Re: [PATCH v5 3/8] bootstd: Support booting EFI where multiple options exist

[Tom Rini <trini@konsulko.com>]

[-- Attachment #1: Type: text/plain, Size: 5626 bytes --]

On Wed, Apr 05, 2023 at 11:56:59PM +0200, Mark Kettenis wrote:
> > Date: Wed, 5 Apr 2023 10:47:59 -0400
> > From: Tom Rini <trini@konsulko.com>
> > 
> > On Wed, Apr 05, 2023 at 05:28:07PM +1200, Simon Glass wrote:
> > > Hi Tom,
> > > 
> > > On Tue, 4 Apr 2023, 02:17 Tom Rini, <trini@konsulko.com> wrote:
> > > 
> > > > On Mon, Apr 03, 2023 at 12:56:49PM +0300, Ilias Apalodimas wrote:
> > > > > On Sat, Apr 01, 2023 at 07:31:49PM +1300, Simon Glass wrote:
> > > > > > Hi Tom,
> > > > > >
> > > > > > On Sat, 1 Apr 2023 at 07:02, Tom Rini <trini@konsulko.com> wrote:
> > > > > > >
> > > > > > > On Fri, Mar 31, 2023 at 10:25:56AM +1300, Simon Glass wrote:
> > > > > > >
> > > > > > > > The current EFI implementation has a strange quirk where it watches
> > > > > > > > loaded files and uses the last-loaded file to determine the device
> > > > that
> > > > > > > > is being booted from.
> > > > > > > >
> > > > > > > > This is confusing with bootstd, where multiple options may exist.
> > > > Even
> > > > > > > > loading a device tree will cause it to go wrong. There is no API
> > > > for
> > > > > > > > passing this information, since the only entry into booting an EFI
> > > > image
> > > > > > > > is the 'bootefi' command.
> > > > > > > >
> > > > > > > > To work around this, call efi_set_bootdev() for EFI images, if
> > > > possible,
> > > > > > > > just before booting.
> > > > > > > >
> > > > > > > > Signed-off-by: Simon Glass <sjg@chromium.org>
> > > > > > >
> > > > > > > Shouldn't this all be a simple wrapper around the EFI Standard
> > > > > > > BootDeviceOrder or whatever that's called?
> > > > > >
> > > > > > I think you are referring to boot manager, which isn't used here. This
> > > > > > is replicating the existing distroboot functionality in standard boot.
> > > > >
> > > > > The distroboot functionality *was* trying to behave like the EFI spec
> > > > > expects the bootmanager to behave.  Unfortunately I haven't had time to
> > > > > review the distroboot patches closely, but back when this started, my
> > > > point
> > > > > was that EFI doesn't need anything.  Whenever the EFI flow is added
> > > > bootstd
> > > > > should 'just' call the bootmanager.
> > > >
> > > > Yes, this. We're trying make things cleaner overall, so the EFI portion
> > > > of bootstd distro boot should just be "call EFI bootmanager" as that has
> > > > a well defined standard way to specify what devices to try in what
> > > > order.
> > > >
> > > 
> > > We already call bootmgr in standard boot, if it is enabled.
> > > 
> > > But I am not sure how widely that is used...
> > > 
> > > This patch is about corner cases in the distro scripts. If we are to turn
> > > these down we do need to try to do the same thing.
> > 
> > We probably need some distro people to chime in about what they're doing
> > / expecting at this point in time? I would have sworn that the long term
> > part of EFI "distro boot" would be using bootmgr since that's the
> > standards based way to set boot order. And if you don't have a device
> > tree in U-Boot, and want the distribution one, aren't you then using
> > something like grub which has a "dtb" keyword to handle that on its own?
> > That's not saying that "distro boot" doesn't need to load the device
> > tree, for when it's then calling booti/bootz/bootm, but not for the EFI
> > case these days? Or no?
> 
> The short anserwer is no.
> 
> The long answer:
> 
> OpenBSD requires EFI (booti/bootz/bootm only support booting Linux
> kernels in various forms) but also relies on a proper device tree
> being provided to its EFI bootloader.  While we have made significant
> progress in having U-Boot provide a fully synched up device tree for
> most of the SoCs that OpenBSD supports, we're not quite there yet, so
> some people rely on U-Boot loading (and tweaking) an updated
> device-tree from the EFI System Partition.
> 
> Our bootloader does have a "mach dtb" command to load a device tree
> from the OpenBSD root partition, but this is considered to be a
> debugging command and can't load a device tree from the EFI System
> Partition.  Loading a device tree this way means the device tree does
> not get tweaked by U-Boot, which is problematic on some SoCs/boards.
> 
> We now have the EFI_DT_FIXUP_PROTOCOL, but our bootloader doesn't use
> this yet.  If the consensus is that distroboot-with-efiboot needs to
> deprecated I can work on supporting that protocol in our bootloader.
> But a release of OpenBSD with support for that will not be available
> until november 2023.
> 
> Alternatively bootmgr could implement the device tree loading and
> fixup.  Or the bootstd stuff could do this before starting bootmgr.  I
> understand the concerns about loading device trees from disk when
> secure boot is enabled.  So maybe this should only be done when secure
> boot is disabled.
> 
> Also note that setting the BootOrder EFI variable from inside the OS
> isn't supported by most (all?) boards currently supported by U-Boot.
> So a convenient way to set the BootOrder EFI variable from within
> U-Boot is needed.
> 
> Cheers,
> 
> Mark
> 
> P.S. Does grub support the EFI_DT_FIXUP_PROTOCOL these days?  If not
>      (or if your Linux distro still ships with an older version of
>      grub) you'll face the same issue as OpenBSD regarding the lack of
>      fixups for loaded devictrees.

I just want to chime in and say thanks for explaining your use case, so
we can make sure what we're doing will actually work for everyone.

-- 
Tom

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 659 bytes --]