From: Simon Glass sjg@chromium.org Date: Mon, 28 Jun 2021 08:18:26 -0600
Hi Tom, Mark,
On Mon, 28 Jun 2021 at 07:37, Tom Rini trini@konsulko.com wrote:
On Mon, Jun 28, 2021 at 10:38:50AM +0200, Mark Kettenis wrote:
From: Simon Glass sjg@chromium.org Date: Sun, 27 Jun 2021 19:48:34 -0600
It has come to light that EFI_LOADER adds an extraordinary amount of code to U-Boot. For example, with nokia_rx51 the size delta is about 90KB. About 170 boards explicitly disable the option, but is is clear that many more could, thus saving image size and boot time.
EFI_LOADER used to be a lot smaller. It is great to see that over the years UEFI support has become more complete, but a lot of that new code implements features that are not at all essential for just booting an OS from storage. If that growth leads to the suggestion to disable EFI_LOADER completely by default, we're putting the cart before the horse.
Well, I see I forgot to prefix my patch with RFC, but I hadn't found EFI_LOADER being used in the wild on armv7, but wasn't sure about the BSD families. I did see that Debian doesn't use it, and that Armbian doesn't even use it on aarch64.
The current situation is affecting U-Boot's image as a svelt bootloader.
Really? I know UEFI has a bad reputation in the Open Source world, and some of its Microsoft-isms are really annoying (yay UCS-2). But it works, it provides a standardized approach across several platforms (ARMv7, AMRv8, RISC-V) and the industry seems to like it. Personally I'd wish the industry had standardized on Open Firmware instead, but that ship sailed a long time ago...
I find it hard to imagine that 90k is a serious amount of storage for something that is going to include a multi-MB Linux kernel. This isn't code that lives in SPL or TPL where severe size restrictions apply.
In one of those cases where I need to pop back in to the other (Nokia N900 specific) thread and see if the big size reduction really was just disabling EFI_LOADER, it's perhaps just one of those "fun" things about Kconfig and anything other than "make oldconfig" for spotting new config options that default to enabled.
Yes it will be interesting to see what you find there. My results on nokia_rx51 were something like this:
default arm: (for 1/1 boards) all +129370.0 bss +1136.0 data +7399.0 rodata +10989.0 text +109846.0
without ebbr arm: (for 1/1 boards) all +38460.0 bss +1040.0 data +2375.0 rodata +5333.0 text +29712.0
with various other things: CONFIG_OF_LIBFDT_ASSUME_MASK=7 # CONFIG_OF_TRANSLATE is not set # CONFIG_SIMPLE_BUS is not set # CONFIG_TI_SYSC is not set # CONFIG_CMD_FDT is not set
arm: (for 1/1 boards) all +19170.0 bss -16.0 data +360.0 rodata+3274.0 text +15552.0
(Mark, in the same email:)
FIT simply isn't fit for purpose (pun intended). It only really works for booting Linux, and forces people to combine u-boot, kernel, initial ramdisk and other firmware components into a single image. That is really undesirable as:
- This makes it sigificantly harder to update individual components of such an image. Making it hard to update a kernel is obviously a serious security risk.
- This makes it impossible to build an OS install image that works om multiple boards/SoCs.
I would really like to understand this better. The whole thing is a complete mystery to me.
Firstly I have sometimes fiddled with booting other OSes using FIT. It seemed OK. I can't see why it only works with Linux.
Well, you could of course rework the boot flow of other OSes such that booting them includes some sort of FIT if you really wanted to. I But an OS like OpenBSD comes with its own bootloader that is essential in the boot flow. On OpenBSD armv7/arm64/riscv64 it adds some essential properties to the device tree. Besides, the kernel itself relies on a valid EFI memory map.
Secondly, I don't expect that U-Boot itself would be in the FIT.
So the FIT would only contain the OS kernel and other OS components? What about the FIT that is used on some arm64 platforms to combine U-Boot and TF-A? I guess you can have multiple FITs...
Thirdly, do you really want the kernel and initrd to be separate? At least in the systems I have used, they are built together, even having the same name, e.g.:
initrd.img-5.10.40-1rodete1-amd64 System.map-5.10.40-1rodete1-amd64 vmlinuz-5.10.28-1rodete2-amd64
I don't really use Linux on these platforms. But I'd expect the normal package management tools of my Linux distribution to build those as necessary and place them in the root file system where the bootloader picks them up. And as a distro developer, I'd like to have the same approach work on all Linux systems, regardless the specific firmware they're running (EDK2, U-Boot or something completely different). Ideally that wouldn't even depend on the architecture.
Now Armbian takes a different approach, and does treat most systems they provide as special snowflakes, providing flash images for each board. But that doesn't scale and makes for a fairly crappy user experience. They don't always support booting a mainline kernel. And updating the kernel often requires installing special packages.
Finally, for the firmware components, do you mean system firmware? If so, I would expect it to be more convenient to distribute updates to that separately, although I suppose they could be combined with the kernel if the combinatorial explosion can be contained. What is the problem, exactly? (If you mean peripheral firmware, I would expect fwupd to handle that.)
I guess I mean system firmware. Essentially everything that runs on the system before my OS bootloader runs. So for me, U-Boot is part of the system firmware even if it sometimes happens to live on removable media.
What exactly is impossible? Can you please be more specific?
So let me explain what we want for OpenBSD. We really want a uniform user experience across platforms, and don't want to maintain different codepaths for special snowflake platforms that might exist for a specific architecture.
Installing OpenBSD on a machine should be as simple as dd'ing the installer to some boot media, plugging it in and powering the machine on. Now this is somewhat tricky to achieve on some hardware targetted by U-Boot as they don't come with usable system firmware on the board itself. But on those boards you can mostly get away with having U-Boot on uSD or eMMC and the OS installer on USB.
Updating the OpenBSD kernel should be as simple as copying the kernel to /bsd. Since the root filesystem uses the UFS/FFS filesystem, this means that whatever we use as a bootloader must be able to read from that filesystem. To make sure the kernel is properly seeded with entropy, the OpenBSD bootloader has some additional tricks up its sleeve. It will replace a special segment in the kernel with random data before handing control to the kernel. On platforms that support it, it will try to use a firmware-provided RNG to do this (and EFI supports this) but also mix in some random data from a file on the UFS/FFS filesystem. It will actually mark that file as "used" after reading it to throw a warning when the file is reused when the machine is rebooted (it should have been filled with fresh new entropy). So you really need to use the OpenBSD bootloader instead loading an OpenBSD kernel directly from system firmware.
Updating the OpenBSD bootloader should be as simple as running installboot(8) from within the OS.
This all works on pretty much any architecture that OpenBSD supports. And right now, thanks to EFI_LOADER support in U-Boot, we have a nearly uniform boot flow on amd64, arm64, armv7 and riscv64.
FIT is just a container. It seems to have been rejected by the EFI crew at some point. Perhaps I just need to try to use it with one of the distros out there, to actually understand what is going on here. But any help is appreciated.
It just doesn't make sense for us to use a container just because the system firmware (U-Boot) insists on it. The kernel lives in the root UFS/FFS filesystem and the bootloader lives on an MS-DOS filesystem on the root disk.
EFI_LOADER is required by EBBR, a new boot standard which aims to bring in UEFI protocols to U-Boot. But EBRR is not required for booting. U-Boot already provides support for FIT, the 'bootm' command and a suitable hand-off to Linux. EBRR has made the decision to create a parallel infrastructure, e.g. does not use FIT, nor U-Boot's signing infrastructure.
EFI_LOADER is required to boot FreeBSD and OpenBSD on several platforms as well as generic Linux distros. For example OpenBSD/armv7, OpenBSD/arm64 and OpenBSD/riscv64 all rely on EFI_LOADER to boot and have done so for the last 4 years. The fact that ARM has embraced UEFI as an embedded boot standard and branded it EBBR really isn't all that relevant.
To be clear here, I like EFI_LOADER. I too do wish some other technologies had become dominant for technical rather than inertia reasons, but here we are. Having played around with it on aarch64, there are some pretty nice comes-along-with parts to it.
What I hadn't seen, and am only a little skeptical of still, is how far backwards in generations it's going to be used on. The general wish is that users nor off the shelf OS groups need to rebuild U-Boot for a given board, and instead it just works. The number of new designs for 32bit parts is no where near the number of new designs for 64bit parts. So what we're seeing in U-Boot now is people updating support on their older designs, and not necessarily caring about using EFI_LOADER.
In a reply to one of the patches in this series, Heinrich mentions a few problems that need resolving (devices for partitions and file handles). Both of those features should first be added to U-Boot, so EFI can then use that support. In general, EFI has tried to work beside driver model, creating its own parallel tables, etc. I have tried to influence this at various points along the way, including at the start and I'm happy to dig out those threads if it helps. But I wasn't kidding. it really needs to be addressed. I would love to see Linaro (for example) organise something here and take this on. I am very happy to help.
Regards, Simon