Re: [PATCH 2/2] x86/boot/compressed: Remove unnecessary sections from bzImage

[Fangrui Song <maskray@google.com>]

On 2020-02-24, Nick Desaulniers wrote:
>On Mon, Feb 24, 2020 at 5:28 AM Michael Matz <matz@suse.de> wrote:
>>
>> Hello,
>>
>> On Sat, 22 Feb 2020, Nick Desaulniers wrote:
>>
>> > > > > In GNU ld, it seems that .shstrtab .symtab and .strtab are special
>> > > > > cased. Neither the input section description *(.shstrtab) nor *(*)
>> > > > > discards .shstrtab . I feel that this is a weird case (probably even a bug)
>> > > > > that lld should not implement.
>> > > >
>> > > > Ok, forget what the tools do for a second: why is .shstrtab special and
>> > > > why would one want to keep it?
>> > > >
>> > > > Because one still wants to know what the section names of an object are
>> > > > or other tools need it or why?
>> > > >
>> > > > Thx.
>> > > >
>> > > > --
>> > > > Regards/Gruss,
>> > > >     Boris.
>> > > >
>> > > > https://people.kernel.org/tglx/notes-about-netiquette
>> > >
>> > > .shstrtab is required by the ELF specification. The e_shstrndx field in
>> > > the ELF header is the index of .shstrtab, and each section in the
>> > > section table is required to have an sh_name that points into the
>> > > .shstrtab.
>> >
>> > Yeah, I can see it both ways.  That `*` doesn't glob all remaining
>> > sections is surprising to me, but bfd seems to be "extra helpful" in
>> > not discarding sections that are required via ELF spec.
>>
>> In a way the /DISCARD/ assignment should be thought of as applying to
>> _input_ sections (as all such section references on the RHS), not
>> necessarily to output sections.  What this then means for sections that
>> are synthesized by the link editor is less clear.  Some of them are
>> generated regardless (as you noted, e.g. the symbol table and associated
>> string sections, including section name string table), some of them are
>> suppressed, and either lead to an followup error (e.g. with .gnu.hash), or
>> to invalid output (e.g. missing .dynsym for executables simply lead to
>> segfaults when running them).

Hi Michael, please see my other reply on this thread: https://lkml.org/lkml/2020/2/24/47

Synthesized sections can be matched as well. For example, SECTIONS { .pltfoo : { *(.plt) }} can rename the output section .plt to .pltfoo
It seems that in GNU ld, the synthesized section is associated with the
original object file, so it can be written as:

   SECTIONS { .pltfoo : { a.o(.plt) }}

In lld, you need a wildcard to match the synthesized section *(.plt)

.rela.dyn is another example.

>> That's the reason for the perceived inconsistency with behaviour on '*':
>> it's application to synthesized sections.  Arguably bfd should be fixed to
>> also not discard the other essential sections (or alternatively to give an
>> error when an essential section is discarded).  The lld behaviour of e.g.
>> discarding .shstrtab (or other synthesized sections necessary for valid
>> ELF output) doesn't make much sense either, though.

I think most input section descriptions *(*) are misuse. They really
should be INPUT_SECTION_FLAGS(SHF_ALLOC) *(*)

>Hi Michael, thank you for the precise feedback.  Do you have a list of
>"synthesized sections necessary for valid ELF output?" Also, could you
>point me to the documentation about `*` and its relation to
>"synthesized sections necessary for valid ELF output?"  This will help
>me file a precise bug against LLD.

https://sourceware.org/binutils/docs/ld/Output-Section-Discarding.html#Output-Section-Discarding

has a few words on this topic. A large part is implementation defined.
In GNU ld, the implementation is mostly in ld/ldlang.c and ld/ldexp.c
(very long).