Re: livepatch: reuse module loader code to write relocations

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Thu, Dec 10, 2015 at 04:33:29PM -0500, Jessica Yu wrote:
> +++ Josh Poimboeuf [10/12/15 08:28 -0600]:
> >On Wed, Dec 09, 2015 at 02:10:14PM -0500, Jessica Yu wrote:
> >>+++ Josh Poimboeuf [08/12/15 12:38 -0600]:
> >>>>+	/* For each __klp_rela section for this object */
> >>>>+	klp_for_each_reloc_sec(obj, reloc_sec) {
> >>>>+		relindex = reloc_sec->index;
> >>>>+		num_relas = pmod->sechdrs[relindex].sh_size / sizeof(Elf_Rela);
> >>>>+		rela = (Elf_Rela *) pmod->sechdrs[relindex].sh_addr;
> >>>>+
> >>>>+		/* For each rela in this __klp_rela section */
> >>>>+		for (i = 0; i < num_relas; i++, rela++) {
> >>>>+			sym = symtab + ELF_R_SYM(rela->r_info);
> >>>>+			symname = pmod->core_strtab + sym->st_name;
> >>>>+
> >>>>+			if (sym->st_shndx == SHN_LIVEPATCH) {
> >>>>+				if (sym->st_info == 'K')
> >>>>+					ret = klp_find_external_symbol(pmod, symname, &addr);
> >>>>+				else
> >>>>+					ret = klp_find_object_symbol(obj->name, symname, &addr);
> >>>>+				if (ret)
> >>>>+					return ret;
> >>>>+				sym->st_value = addr;
> >>>
> >>>So I think you're also working on removing the 'external' stuff.  Any
> >>>idea how this code will handle that?  Specifically I'm wondering how the
> >>>objname and sympos of the rela's symbol will be specified.  Maybe it can
> >>>be encoded somehow in one of the symbol fields (st_value)?
> >>
> >>Thanks for bringing this up. I think we can just encode the symbol's
> >>position in kallsyms in the symbol's st_other field. It isn't used
> >>anywhere and has size char, which is plenty of bits to represent the
> >>small ints.
> >
> >st_other does seem to at least have some trivial usage in the kernel,
> >see print_absolute_symbols() and sym_visibility() in
> >arch/x86/tools/relocs.c.  Two of the bits are used to specify the
> >"visibility" of a symbol.  Also readelf shows a "Vis" column in the
> >symbol table.
> 
> Yeah, for x86 it looks like st_other is used only for SHN_ABS symbols
> in print_absolute_symbols(). Technically SHN_LIVEPATCH symbols
> shouldn't be affected in this case...but despite its sparse usage in the
> kernel it does look like using st_other to encode sympos is out of the
> question as its meaning is architecture specific..
> 
> >>For objname, the simplest solution might be to append ".klp.objname"
> >>to the symbol name, and extract out this suffix when resolving the
> >>symbol. Another way might be to have st_value contain the index into
> >>the strtab (or .kpatch.strings) that contains the objname. Then we'd
> >>access the objname just like how we access the symbol's name (strtab +
> >>sym->st_name). After we find the objname we can then overwrite
> >>st_value with the real address. I think this second method is cleaner.
> >>Thoughts?
> >
> >Yeah, I guess there are a lot of possibilities for ways to encode it.
> >
> >Personally I think it would be nice if the encoding were something that
> >could easily be seen when dumping the symbol table with readelf.  So,
> >for example, the objname could be encoded in the symbol name (as you
> >suggested), and the sympos could be in st_value.
> 
> Sure, that should be doable. So the new process might look like this:
> 
> For every livepatch symbol referenced by a rela..
> 
> 1) Save the sympos encoded in st_value
> 
> 2) Save the sym_objname that is encoded in the symbol's name with the
> 'klp' suffix (Just to clarify: the sym_objname specifies the object
> in which the symbol lives, and recall that we need this to remove the
> need for the "external" flag)
> 
> 3) Resolve the symbol by using its name (without the klp suffix),
> sympos, and sym_objname
> 
> 4) Set st_value to the found address

Sounds right to me.

> >If we do that, it'd probably be good to keep the naming consistent with
> >the '__klp_rela_objname.symname' thing.  So something like
> >'_klp_sym_objname.symname'.
> 
> How about 'symname.klp.objname', and renaming the klp reloc sections
> to '.klp.objname.rela.section_name'? Special symbol suffixes and
> section names seem to always use '.', so maybe this would look better?
> :-) But we can keep the underscores if people like that more. Both
> naming methods would work, it is only a matter of preference.

It's your patches, so I'd say you get to pick ;-)  My only request would
be some consistency between the symbol names and the rela section names.

> >But... would there be any side effects associated with renaming it?  For
> >example, would it cause problems with the s390 PLT?

Just to verify, did you see this question? :-)

-- 
Josh