablet screens in a landscape fashion /
a landscape case, e.g. the GPD pocket, GPD win, Asus T100HA and the VIOS LTH17
(so the didplay is mounted rotated 90 degrees clock-wise or counter-clock-wise).
Starting with kernel 4.16 there is code in the kernel to:
-Detect affected devices
-Automatically rotate the fbconsole to compensate
-Let userspace know about this with a new "panel orientation" property on the drm connector of the panel
The patch series I'm attaching to this bug adds support for this to
plymouth, so that the boot-splash and diskcrypt-ask-password are shown
with the correct orientation / the right way up on these devices.
On devices where the (LCD) panel is mounted upside-down in the case
the kernel's drm_fb_helper code may have set up rotation on the primary
plane to make the text-console (and other fbdev using apps) show the right
way up.
We inherit this rotation from the text-mode and since we do our own rotation
where necessary we end up rotating twice and showing the boot-splash
upside-down again.
Dealing with hardware rotation may require using a specific framebuffer
tiling which we do not support, so we should just disable the hardware
rotation and keep using our own software rotation.
This commit adds code to find the primary plane and its rotation property
and if it is not DRM_MODE_ROTATE_0 then sets it to DRM_MODE_ROTATE_0. fixing
the double rotation issue.
https://bugs.freedesktop.org/show_bug.cgi?id=104714
On some devices the LCD panel is mounted in the casing in such a way
that the up/top side of the panel does not match with the top side of
the device (e.g. it is mounted upside-down).
Kernel 4.16 introduces a new "panel-orientation" property on the drm
connector which allows modesetting applications / code to check for
such LCD panels.
This commit adds support for this new property and passes this to the
pixel_buffer code using the new ply_pixel_buffer_new_with_device_rotation
method, so that the pixel_buffer code will automatically rotate the
image to correct for the panel orientation.
https://bugs.freedesktop.org/show_bug.cgi?id=104714
On some devices the LCD panel is mounted in the casing in such a way
that the up/top side of the panel does not match with the top side of
the device (e.g. it is mounted upside-down).
This commit adds support to the ply-pixel-buffer code to create
buffers which take device rotation into account and which will rotate
the picture to compensate.
https://bugs.freedesktop.org/show_bug.cgi?id=104714
On machines with a slow CPU (Atom) and a highres screen drawing the
diskcrypt dialog may take longer then the keyrepeat speed, this leads to
a long delay before showing keypresses when doing the following:
1) Type long password
2) Realize it is wrong, press + hold backspace
the key-repeat will now generate backspace key presses faster then we
process them as main.c does an update_display for each press
3) Users releases backspace when we've processed input-length backspace
key-presses, but since we were drawing slower then key-presses were
coming in many more backspace keypresses are in the keyboard buffer
4) User types first character of the right password, this shows up up to
a couple of seconds later because first we are still processing all
the queued up backspace presses and doing a redraw for each.
This commit fixes this by skipping the redraws in on_backspace when there
is no more input left in the input buffer.
https://bugs.freedesktop.org/show_bug.cgi?id=104714
