systemd tries to bring down the world when going from
the initramfs to /
commit 9e5a276f32 tried
to prevent plymouthd from getting killed by settings
argv[0][0] to @
This isn't seemingly sufficient. Throw some lines into
the plymouth-start service file that should hopefully help.
The previous systemd commit introduced a file named
systemd-ask-password-plymouth.path.in
The makefile was only looking for .service.in files
when stripping the .in suffix, so it got installed
incorrectly.
This commit fixes up the Makefile.
When plymouth service files were moved from systemd to plymouth
two files got lost in the shuffle.
This commit adds them.
http://bugzilla.freedesktop.org/51573
plymouth uses a shadow framebuffer to cache screen contents for
quick compositing. This shadow framebuffer may or may not have
the same memory layout as the hardware framebuffer. In cases
where the size and layout of pixels are the same between the shadow
framebuffer and the hardware framebuffer we can memcpy()
the pixels line-by-line to the hardware. If the width of area being
flushed is the same number of bytes as the width of the hardware buffer,
then we can memcpy() the entire flush area in one call to memcpy.
The check for this latter fast-path has a miscalculation that tests
the number of pixels in the flush area width to number of bytes in the
buffer width. This commit adds the * 4 multiplier to correctly compare
bytes with bytes instead of pixels with bytes.
This commit also adds a sanity check to make sure the byte size of the
hardware framebuffer width is equal to the advertised row stride.
The udev trigger calls that are there, are actually
spurious. udev will coldplug those subsystems at
start up anyway, so doing it explicitly is wrong.
This commit drops those trigger calls.
What does matter is the udevadm settle call. It's
what makes things block until the graphics driver
is loaded.
udevadm settle is a big sledgehammer, though. It blocks
until all the triggered events (even stuff unrelated to
graphics) are finished.
This commit adds a --exit-if-exists argument to udevadm settle,
so it will bail early as soon as the graphics devices
are up.
plymouth-start.service does this sort of hacky
"udevadm trigger" stuff before doing plymouth show-splash,
to ensure plymouth show-splash is called after the
graphcis subsystem is up.
It actually does two calls:
- one call that triggers any pci devices with the class
0x030000 (which is "vga compatible display device")
- another call that triggers the gpu subsystem
The first call is borrowed from dracut:
http://dracut.git.sourceforge.net/git/gitweb.cgi?p=dracut/dracut;a=commitdiff;h=2c02c8318
and I can't find any historical context on why its needed. As I
understand things, the latter should be a superset of the former.
Furthermore, the first trigger is missing a --subsystem-match=pci call
so it's matching the "class" attribute in every subsystem, which is slow.
I'm going to drop the first trigger until I start hitting problems
and need to add it back.
plymouth-start.service does this sort of hacky
"udevadm trigger" call before doing plymouth show-splash,
to ensure plymouth show-splash is called after the
graphcis subsystem is up. Because this udev trigger call
only passes --subsystem-match calls with no corresponding
--subsystem call, it will trawl through all subsystems in
/sys rather than just the pci subsystem (where all the matches
are guaranteed to be).
This commit adds --subsystem=pci to stop doing that extra work.
All of this will eventually be replaced with plymouthd either
listening for the udev events itself (or potentially logind,
if logind gains a way to tag a "graphics" capability on a seat).
get_os_string either returns a string describing the distribution,
or NULL on error.
This NULL later causes problems since we run strlen on it.
This commit makes get_os_string return "" instead of NULL for error
cases.
Spotted by Andreas Henriksson.
Having the dependency is awkward and causes problems, so
don't bother.
This commit just copy and pastes the relevant bits for
inst().
Ideally, this would be an installroot command implemented in
C in util-linux, but shrug
Right now, we do the end animation immediately if in
shutdown mode. This is because shutdown is fast,
and we don't try to estimate how long it will take.
The code depends on is_idle being true after the end
animation is run, but we neglect to do that in the
shutdown case.
This commit makes shutdown mode just call
"become_idle" right away, and deletes the duplicated
shutdown code path.
When plymouth's VT is not the foreground VT, we tell the
kernel not to scan out from the kernel buffer we manage.
Also, we don't bother dispatching rendering to the kernel
buffer (since it's not getting scanned out anyway).
When plymouth's VT becomes the foreground VT, we tell the
kernel to start scanning out from our buffer again.
Unfortunately, we neglect to flush all the pending drawing
that happens while VT switched away.
This means we briefly show stale contents.
This commit flushes all pending rendering to the kernel,
before resetting the scan out buffer, so we get a current
view of the splash shown immediately.
In the shutdown path we start out on a different
vt than where the splash is drawn. Make sure we
don't set a mode until the jump.
This prevents unnecessary flicker.
This reverts commit 3f14f2700e.
Forcing frame-buffer plugin makes multi-monitor look really bad.
Instead, depend on the init system to make sure we stay alive
until the end.
Systemd will cleanse the system of running processes
in the hand over from the initrd to the main root filesystem,
and at shutdown.
In both cases we want to keep on chugging, so we tag ourselves
in a way that systemd won't kill us.
See http://www.freedesktop.org/wiki/Software/systemd/RootStorageDaemons
The drm driver tells us the minimum dimensions it supports
for buffer objects. We use this minimum for creating a
small temporary 32-bit buffer to test if 32-bit buffers
are supported.
Unfortunately, some drivers neglect to fill in
min_width/min_height and then we try to allocate a buffer
with 0 sized dimensions.
This commit checks for min_width/min_height being 0, and then
bumps them to 1.
Minor changes to initial patch by Ray Strode.
In some cases, clients may need a way to ensure that all queued
messages have been flushed and are sent to the daemon
(before, for instance, exiting).
This commit adds an API to block until the outgoing request queue is
empty.
This commit adds four keyboard shortcuts to the developer
documentation that were found from reading the source:
Esc: Toggle between system console and plymouth boot animation.
Ctrl-V: Toogle verbose mode on and off.
Ctrl-U or Ctrl-W: erase a line
This commit starts to document plymouth to help
new contributors get into the code.
The aim of the document is to provide useful information,
such as the overall architecture, the most
important data structures, and howto recipes for
typical user cases, like debugging.
The document is explicitly not meant to be detailed API documentation.
In the future, that type of documentation may be provided with gtk-doc
style annotations.
The documentation is written in asciidoc
(http://www.methods.co.nz/asciidoc/) and is therefore easily readable as
its ascii text and can also be translated into more rich formats
(like HTML).
For example, to get an HTML version of the document:
yum install asciidoc (or equivalent for your distribution)
cd docs
make development.html
When the computer is progressing through its boot up process, plymouth
calls into the splash plugin's on_boot_progress function at regular
intervals with increasing values for "percent_done". At some point, it
gets to 90% done, and that's when two-step begins its finishing
animation sequence. As soon as this sequence finishes, two-step pulls
its stop trigger, which
1) sets its "is_idle" flag to true and
2) pulls the core plymouthd code's idle trigger, to notify that
code that it's at a good animation frame to quit (if the core
plymouthd code has an idle trigger set up)
During the boot process, the user may need to enter a password
(the "plymouth ask-for-password" command). When that happens,
the splash waits for the user to enter a password, but boot progresses
in the background.
If the user then enters a password, the boot animation restarts again
(from the display_normal function). This restarting of the boot
animation will cause the "is_idle" flag of the splash to get set back
to false.
Later when plymouthd wants to quit, it calls the become_idle function
of the splash plugin. That function notices "is_idle" is false, and
the stop_trigger is not NULL. The function isn't suited to work
with this combination, and so at this point the splash never
tells the code daemon code it's idle.
This commit changes on_boot_progress to return before looking at
percent_done, if the user is getting asked a question. This way
the stop_trigger won't get created prematurely, and is_idle won't
get out of sync.
https://bugs.freedesktop.org/show_bug.cgi?id=49355
Most distributions no longer use this transition, which relies
on plymouth quitting before X starts.
For clarity, rename the #define to include the word DEPRECATED
Since around 2.6.39, the kernel has offered a generic ioctl interface
for modesetting (the KMS 'dumb' interface). This interface is now
provided by most all of the modesetting drivers.
Adding support for this interface means that plymouth will automatically
gain support for new modesetting drivers going forward.
This commit adds that support. Some changes made by Ray Strode.
A ply_label can now be alignment with ply_label_set_alignment()
taking one of PLY_LABEL_ALIGN_LEFT, PLY_LABEL_ALIGN_CENTER, or
PLY_LABEL_ALIGN_RIGHT.
This alignment is made within a horizontal box which width can be set with
ply_label_set_width().
Both functions are now used by the plugin "two-step" to make sure, that
the prompt for "ask-for-password" gets centered on the screen. Previously
the prompt started in the middle of the screen and was cropped at the right
border of the screen (fixes bz #681513). This lead to unreadable prompts
for disk encryption.
The previous commit tries to create a 32bpp 1x1 dummy buffer
to know whether or not the modesetting driver can handle
our needs. It's concievable (though not realized in practice) that
the driver could support our needs fine, but not support buffers
as small as 1x1.
This commit uses the queried minimum buffer dimensions for the dummy
buffer instead of assuming 1x1.
Now that more and more hardware is getting kms support,
we need to be more careful about what hardware we try to
run on. Cirrus cards can't make 32bpp framebuffers, which
is all we support in the drm renderer.
This commit tries to create a small buffer up front, just
to see if the driver will allow it. If the driver chokes,
then we bail and fall back to the /dev/fb renderer, which
is more tolerant of aging hardware.
