Re: [PATCH] RISC-V: Add an Image header that boot loader can parse.

[Karsten Merker <merker@debian.org>]

On Wed, May 01, 2019 at 10:41:52PM +0530, Anup Patel wrote:
> On Wed, May 1, 2019 at 10:30 PM Mark Rutland <mark.rutland@arm.com> wrote:
> > On Mon, Apr 29, 2019 at 10:42:40PM -0700, Atish Patra wrote:
> > > On 4/29/19 4:40 PM, Palmer Dabbelt wrote:
> > > > On Tue, 23 Apr 2019 16:25:06 PDT (-0700), atish.patra@wdc.com wrote:
> > > > > Currently, last stage boot loaders such as U-Boot can accept only
> > > > > uImage which is an unnecessary additional step in automating boot flows.
> > > > >
> > > > > Add a simple image header that boot loaders can parse and directly
> > > > > load kernel flat Image. The existing booting methods will continue to
> > > > > work as it is.
> > > > >
> > > > > Tested on both QEMU and HiFive Unleashed using OpenSBI + U-Boot + Linux.
> > > > >
> > > > > Signed-off-by: Atish Patra <atish.patra@wdc.com>
> > > > > ---
> > > > >   arch/riscv/include/asm/image.h | 32 ++++++++++++++++++++++++++++++++
> > > > >   arch/riscv/kernel/head.S       | 28 ++++++++++++++++++++++++++++
> > > > >   2 files changed, 60 insertions(+)
> > > > >   create mode 100644 arch/riscv/include/asm/image.h
> > > > >
> > > > > diff --git a/arch/riscv/include/asm/image.h b/arch/riscv/include/asm/image.h
> > > > > new file mode 100644
> > > > > index 000000000000..76a7e0d4068a
> > > > > --- /dev/null
> > > > > +++ b/arch/riscv/include/asm/image.h
> > > > > @@ -0,0 +1,32 @@
> > > > > +/* SPDX-License-Identifier: GPL-2.0 */
> > > > > +
> > > > > +#ifndef __ASM_IMAGE_H
> > > > > +#define __ASM_IMAGE_H
> > > > > +
> > > > > +#define RISCV_IMAGE_MAGIC        "RISCV"
> > > > > +
> > > > > +#ifndef __ASSEMBLY__
> > > > > +/*
> > > > > + * struct riscv_image_header - riscv kernel image header
> > > > > + *
> > > > > + * @code0:               Executable code
> > > > > + * @code1:               Executable code
> > > > > + * @text_offset: Image load offset
> > > > > + * @image_size:          Effective Image size
> > > > > + * @reserved:            reserved
> > > > > + * @magic:               Magic number
> > > > > + * @reserved:            reserved
> > > > > + */
> > > > > +
> > > > > +struct riscv_image_header {
> > > > > + u32 code0;
> > > > > + u32 code1;
> > > > > + u64 text_offset;
> > > > > + u64 image_size;
> > > > > + u64 res1;
> > > > > + u64 magic;
> > > > > + u32 res2;
> > > > > + u32 res3;
> > > > > +};
> > > >
> > > > I don't want to invent our own file format.  Is there a reason we can't just
> > > > use something standard?  Off the top of my head I can think of ELF files and
> > > > multiboot.
> > >
> > > Additional header is required to accommodate PE header format. Currently,
> > > this is only used for booti command but it will be reused for EFI headers as
> > > well. Linux kernel Image can pretend as an EFI application if PE/COFF header
> > > is present. This removes the need of an explicit EFI boot loader and EFI
> > > firmware can directly load Linux (obviously after EFI stub implementation
> > > for RISC-V).
> >
> > Adding the EFI stub on arm64 required very careful consideration of our
> > Image header and the EFI spec, along with the PE/COFF spec.
> >
> > For example, to be a compliant PE/COFF header, the first two bytes of
> > your kernel image need to be "MZ" in ASCII. On arm64 we happened to find
> > a valid instruction that we could rely upon that met this requirement...
> 
> The "MZ" ASCII (i.e. 0x5a4d) is "li s4,-13" instruction in RISC-V so this
> modifies "s4" register which is pretty harmless from Linux RISC-V booting
> perspective.
> 
> Of course, we should only add "MZ" ASCII in Linux RISC-V image header
> when CONFIG_EFI is enabled (just like Linux ARM64).

Probably I'm missing something obvious, but I cannot completely
follow you here. My understanding is as follows:

The kernel gets executed by jumping to the _start label, where it
currently immediately starts with setting up everything (mask
interrupts, disable FPU, etc).  Now we insert a structure before
the actual init code where the first 64 bits of the structure
(the code0 and code1 fields) are filled with values that
constitute a valid jump opcode to the actual init code behind the
end of the new structure.

If the first byte in a PE/COFF header has to be an ASCII "M",
that is 01001101 in binary.  RISC-V is little-endian and the last
two bits of the lowest-value byte define the type of instruction. 
According to the chapter "Base Instruction-Length Encoding" in
the RISC-V ISA spec everything except 11 as the lowest bits
denotes a compressed instruction and if I have puzzeled together
the the various instruction bits correctly, ASCII "MZ" would be
excuted as a compressed load immediate to x9/s1, wouldn't it?

If my previous thoughts are correct, this would also mean that
having a PE/COFF-compatible header for EFI boot would make a
kernel image inherently incompatible with performing a non-EFI
boot of the same image on systems that don't implement the C
extension.  As the work-in-progress RISC-V Unix platform spec
will probably mandate C, that probably won't be much of a
practical problem, though.

Regards,
Karsten
-- 
Ich widerspreche hiermit ausdrücklich der Nutzung sowie der
Weitergabe meiner personenbezogenen Daten für Zwecke der Werbung
sowie der Markt- oder Meinungsforschung.