The Boot Loader Specification
TL;DR: Currently there's little cooperation between multiple distributions in dual-boot (or triple, ... multi-boot) setups, and we'd like to improve this situation by getting everybody to commit to a single boot configuration format that is based on drop-in files, and thus is robust, simple, works without rewriting configuration files and is free of namespace clashes.
The Boot Loader Specification defines a scheme how different operating systems can cooperatively manage a boot loader configuration directory, that accepts drop-in files for boot menu items that are defined in a format that is shared between various boot loader implementations, operating systems, and userspace programs. The target audience for this specification is:
- Boot loader developers, to write a boot loader that directly reads its configuration at runtime from these drop-in snippets
- Distribution and Core OS developers, in order to create these snippets at OS/kernel package installation time
- UI developers, for implementing a user interface that discovers the available boot options
- OS Installer developers, for setting up the initial drop-in directory
Why is there a need for this specification?
Of course, without this specification things already work mostly fine. But here's why we think this specification is needed:
- To make the boot more robust, as no explicit rewriting of configuration files is required any more
- To improve dual-boot scenarios. Currently, multiple Linux installations tend to fight over which boot loader becomes the primary one in possession of the MBR, and only that one installation can then update the boot loader configuration of it freely. Other Linux installs have to be manually configured to never touch the MBR and instead install a chain-loaded boot loader in their own partition headers. In this new scheme as all installations share a loader directory no manual configuration has to take place, and all participants implicitly cooperate due to removal of name collisions and can install/remove their own boot menu entries at free will, without interfering with the entries of other installed operating systems.
- Drop-in directories are otherwise now pretty ubiquitous on Linux as an easy way to extend configuration without having to edit, regenerate or manipulate configuration files. For the sake of uniformity we should do the same for extending the boot menu.
- Userspace code can sanely parse boot loader configuration which is essential with modern BIOSes which do not necessarily initialize USB keyboards anymore during boot, which makes boot menus hard to reach for the user. If userspace code can parse the boot loader configuration, too, this allows for UIs that can select a boot menu item to boot into, before rebooting the machine, thus not requiring interactivity during early boot.
- To unify and thus simplify configuration of the various boot loaders around, which makes configuration of the boot loading process easier for users, administrators and developers alike
- For boot loaders with configuration scripts such as grub2, adopting this spec allows for mostly static scripts that are generated only once at first installation, but then do not need to be updated anymore as that is done via drop-in files exclusively.
Why not simply rely on the EFI boot menu logic?
The EFI specification provides a boot options logic that can offer similar functionality. Here's why we think that it is not enough for our uses:
- The various EFI implementations implement the boot order/boot item logic to different levels. Some firmware implementations do not offer a boot menu at all and instead unconditionally follow the EFI boot order, booting the first item that is working.
- If the firmware setup is used to reset all data usually all EFI boot entries are lost, making the system entirely unbootable, as the firmware setups generally do not offer a UI to define additional boot items. By placing the menu item information on disk it is always available, regardless if the BIOS setup data is lost.
- Harddisk images should be moveable between machines and be bootable without requiring explicit EFI variables to be set. This also requires that the list of boot options is defined on disk, and not in EFI variables alone.
- EFI is not universal yet (especially on non-x86 platforms), this specification is useful both for EFI and non-EFI boot loaders.
- Many EFI systems disable USB support during early boot to optimize boot times, thus making keyboard input unavailable in the EFI menu. It is thus useful if the OS UI has a standardized way to discover available boot options which can be booted to.
We define two directories:
$BOOT/loader/is the directory containing all files defined by this specification
$BOOT/loader/entries/is the directory containing the drop-in snippets. This directory contains one
.conffile for each boot menu item.
These directories are defined below the placeholder file system
$BOOT. This placeholder file system shall be determined during installation time, and an fstab entry for it shall be created mounting it to /boot. The installer program should pick
$BOOT according to the following rules:
- If the OS is installed on a disk with MBR disk label, and a partition with the MBR type id of 0xEA already exists it should be used as $BOOT.
- Otherwise, if the the OS is installed on a disk with MBR disk label, a new partition with MBR type id of 0xEA shall be created, of a suitable size (let's say 500MB), and it should be used as $BOOT.
- If the OS is installed on a disk with GPT disk label, and a partition with the GPT type GUID of bc13c2ff-59e6-4262-a352-b275fd6f7172 already exists, it should be used as $BOOT.
- Otherwise, if the OS is installed on a disk with GPT disk label, and an ESP partition (i.e. with the GPT type UID of c12a7328-f81f-11d2-ba4b-00a0c93ec93b) already exists and is large enough (let's say 250MB) and otherwise qualifies, it should be used as $BOOT.
- Otherwise, if the OS is installed on a disk with GPT disk label, and if the ESP partition already exists but is too small, a new suitably sized (let's say 500MB) partition with GPT type GUID of bc13c2ff-59e6-4262-a352-b275fd6f7172 shall be created and it should be used as $BOOT.
- Otherwise, if the OS is installed on a disk with GPT disk label, and no ESP partition exists yet, a new suitably sized (let's say 500MB) ESP should be created and should be used as $BOOT.
Note: In all cases the /loader directory should be located directly in the root of the file system. Specifically, if $BOOT is the ESP, then /loader directory should be located directly in the root directory of the ESP, and not in the EFI/ subdirectory.
$BOOT should be considered shared among all OS installations of a system. Instead of maintaining one
$BOOT per installed OS (as
/boot was traditionally handled), all installed OS share the same place to drop in their boot-time configuration.
$BOOT can either be a FAT file system (which is probably the best choice, to allow acess/editing of the configuration from Windows and the EFI shell, too), or ext2, ext3, or ext4. Other file systems should not be used. Applications accessing
$BOOT should hence not assume that fancier file system features such as symlinks, hardlinks, access control or case sensitivity are supported.
$BOOT/loader/entries/ directory each OS vendor may drop one or more configuration snippets with the suffix ".conf", one for each boot menu item. The file name of the file is used for identification of the boot item, but shall never be presented to the user in the UI. The file name may be chosen freely but should be unique enough to avoid clashes between OS installations. More specifically it is suggested to include the machine ID (
/etc/machine-id or the D-Bus machine ID for OSes that lack
/etc/machine-id), the kernel version (as returned by
uname -r) and an OS identifier (The ID field of
/etc/os-release). Example: $BOOT/loader/entries/6a9857a393724b7a981ebb5b8495b9ea-3.8.0-2.fc19.x86_64.conf`.
These configuration snippets shall be Unix-style text files (i.e. line separation with a single newline character), in the UTF-8 character sets. The configuration snippets are loosely inspired on Grub1's configuration syntax. Lines beginning with '#' shall be ignored and used for commenting. The first word of a line is used as key, and shall be separated by a space from its value. The following keys are known:
titleshall contain a human readable title string for this menu item. This will be displayed in the boot menu for the item. It is a good idea to initialize this from the
/etc/os-release. This name should be descriptive and does not have to be unique. If a boot loader discovers two entries with the same title it is a good idea to show more than just the raw title in the UI, for example by appending the
versionfield. This field is optional. Example: "Fedora 18 (Spherical Cow)".
versionshall contain a human readable version string for this menu item. This is usually the kernel version and is intended for use by OSes to install multiple kernel versions at the same time with the same
titlefield. This field shall be in a syntax that is useful for Debian-style version sorts, so that the boot loader UI can determine the newest version easily and show it first or preselect it automatically. This field is optional. Example:
machine-idshall contain the machine ID of the OS
/etc/machine-id. This is useful for boot loaders and applications to filter out boot entries, for example to show only a single newest kernel per OS, or to group items by OS, or to maybe filter out the currently booted OS in UIs that want to show only other installed operating systems. This ID shall be formatted as 32 lower case hexadecimal characters (i.e. without any UUID formatting). This key is optional. Example:
linuxrefers to the kernel to spawn, and shall be a path relative to the
$BOOTdirectory. It is recommended that every distribution creates a machine id and version specific subdirectory below
$BOOTand places its kernels and initial RAM disk images there. Example:
initrdrefers to the initrd to use when executing the kernel. This also shall be a path relative to the
$BOOTdirectory. This key is optional. Example:
efito spawn arbitrary EFI programs. This also takes a path relative to
$BOOT. This key is only available on EFI systems.
optionsshall contain kernel parameters to pass to the Linux kernel to spawn. This key is optional.
devicetreerefers to the binary device tree to use when executing the kernel. This also shall be a path relative to the $BOOT directory. This key is optional. Example:
Each configuration drop-in snippet must include at least a
linux or an
efi key, and is otherwise not valid. Here's an example for a complete drop-in file:
/boot/loader/entries/6a9857a393724b7a981ebb5b8495b9ea-3.8.0-2.fc19.x86_64.conf title Fedora 19 (Rawhide) version 3.8.0-2.fc19.x86_64 machine-id 6a9857a393724b7a981ebb5b8495b9ea options root=UUID=6d3376e4-fc93-4509-95ec-a21d68011da2 linux /6a9857a393724b7a981ebb5b8495b9ea/3.8.0-2.fc19.x86_64/linux initrd /6a9857a393724b7a981ebb5b8495b9ea/3.8.0-2.fc19.x86_64/initrd
On EFI systems all kernel images shall be EFI images. In order to be compatible with EFI systems it is highly recommended only to install EFI kernel images, even on non-EFI systems, if that's applicable and supported on the specific architecture.
Note that these configurations snippets do not need to be the only configuration source for a boot loader. It may extend this list of entries with additional items from other configuration files (for example its own native configuration files) or automatically detected other entries without explicit configuration.
Note that these configuration snippets may only reference kernels (and EFI programs) that reside on the same file system as the configuration snippets, i.e. everything referenced must be contained in the same file system. This is by design, as referencing other partitions or devices would require a non-trivial language for denoting device paths. If kernels/initrds are to be read from other partitions/disks the boot loader can do this in its own native configuration, using its own specific device path language, and this is out of focus for this specification. More specifically, on non-EFI systems configuration snippets following this specification cannot be used to spawn other operating systems (such as Windows).
To make this explicitly clear: this specification is designed with "free" operating systems in mind, starting Windows or MacOS is out of focus with these configuration snippets, use boot-loader specific solutions for that. In the text above, if we say "OS" we hence imply "free", i.e. primarily Linux (though this could be easily be extended to the BSDs and whatnot).
Note that all paths used in the configuration snippets use a Unix-style "/" as path separator. This needs to be converted to an EFI-style "\" separator in EFI boot loaders.
A boot loader needs a file system driver to discover and read
$BOOT, then simply reads all files
$BOOT/loader/entries/*.conf, and populates its boot menu with this. Optionally it might sort the menu based on the
version fields, and possibly others. It uses the file name to identify specific items, for example in case it supports storing away default entry information somewhere. A boot loader should generally not modify these files.
A kernel installer package installs the kernel and initrd images to
$BOOT (it is recommended to place these files in a vendor and OS and installation specific directory) and then generates a configuration snippet for it, placing this in
$BOOT/loader/entries/xyz.conf, with xyz as concatenation of machine id and version information (see above). The files created by a kernel package are private property of the package, and should be deleted along with it.
A UI application intended to show available boot options shall operate similar to a boot loader, but might apply additional filters, for example by filtering out the booted OS via the machine ID, or by suppressing all but the newest kernel versions.
An OS installer picks the right place for
$BOOT as defined above (possibly creating a partition and file system for it), and pre-creates the
/loader/entries/ directory in it. It then installs an appropriate boot loader that can read these snippets. Finally it installs one or more kernel packages.
Out of Focus
There are a couple of items that are out of focus for this specifications:
- If userspace can figure out the available boot options, then this is only useful so much: we'd still need to come up with a way how userspace could communicate to the boot loader the default boot loader entry temporarily or persistently. Defining a common scheme for this is certainly a good idea, but out of focus for this specifications.
- This specifications is just about "Free" Operating systems. Hooking in other operating systems (like Windows, MacOS) into the boot menu is a different story, and should probably happen outside of this specification. For example, boot loaders might choose to detect other available OSes dynamically at runtime without explicit configuration (like Gummiboot does it), or via natvie configuration (for example via explicit Grub2 configuration generated once at installation).
- This specification leaves undefined what to do about systems which are upgraded from an OS that does not implement this specification. As the previous boot loader logic was largely handled by in distribution-specific ways we probably should leave the upgrade path (and whether there actually is one) to the distributions. The simplest solution might be to simply continue with the old scheme for old installations and use this new scheme only for new installations.