Better Startup Experience (Graphical Boot Sequence)
The startup experience needs to be flicker-free, seamless and shiny. To do this we are getting rid of RHGB and writing a new program, Plymouth, that starts earlier (even before / is mounted!), doesn't require an X server, and gets rid of a lot of the noise during startup.
Plymouth will requires DRM kernel modesetting drivers to get pretty graphics, but will have a text mode fallback for systems without driver support.
- Name: RayStrode, KristianHoegsberg, JonMcCann, PeterJones, DavidAirlie, JeremyKatz et al
- Targeted release: Fedora 10
- Last updated: 2008-07-08
- Percentage of completion: 52%
F9 had kernel modesetting for Intel hardware. Ray has been working on a prototype rhgb replacement called Plymouth. It was put on hold while kernel modesetting was maturing, and has been revived after F9. Modesetting is again in flux, though, in rawhide, since there are two memory managers in play. These are the long anticipated TTM memory manager from tungsten graphics, and the newer GEM memory manager from Eric Anholt at Intel that supercedes TTM. Kristian is trying to get GEM into rawhide soon.
The current plymouth sources can be found here: http://gitweb.freedesktop.org/?p=plymouth;a=summary . There are plymouth packages in rawhide already, modesetting drivers, aren't, so most people won't be able to see graphical boot up yet. Having said that, we enable the text plugin by default now. Anyone using rawhide will see it, although it's sort of ugly and flickery. Live images now also run plymouth in their initrd.
The new startup sequence (including quiet kernel, quiet grub, plymouth, smooth transition to the login screen) has been demonstrated at the Red Hat Summit and was shown informally to various people at FUDcon.
Requires changes to grub, mkinitrd, initscripts, gdm, X, possibly other packages. rhgb will be replaced by Plymouth. The extent of this project is getting a graphical boot sequence that is polished and professional looking.
Boot up could be made faster, too, but that issue is orthogonal and should be tracked separately .
- Verify that holding a key during power-on gets you to the grub menu
- Verify that boot without rhgb in the kernel cmdline gives a traditional text mode boot sequence
- Verify that failure to set a mode in initrd falls back to a clean and simple text mode boot sequence
- Verify that the graphical boot sequence contains only a single mode switch from blank text mode to graphical mode, and that there is a smooth transition from the startup animation to the login screen
- Verify that entering encryption passwords works both in the graphical case and in the fallback case
- Verify that specifying a mode in the kernel commandline works
- Verify that failure messages from service startup can inspected after boot
- Verify that dropping to a shell on critical errors works
- Verify that pressing Escape in plymouth brings you to the detailed mode showing boot messages
This feature depends on plymouth getting into F10. For other packages involved in realizing this feature, see Scope.
For a smooth and shiny experience, we want
- the mode should be changed early from text mode to the most optimal graphics mode available
- the mode should not be changed more than once
- there should be no other uncontrolled transitions
- there should be no text messages, unless the user presses Escape to see them
- there should be good, smooth transitions
The fallback if mode-setting fails or is unavailable is text mode.
Timeline of the current bootup
Plan for improving this
- do away with mode switch for grub menu by default (ie, don't show the image unless "needed")
- Kristian currently has a patch for this. You press and hold ctrl (or any key) during boot up and the menu appears. It's in rawhide now, but the default configuration isn't making use of it.
- make kernel, nash and init obey the quiet option
- There are patches for these, though the kernel keeps sneaking messages in.
- switch to graphical mode in the initrd, draw an animation and keep it running until gdm starts up
- We have an animation but need there are some changes we need to make to the X server and GDM to keep the transitions smooth
- rhgb goes away
- rhgb is already gone
- make the X server started by gdm take over the existing vt, mode and framebuffer content
- krh has a patch which adds a -nr ("no root") option which makes the root window keep the framebuffer contents that were around when X started
- improve the transitions animation -> gdm -> session
- firstboot runs on the gdm X server
This lets us start with a blank screen in text mode, then a single switch to graphics mode, then an animation, then the login screen fades in.
One complication is support for encrypted disks. This may require us to ask for a password in the initrd, before we have X or toolkits. The proposal for acceptable input-handling (at least as far as keyboard layouts are concerned) is to write an xkb-to-console-converter.
- Make grub not display a splash image unless the user holds down ctrl to get to a menu. Grub will have a near zero timeout, but users will be able to hold the key down at the bios before grub shows up.
- Make grub not report which choice was selected if the user didn't pick a choice #239760
- Make drm drivers/mode-setting happen in the upstream kernel (see http://www.ussg.iu.edu/hypermail/linux/kernel/0705.2/0893.html)
- Put drm drivers in the initrd
- If rhgb is in the kernel commandline, switch mode in initrd, use drmfb on vt7
- If rhgb is not in the kernel commandline, we still run plymouthd and it handles input but we just don't do the graphical plugins. This lets us consistently be able to count on plymouth being available
- Put animation drawing app in the initrd, keep initrd mounted until that app exits
- For selecting the mode, make s-c-d set a kernel commandline like rhgb:mode=1600x1200 or similiar
- Open question: what about multihead ? answer for now: all heads painted in background color except one chosen randomly with animation
- Display animation all the way from initrd until gdm comes up, keep initrd mounted until it exits
- Redirect init output early on to a pseudo-terminal
- Have something that buffers the output, can buffer it on the initrd
- If something happens that drops us to a shell (filesystem errors), hook up a framebuffer console with the pseudo-terminal
- If there are warnings in the saved bootup messages, show a "Show boot messages" button on the login screen and/or in the notification area
- Need to start X on the existing vt, not set mode (unless we are still in text mode), preserve existing framebuffer content
- Coordinate session startup. fade in ? might be able to do something nice if we have compositing
- krh to get GEM in rawhide. GEM is a key component to getting shared graphics between X and plymouth.
- Updating the kernel. We need to update the drm module shipped in the kernel with the drm that ships in the gem-modesetting branch of Mesa.
- Update libdrm. We need to update the libdrm userspace interface to the kernel module again from the gem-modesetting branch of Mesa.
- We need to update the X server intel dri driver from gem-modesetting branch.
- we need to update rc.sysinit to call plymouth --hide-splash before jumping to a sulogin. Right now plymouth keeps going on top of the sulogin, so the user never knows what's happening and key presses are read both processes. rc.sysinit goes to sulogin in a few places. Also, we need to --show-splash after the login finishes (if it finishes). Patch should get sent to Bill. --hide-splash and --show-splash may need to some bugfixes wrt to input handling, not sure. --hide-splash should close the tty and --show-splash should make sure it gets reopened. It may already be right, don't remember.
- We need to update mkinitrd to use plymouth ask-for-password --command="cryptsetup luksOpen ..." instead of doing plymouth --ask-for-password > password.txt; cryptsetup luksOpen ... password.txt. The new --command interface allows plymouth to call cryptsetup in a loop if the user mistypes password. Peter is looking into this. We probably don't need to make a similiar change for mkliveinitrd because we don't encrypt live images normally.
- plymouth needs to be ported to use libdrm api for modesetting. It currently uses standard kernel ioctls on /dev/fb for getting a front buffer to render to. By using libdrm we'll be able to set the mode intellibly in multihead environments, place X's mouse cursor off center before X is even started, etc. This api is still in flux, though. It's important we use the API exported by the gem branch. This should happen almost entirely in ply-frame-buffer.c. The old vesafb interface should remain as a fallback. Ultimately, either api will end up with a block of memory that is the scan out buffer that goes to the screen. Either api will keep the shadow buffer code there that's in a well known format. At flush time that shadow buffer is converted into a format that's compatible with the scanout buffer.
Any cursor setting api logically belongs in ply-window.c, but the logistics may make that hard. If we hide vesafb versus drm modesetting entirely in ply-frame-buffer.c, and cursor setting only works with drm modesetting, then we'll need to provide some sort of hook from ply-frame-buffer (say ply_frame_buffer_set_cursor_position). It might make sense to have a wrapper api in ply-window that just calls the frame buffer api, or maybe just have splash plugins use the frame buffer api directly. Once the cursor setting api is in place, we should make sure the cursor gets moved below the logo on screen in splash plugin and make sure that the X server doesn't recenter the cursor when it starts.
- plymouth's password entry needs to be smarter about boundries. You can currently type past the edge of the entry. What should happen first is the entry shouldn't be done ad hoc in the splash plugin. It should be moved to a libplybootsplash like the throbber control. Then it detect if the characters are out of bounds and only show the tail of the typed text. It might make sense to clip one character in half at the very beginning to make it clear that there is some clipped text. The entry api should probably take the images used to draw the entry box and the bullets in so that it's configurable for different splash plugins. Note right now we add a little padding to the entry box image. That will make it difficult to compute how many bullets can fit before we need to clip some bullets. We should strip that padding off and just use a separate spacer image from within the splash plugin. One other problem we'll run in to is there is no mechanism for controls in libplybootsplash to clear to the background of splash plugins. There should be some api added to ply-window.c like ply_window_draw_area or some such that will redraw an area of the screen. ply-window will also need a ply_window_set_draw_handler function or some such that lets' the splash plugin register the draw handler. As an added bonus, we can make ply-window watch for ctrl-L to perform a full screen redraw.
- rc.sysinit also needs to use plymouth ask-for-password when it asks for passwords. Bill is looking into this. It's for unlocking auxillary devices like encrypted /home partitions. One complication with these auxillary devices is we need to let the user know which device it is. Plymouth currently has no way of doing that. This leads to two prerequisite TODOS:
- We need to make plymouth support font rendering after root fs is mounted, so auxillary encrypted block devices can have a prompt string. At a minimum we should have limited ascii support, but we probably want to dlopen pango/cairo and draw text to an image surface that's copied over to the framebuffer. This could all happen as an implementation detail in a new ply-text.c. If we're going to be using cairo *anyway* though, we may want to consider shipping it in the initrd and making a stock feature for splash plugins to use. Hard to say what's right. It's a bit of a slippery slope, because if we gain a cairo dependency in the initrd then we'll also gain a pixman dependency which could give us cpu accelerated pixel conversion operations when going from the shadow buffer to the scanout buffer.
- We need to add a -hidecursor (ajax already has a patch from someone). The idea is we don't want to show an "x" cursor when X first starts before GDM loads.
- We need to add a -nr (don't clear frontbuffer, krh has a patch) options to X to make the transition from plymouth to X smoother.
- We need to add a vt7 argument to the X server so X starts on vt7 even though that vt is already open from plymouth
- We need to update GDM to use those option. One problem with this is, we only want to use the vt7 option for the login greeter. We don't want it for user switching greeters. It's tempting to turn on factory mode in the GDM source code and just make the factory X server use vt7 while all other non-factory greeters wouldn't get it.
- We need to update GDM to not set it's background, and instead just use what was shown during plymouth
- We need to make plymouth always run, even if rhgb isn't on the command line, so we still get boot logging and nash's daemonize command doesn't get confused.
- Turn on silent grub mode. We have all the bits in rawhide. We just need to adjust the default config to use a timeout of 0.
- Make sure grub patch works with efi, too. This would probably have to be either Peter or krh, since Peter knows about EFI, krh wrote the silent patch for legacy bioses, and both have efi hardware.
- Need to run art by fedora-art-list for opinions and potential changes. They may have some useful suggestions on making the process more polished.
- Need to figure out what to do on shutdown. This probably means having a plymouth client command to activate the daemon and show splash
- Need to figure out what to do on wake up from hibernate. When coming out of hibernate we'll get a brief period where we can show animation, but we need to end on a still image. We need to make sure the kernel doesn't clear that still image if possible. That might be tricky.
- Right now we get events about boot from the legacy init scripts. Might be nice to get events in a more upstart friendly way. The latest upstream upstart gives limited events as jobs start, finish starting, and stop.
- Might be good to detect a failure during boot up and show it to the user at the login screen via a notification area icon. For now this would. Right now on failure rhgb-client --details=yes is called which is a noop. One idea is to make it hard link /var/log/boot.log to /var/spool/somewhere/something . Then there would be a small app that is started in the greeter via /usr/share/gdm/autostart/LoginWindow that would look at the spool directory on startup, read any file that's there into memory, remove it from the spool directory and display an icon in gdm's notification area. If the icon gets clicked on the program would display a little vte widget with the buffered boot log.
- Investigate why kernel messages are no longer getting redirected to boot.log and instead overwriting text plugin. Kernel messages go to the current active console which should be a pty plymouth owns. I think when we switched to upstart this broke--Bill mentioned the other day he saw kernel messages on top of the text plugin which shouldn't happen. Maybe upstart is calling the TIOCCONS breaking us.
If things don't work out as expected, we can bring back rhgb from the dead.
The release notes need to explain how to get to the grub menu by holding a key during power-on. Should also explain that the graphical mode of plymouth requires kernel modesetting and will only work on certain hardware (likely only intel for F10). Also, switching to detailed mode with Escape should be mentioned.