Re: [PATCH v4 20/26] x86/build: Make generated PE more spec compliant

[Evgeniy Baskov]

On 2023-03-11 20:31, Ard Biesheuvel wrote:
> On Sat, 11 Mar 2023 at 16:02, Evgeniy Baskov <baskov@xxxxxxxxx> wrote:
> > On 2023-03-10 18:17, Ard Biesheuvel wrote:
> > > On Thu, 15 Dec 2022 at 13:42, Evgeniy Baskov <baskov@xxxxxxxxx> wrote:
> > >>
> > >> Currently kernel image is not fully compliant PE image, so it may
> > >> fail to boot with stricter implementations of UEFI PE loaders.
> > >>
> > >> Set minimal alignments and sizes specified by PE documentation [1]
> > >> referenced by UEFI specification [2]. Align PE header to 8 bytes.
> > >>
> > >> Generate PE sections dynamically. This simplifies code, since with
> > >> current implementation all of the sections needs to be defined in
> > >> header.S, where most section header fields do not hold valid values,
> > >> except for their names. Before the change, it also held flags,
> > >> but now flags depend on kernel configuration and it is simpler
> > >> to set them from build.c too.
> > >>
> > >> Setup sections protection. Since we cannot fit every needed section,
> > >> set a part of protection flags dynamically during initialization.
> > >> This step is omitted if CONFIG_EFI_DXE_MEM_ATTRIBUTES is not set.
> > >>
> > >> [1]
> > >> https://download.microsoft.com/download/9/c/5/9c5b2167-8017-4bae-9fde-d599bac8184a/pecoff_v83.docx
> > >> [2]
> > >> https://uefi.org/sites/default/files/resources/UEFI_Spec_2_9_2021_03_18.pdf
> > >>
> > >> Tested-by: Peter Jones <pjones@xxxxxxxxxx>
> > >> Signed-off-by: Evgeniy Baskov <baskov@xxxxxxxxx>
> > >
> > > I would prefer it if we didn't rewrite the build tool this way.
> > >
> > > Having the sections in header.S in the order they appear in the binary
> > > is rather useful, and I don't think we should manipulate the section
> > > flags based on whether CONFIG_DXE_MEM_ATTRIBUTES is set. I also don't
> > > think we need more than .text / .,data (as discussed in the other
> > > thread on linux-efi@)
> > >
> > > Furthermore, I had a look at the audk PE loader [0], and I think it is
> > > being overly pedantic.
> > >
> > > The PE/COFF spec does not require that all sections are virtually
> > > contiguous, and it does not require that the file content is
> > > completely covered by either the header or by a section.
> > >
> > > So what I would prefer to do is the following:
> > >
> > > Sections:
> > > Idx Name          Size     VMA              Type
> > >   0 .reloc        00000200 0000000000002000 DATA
> > >   1 .compat       00000200 0000000000003000 DATA
> > >   2 .text         00bee000 0000000000004000 TEXT
> > >   3 .data         00002200 0000000000bf2000 DATA
> > >
> > > using 4k section alignment and 512 byte file alignment, and a header
> > > size of 0x200 as before (This requires my patch that allows the setup
> > > header to remain unmapped when running the stub [1])
> > >
> > > The reloc and compat payloads are placed at the end of the setup data
> > > as before, but increased in size to 512 bytes each, and then mapped
> > > non-1:1 into the RVA space.
> > >
> > > This works happily with both the existing PE loader as well as the
> > > audk one, but with the pedantic flags disabled.
> > >
> >
> > This makes sense. I'll change this patch to use this layout and
> > to keep sections in headers.S before sending v5. (and I guess I'll
> > make the compressed kernel a part of .text). I have a few questions
> > though:
> >
> > This layout assumes having the local copy of the bootparams as
> > in your RFC patches, right?
>
> Indeed. Otherwise, the setup header may not have been copied to memory
> by the loader.
>
> > Can I keep the .rodata -- 5th section fits in the section table
> > without much work?
>
> You could, but at least the current PE/COFF loader in EDK2 will map it
> read/write, as it only distinguishes between executable sections and
> non-executable sections.

At least it will slightly improve security for some implementations
(e.g. audk, while being overly strict support RO sections)

> > Also, why .reloc is at offset 0x2000 and not just 0x1000, is there
> > anything important I am missing? I understand that is cannot be 0
> > and should be aligned on page size, but nothing else comes to my
> > mind...
>
> That was just arbitrary, because the raw allocations of reloc and
> compat are also allocated towards the end. But I guess starting at
> 0x1000 for .reloc makes more sense so feel free to change that.

Thanks for clarifications!