Re: SYMTAB and DYNSYM tables relationship question

[Vladimir Simonov <sv at sw dot ru>]

On 07/01/2010 04:35 AM, Ian Lance Taylor wrote:
> Vladimir Simonov<sv@sw.ru>  writes:
>
> > I have ELF executable (EXE) dynamically linked with DSO (SO) on
> > some dos-like OS with ELF support. Some symbol A (FUNC GLOBAL DEFAULT)
> > is referenced only from SO(not from EXE itself). After link
> > (by ld, binutils 2.20.1) EXE contains info about symbol A only
> > in DYNSYM table. There is nothing about A in SYMTAB and .rel.text.
> > As result our runtime linker is unable to find symbol A during SO
> > loading. If I reference A from EXE, A appears in SYMTAB, DYNSYM
> > and .rel.text and runtime works OK.
> >
> > Questions:
> > 1. Must symbol A be in SYMTAB and .rel.text if it is not
> > referenced from EXE directly? By other words, is the problem
> > in ld or in runtime linker?
>
> An ELF dynamic linker can be implemented to look only at the dynamic
> symbol table.  In fact, an ELF dynamic linker should normally not look
> at the normal symbol at all.  So I think there is something wrong with
> your dynamic linker.
>
> > 2. If the problem is in runtime linker, then is it possible
> > to force such records creation without explicit reference on A
> > from EXE sources? Via ld's version-script, may be options...
>
> I'm not aware of any way to do this.
>
> Ian

Thank you for answers.

In reality all not so simple - our dos-like OS EXE requires TEXT and PLT
sections relocation (a program works in physical address space and can't
be loaded at desired address).

Investigations show that "strip --strip-all" and even "strip --strip-unneeded"
remove symtab and necessary relocation sections. As I understand EXE's "text"
relocation can't be done without symtab because relocation entry references
symtab.

Currently, looks like "strip --strip-debug --discard-all" keeps
all required info. Runtime somehow works :)

Thanks again
Regards
Vladimir Simonov