Re: [RFC PATCH] efi: Put Linux specific magic number in the DOS header

[Daniel Kiper]

On Wed, Nov 09, 2022 at 04:01:27PM +0100, Heinrich Schuchardt wrote:
> On 11/9/22 15:16, Ard Biesheuvel wrote:
> > GRUB currently relies on the magic number in the image header of ARM and
> > arm64 EFI kernel images to decide whether or not the image in question
> > is a bootable kernel.
> >
> > However, the purpose of the magic number is to identify the image as one
> > that implements the bare metal boot protocol, and so GRUB, which only
> > does EFI boot, can only boot images that could potentially be booted in
> > a non-EFI manner as well.
> >
> > This is problematic for the new zboot decompressor image format, as it
> > can only boot in EFI mode, and must therefore not use the bare metal
> > boot magic number in its header.
> >
> > For this reason, the strict magic number was dropped from GRUB, to
> > permit essentially any kind of EFI executable to be booted via the
> > 'linux' command, blurring the line between the linux loader and the
> > chainloader.
> >
> > So let's use the same field in the DOS header that RISC-V and arm64
> > already use for their 'bare metal' magic numbers to store a 'generic
> > Linux kernel' magic number, which can be used to identify bootable
> > kernel images in PE format which don't necessarily implement a bare
> > metal boot protocol in the same binary.
> >
> > Let's set the generic magic number for x86 images as well: existing
> > bootloaders already have their own methods to identify x86 Linux images
> > that can be booted in a non-EFI manner, and having the magic number in
> > place there will ease any future transitions in loader implementations
> > to merge the x86 and non-x86 EFI boot paths.
> >
> > Note that 32-bit ARM already uses the same location in the header for a
> > different purpose, but the ARM support is already widely implemented and
> > the EFI zboot decompressor is not available on ARM anyway, so we just
> > disregard it here.
> >
> > Cc: Huacai Chen <chenhuacai@kernel.org>
> > Cc: Atish Patra <atishp@rivosinc.com>
> > Cc: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
> > Cc: Daniel Kiper <daniel.kiper@oracle.com>
> > Cc: Leif Lindholm <quic_llindhol@quicinc.com>
> > Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
> > ---
> > The idea is that, with this in place, the existing arm64 support in GRUB
> > can be made generic, with the arm64 variant of the arch image header
> > typedef being made generic as well.
> >
> > Any code that attempts to identify EFI images as kernel images should
> > check for the arm64, RISC-V and generic values, and if the latter, look
> > at the PE machine type if it wants to know the architecture.
> >
> >   arch/loongarch/kernel/head.S                | 3 ++-
> >   arch/x86/boot/header.S                      | 3 ++-
> >   drivers/firmware/efi/libstub/zboot-header.S | 3 ++-
> >   include/linux/pe.h                          | 7 +++++++
> >   4 files changed, 13 insertions(+), 3 deletions(-)
>
> We need files in Documentation/ that describes the meaning and value of the
> field per architecture.
>
> We already have:
>
> riscv: Documentation/riscv/boot-image-header.rst
> arm64: Documentation/arm64/booting.rst

I think we should add something to the Documentation/x86/boot.rst too.
Probably in or after "EFI Handover Protocol (deprecated)" section.

> But other UEFI architectures are missing.
>
> Otherwise I am fine with the change.
>
> Best regards
>
> Heinrich
>
> >
> > diff --git a/arch/loongarch/kernel/head.S b/arch/loongarch/kernel/head.S
> > index 97425779ce9f3499..e1deea93aaafa069 100644
> > --- a/arch/loongarch/kernel/head.S
> > +++ b/arch/loongarch/kernel/head.S
> > @@ -25,7 +25,8 @@ _head:
> >     .dword  kernel_entry            /* Kernel entry point */
> >     .dword  _end - _text            /* Kernel image effective size */
> >     .quad   0                       /* Kernel image load offset from start 
> > of RAM */
> > -   .org    0x3c                    /* 0x20 ~ 0x3b reserved */
> > +   .org    0x38                    /* 0x20 ~ 0x38 reserved */
> > +   .long   LINUX_PE_MAGIC
> >     .long   pe_header - _head       /* Offset to the PE header */
> >   pe_header:
> > diff --git a/arch/x86/boot/header.S b/arch/x86/boot/header.S
> > index d31982509654dcb1..9338c68e7413d6e6 100644
> > --- a/arch/x86/boot/header.S
> > +++ b/arch/x86/boot/header.S
> > @@ -80,10 +80,11 @@ bs_die:
> >     ljmp    $0xf000,$0xfff0
> >   #ifdef CONFIG_EFI_STUB
> > -   .org    0x3c
> > +   .org    0x38
> >     #
> >     # Offset to the PE header.
> >     #
> > +   .long   LINUX_PE_MAGIC
> >     .long   pe_header
> >   #endif /* CONFIG_EFI_STUB */
> > diff --git a/drivers/firmware/efi/libstub/zboot-header.S 
> > b/drivers/firmware/efi/libstub/zboot-header.S
> > index 9e6fe061ab07a008..97c2637337d79913 100644
> > --- a/drivers/firmware/efi/libstub/zboot-header.S
> > +++ b/drivers/firmware/efi/libstub/zboot-header.S
> > @@ -20,7 +20,8 @@ __efistub_efi_zboot_header:
> >     .long           __efistub__gzdata_size - ZBOOT_SIZE_LEN // payload size
> >     .long           0, 0                                    // reserved
> >     .asciz          COMP_TYPE                               // compression 
> > type
> > -   .org            .Ldoshdr + 0x3c
> > +   .org            .Ldoshdr + 0x38
> > +   .long           LINUX_PE_MAGIC
> >     .long           .Lpehdr - .Ldoshdr                      // PE header 
> > offset
> >   .Lpehdr:
> > diff --git a/include/linux/pe.h b/include/linux/pe.h
> > index 1d3836ef9d92dcd8..fa176c24167c301c 100644
> > --- a/include/linux/pe.h
> > +++ b/include/linux/pe.h
> > @@ -31,6 +31,13 @@
> >   #define LINUX_EFISTUB_MAJOR_VERSION               0x1
> >   #define LINUX_EFISTUB_MINOR_VERSION               0x0
> > +/*
> > + * LINUX_PE_MAGIC appears at offset 0x30 into the MSDOS header of EFI 
> > bootable

s/0x30/0x38/?

However, it means you go into reserved part of MS-DOS header [1]. This
should be safe here. Though I think it should be mentioned in the commit
message and/or documentation.

Another question which bothers me is: why do not use "OEM identifier"
[1] and "OEM info" [1] to achieve what we want? IMO this would be more
natural but on the other hand both fields are not well defined...

Finally, current approach do not allow us easily encode info about
images which support more than one boot protocol. Maybe not big deal at
this point but I would ponder on it at least a bit...

Daniel

[1] https://wiki.osdev.org/MZ#PE_Extension

> > + * Linux kernel images that target the architecture as specified by the 
> > PE/COFF
> > + * header machine type field.
> > + */
> > +#define LINUX_PE_MAGIC     0x818223cd
> > +
> >   #define MZ_MAGIC  0x5a4d  /* "MZ" */
> >   #define PE_MAGIC          0x00004550      /* "PE\0\0" */