Re: [PATCH 1/7] multiboot2: Add load type header and support for the PE binary type

[Ross Lagerwall <ross.lagerwall@xxxxxxxxxx>]

On Tue, Mar 19, 2024 at 1:18 PM Roger Pau Monné <roger.pau@xxxxxxxxxx> wrote:
>
> On Wed, Mar 13, 2024 at 03:07:42PM +0000, Ross Lagerwall wrote:
> > Currently, multiboot2-compatible bootloaders can load ELF binaries and
> > a.out binaries. The presence of the address header tag determines
> > how the bootloader tries to interpret the binary (a.out if the address
> > tag is present else ELF).
> >
> > Add a new load type header tag that explicitly states the type of the
> > binary. Bootloaders should use the binary type specified in the load
> > type tag. If the load type tag is not present, the bootloader should
> > fall back to the previous heuristics.
> >
> > In addition to the existing address and ELF load types, specify a new
> > optional PE binary load type. This new type is a useful addition since
> > PE binaries can be signed and verified (i.e. used with Secure Boot).
> >
> > Signed-off-by: Ross Lagerwall <ross.lagerwall@xxxxxxxxxx>
> > ---
> >  doc/multiboot.texi | 39 ++++++++++++++++++++++++++++++++++-----
> >  doc/multiboot2.h   | 13 +++++++++++++
> >  2 files changed, 47 insertions(+), 5 deletions(-)
> >
> > diff --git a/doc/multiboot.texi b/doc/multiboot.texi
> > index df8a0d056e76..d12719c744eb 100644
> > --- a/doc/multiboot.texi
> > +++ b/doc/multiboot.texi
> > @@ -511,11 +511,12 @@ assumes that no bss segment is present.
> >
> >  Note: This information does not need to be provided if the kernel image
> >  is in @sc{elf} format, but it must be provided if the image is in a.out
> > -format or in some other format. When the address tag is present it must
> > -be used in order to load the image, regardless of whether an @sc{elf}
> > -header is also present. Compliant boot loaders must be able to load
> > -images that are either in @sc{elf} format or contain the address tag
> > -embedded in the Multiboot2 header.
> > +format or in some other format. If the load type tag is not specified
> > +and the address tag is present it must be used in order to load the
> > +image, regardless of whether an @sc{elf} header is also present.
> > +Compliant boot loaders must be able to load images that are either in
> > +@sc{elf} format or contain the address tag embedded in the Multiboot2
> > +header.
> >
> >  @subsection The entry address tag of Multiboot2 header
> >
> > @@ -732,6 +733,34 @@ and @samp{2} means load image at highest possible 
> > address but not
> >  higher than max_addr.
> >  @end table
> >
> > +@node Load type tag
> > +@subsection Load type tag
> > +
> > +@example
> > +@group
> > +        +-------------------+
> > +u16     | type = 11         |
> > +u16     | flags             |
> > +u32     | size = 12         |
> > +u32     | load_type         |
> > +        +-------------------+
> > +@end group
> > +@end example
> > +
> > +This tag indicates the type of the payload and how the boot loader
> > +should load it.
> > +
> > +The meaning of each field is as follows:
> > +
> > +@table @code
> > +@item load_type
> > +Recognized load types are @samp{0} for address (i.e. load a.out using
> > +the address tag), @samp{1} for ELF, and @samp{2} for PE. Compliant
> > +bootloaders should implement support for a.out and ELF as a minimum.  If
> > +this tag is not specified, the boot loader should attempt to load the
> > +payload using the information specified in the address tag if present,
> > +else it should load the payload as an ELF binary.  @end table
>
> I wonder if it would be simpler to use the following order instead:
>
> 1. Address tag
> 2. Load type tag
> 3. ELF header
>
> It's pointless to add a Loader type tag with load_type == 0, as that's
> already mandated by the Address tag.  IOW: signaling the use of the
> Address tag here is kind of pointless, if the fields in the Address
> tag are set, that's the only signaling required for the data in the
> Address tag to be used.
>
> Or are we attempting to support images that set Address tag and Load
> type tag != 0 in order to use the Address tag when the loader doesn't
> recognize the Load type tag, and otherwise use a different format?

No, that wasn't my intention. My intention for the load type tag was
to have one tag that unambiguously describes the payload format and if
that is missing, fall back to the previous logic for compatibility.
It avoids more complicated heuristics if different payload types are
added in the future.

>
> Would it be sensible for multiboot2 to use PE in preference to ELF if
> present on the image?  (without requiring any signaling).  I would
> think this could be troublesome if kernels are so far expecting the
> ELF header to be used with multiboot2, even if they also expose a PE
> header.
>

AFAIK an ELF image can't also be a valid PE/COFF image since they have
different magic numbers at the start of the image. Perhaps it would
be simpler to avoid introducing the load type tag and just load the
payload based on the magic number?

Ross