EFI bootloader from systemd, with hacks to build it on Alpine/pmOS.
Cross compilation (using a meson cross file) is used for building
32-bit version on x86_64, for systems that have a 32-bit EFI. Everything
else assumes that the EFI arch matches the CPU arch.
Besides supporting all the archs we need, another major goal was to
minimize the number of changes to systemd's build system required to
build only the bootloader, so that maintaining/rebasing isn't *too*
painful...
I am adding this to the "main" category, because I don't think there's a
way to add it to Alpine. It requires cross compiling to x86 on x86_64
(to support 32-bit EFI on this arch), and Alpine doesn't support this.
It requires stuff in pmaports/cross.
--- Research notes ---
I started looking at all of this because I wanted to come up with a
single way to boot Linux via EFI, that supports all (or as many as possible)
devices in pmaports. I looked at quite a few different options, and have
some notes below about my observations and conclusions for each.
Of everything I looked at, systemd-boot was the clear winner that met
the most requirements ("pro" below) with the fewest downsides ("con"
below).
Using a Unified Kernel Image (UKI) was a close second place, however
systemd-boot can also support booting UKI images quite easily (while
also giving us more flexibility to boot other things easily too), so I
think it wins over UKI.
The capitalization (or lack thereof) of the "pro" and "con" markers
below is significant: "PRO" / "CON" are major pros or cons for each
point (e.g. a major downside that blocks using the option), and
"pro"/"con" are minor (e.g. a downside that I'm willing to overlook.)
---- Requirements ----
- Arch support:
- x86_64
- x86 (nice to have, but not sure if necessary...)
- armv7
- aarch64
- riscv64
- EFI support:
- support 32-bit EFI on x86_64 CPU (includes being able to build
32-bit .efi app on x86_64)
- Easy to configure
- Easy to maintain
- Any changes to the bootloader required to get it working in
pmOS
- Config for it
---- Evaluated options ----
------ grub ------
- (PRO) can target all required archs
- (CON) grub can't be installed in pmb chroot, it calls grub-install and
that fails due to something missing in /dev. Maybe this could be worked
around in pmb?
- (CON) grub-mkimage exe is integrated in grub package, grub-efi depends on
grub
- don't want to install all of grub just for 1 exe and/or the EFI modules
- downsides of installing all of grub is that I think it can mislead
users into thinking we use grub the "normal way". this might cause them
to have the wrong expectations and break pmOS boot on their
system
- have POC "fixing" this
- I'm not sure upstream Alpine will like this, it's ugly
- (CON) grub x86 EFI support for x86_64 is currently in pmaports, that's
pretty ugly.
- IMHO forking grub (or grub components) for this purpose signals
to me that grub is the wrong tool for this job
------- kernel's efistub -------
- (PRO) already included in the kernel, nothing else required
- (pro) initrd and dtb can be passed in the kernel cmdline...
however....
- (CON) kernel cmdline can only be set at compile time
- (con) not all kernels may have EFISTUB set?
- (con) can't do measured/secure boot
- (con) requires a fairly recent kernel on aarch w/ efi_zboot support
enabled since we compress the kernel
------- UKI -------
- (PRO) very simple, 1 file thing
- (PRO) supports adding dtb, setting kernel cmdline and so on
- (pro) can do measured/secure boot
- (CON) requires an EFI stub loader
- can't find a stub loader that meets all requirements (other than
the one from systemd-boot...)
- (con) requires efi-mkuki or dealing with objcopy directly (eww)
- (con) requires a fairly recent kernel on aarch w/ efi_zboot support
------- limine -------
- (PRO) easy to install/configure, already have boot-deploy and pmaports
patches
- (PRO) can be cross compiled easily
- evidence is in aports
- ...but I couldn't reproduce building aarch64 and riscv64 on x86_64
- (pro) can do measured/secure boot (I think?)
- (CON) doesn't target all required archs
- can't do "linux boot" on aarch64, only "chainload"
- what about using chainload everywhere?
- requires using efistub in kernel
- what about dtb= and upstream recommendation to not use it except for
debug?
- no kernel compression support on aarch64
- see efi-stub.txt kernel doc
- (CON) vendors libgcc to support cross compilation
- probably not a good idea to trust binaries produced in microsoft
github's CI for some random project
------- stubbyboot -------
- (PRO) a straight forward stub loader
- (pro) can do measured/secure boot
- (CON) doesn't target all required archs
- (CON) cross compiling doesn't work.
- gcc can't do 32-bit on x86_64 Alpine...
- gnu-efi-dev needs to be fixed to package both 32-bit and 64-bit on x86_64...
- have patch in ~/src/aports that kinda does it.. but needs to be
fixed/finished
- maybe limine-efi works with it?
- tried, but fails due to missing efilib.h in limine-efi
------- systemd-stub -------
- (PRO) another straight forward stub loader
- (PRO) many (many) people using it, as part of systemd-boot
- (pro) can do measured/secure boot
- (con) requires a fairly recent kernel on aarch w/ efi_zboot support
enabled since we compress the kernel
- (con) doesn't target all required archs
- but does claim to support most... missing armv7.. maybe it
works?
- (con) will end up maintaining some downstream patch to build it
- hopefully the patch (if I can even make a working one!) is not too
complex!
- (CON) can't be built outside of systemd's silly large build system.
- UPDATE: largely resolved this in pmaports
- was able to build for native arch!
- can't build 32-bit on x86_64, no gcc multilib support in Alpine...
Couldn't get clang to work properly, but maybe it can somehow...
- https://github.com/mintsuki/libgcc-binaries ? NO! (don't want
bootloader binaries that depend on code compiled by microsoft /
github...)
------- DIY stub / bootloader -----
- (PRO) **might** target all required archs and other meet
requirements
- (CON) lots of time required to learn, design, do, debug, test
- (CON) lots of time required to learn, design, do, debug, test
- (CON) lots of time required to learn, design, do, debug, test
- (CON) (get the hint yet???)
- (CON) written in C, probably (there's a rust EFI lib, lol...)
[ci:skip-build]: Already built successfully in CI
Clayton Craft