Re: [v2 PATCH] RISC-V: Add a PE/COFF compliant Image header.

From: Paul Walmsley
Date: Thu May 16 2019 - 12:22:53 EST


+ ARM64 maintainers, Tom, Marek

Hi Atish,

On Mon, 13 May 2019, Atish Patra wrote:

> On 5/13/19 5:40 PM, Paul Walmsley wrote:
> > On Mon, 13 May 2019, Atish Patra wrote:
> > > On 5/13/19 5:09 PM, Paul Walmsley wrote:
> > >
> > > > What are the semantics of those reserved fields?
> > >
> > > +struct riscv_image_header {
> > > + u32 code0;
> > > + u32 code1;
> > > + u64 text_offset;
> > > + u64 image_size;
> > > + u64 res1;
> > > + u64 res2;
> > > + u64 res3;
> > > + u64 magic;
> > > + u32 res4; ---> We can use this for versioning when required
> > > + u32 res5; ---> This is reserved for PE/COFF header
> > > +};
> >
> > I saw that in your patch. The problem is that this doesn't describe what
> > other software might expect in those fields. Can anything at all be
> > placed in those reserved fields?
>
> Yes. The reserved fields can be used for anything that boot loaders and Linux
> kernel can agree with each other. If you look at the ARM64, they have
> "Informative flags" in place of res1.
>
> > > > > Do we need to add it now or add it later when we actually need a
> > > > > version
> > > > > number. My preference is to add it later based on requirement.
> > > >
> > > > If it isn't added now, how would bootloaders know whether it was there
> > > > or
> > > > not?
> > > >
> > > >
> > > Here is the corresponding U-Boot Patch
> > > https://patchwork.ozlabs.org/patch/1096087/
> > >
> > > Currently, boot loader doesn't care about versioning. Since we are
> > > updating a
> > > reserved field, offsets will not change. If a boot loader want to use the
> > > versioning, it should be patched along with the kernel patch.
> > >
> > > Any other boot loader that doesn't care about the version, it can continue
> > > to
> > > do so without any change.
> > >
> > > My idea is to enable the minimum required fields in this patch and keep
> > > everything else as reserved so that it can be amended in future as
> > > required.
> >
> > If those fields really are reserved for implementors to do whatever they
> > want with them, then that might be a reasonable approach. That seems
> > unlikely, however, since specification authors usually reserve the right
> > to use reserved fields for their own purposes in later versions.
> >
> Technically, we are just implementing the "DOS" header part of PE/COFF format
> for now. It only mandates a magic string "MZ" at the top and a 32bit value at
> offset 0x3c tells us offset of PE/COFF header in image.
> Anything in between is implementation specific.
>
> For example, it will be updated to support EFI stub as described in the commit
> text,
> "In order to support EFI stub, code0 should be replaced with "MZ" magic string
> and res5(at offset 0x3c) should point to the rest of the PE/COFF header (which
> will be added during EFI support)."

OK. I think we should try to share this header format with other
architectures. This one after all is copied from ARM64, and some of the
core fields will be the same across multiple architectures. That way we
can try to avoid proliferating different boot header formats for each
architecture, which should be better for both the kernel and the
bootloaders. ARM64 folks, would you be interested in working together on
this?

Meanwhile, to unblock RISC-V, and to make this header durable for future
extensions and to match the existing ARM64 usage, I think we should make
the following technical changes to what you proposed:

1. Reserve all of the existing ARM64 fields in the same way ARM64 does
now. This keeps open the possibility that we can merge this format
with the one used with ARM64, and reuse the same bootloader code.
Based on our discussions, it sounds like the primary difference between
what you're proposing and the ARM64 format involves the flags/res1
field. Let's keep that as a flag field, reuse ARM64's endianness bit
as architecture-independent, then define the rest of the flags in that
field as architecture-defined.

2. Allocate another set of reserved bits for a format version number.
Probably 16 bits is sufficient. This tells bootloaders how to
interpret the header fields in future extensions. The goal is to
preserve compatibility across newer and older versions of the header.
The existing ARM64 header would be version 0. This format that
incorporates these changes would be version 1. The thought here is to
preserve all of the semantics of existing fields in newer versions
(except for any remaining reserved fields), since many people often do
not replace their bootloaders.

3. Define a way to point to additional fields outside this existing
header. Another 32 bits of previously reserved data can be defined as
a file offset to additional fields (defined as 32-bit words from the
beginning of the header). This should make it technically simple to
add additional fields in the future. For example, RISC-V, and probably
other architectures, will want to add some way to indicate which ISA
extensions are necessary to run the kernel image. Right now there
won't be any fields defined, so we can leave the format undefined for
the moment also. Let's stipulate for version 1 that this field
should be fixed at 0, indicating no additional fields.

4. Document all of this, in this patch, in a file such as
Documentation/riscv/boot-image-header.txt. If
we're able to reach agreement with other maintainers, then we
can move this file out into a common, non-architecture-specific
documentation location.


thanks

- Paul