Re: [PATCH v3 16/16] objtool,x86: Rewrite retpoline thunk calls

From: Fāng-ruì Sòng
Date: Mon Jun 07 2021 - 13:23:34 EST


On 2021-06-07, Peter Zijlstra wrote:
On Mon, Jun 07, 2021 at 11:22:11AM +0200, Peter Zijlstra wrote:
On Mon, Jun 07, 2021 at 09:56:48AM +0200, Peter Zijlstra wrote:
> On Sat, Jun 05, 2021 at 06:58:39PM -0700, Fāng-ruì Sòng wrote:
> > On Sat, Jun 5, 2021 at 3:39 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> > > I think you've absolutely nailed it; but would you have more information
> > > or a code reference to what you're speaking about? My complete ELF
> > > and libelf knowledge is very limited and as demonstrated here, I'm not
> > > at all sure how all that extended index stuff is supposed to work.
> >
> > The section index field of an Elf{32,64}_Sym (st_shndx) is 16-bit, so
> > it cannot represent a section index greater than 0xffff.
> > ELF actually reserves values in 0xff00~0xff00 for other purposes, so
> > st_shndx cannot represent a section whose index is greater or equal to
> > 0xff00.
>
> Right, that's about as far as I got, but never could find details on how
> the extension worked in detail, and I clearly muddled it :/

OK, so I'm all confused again...

So a .symtab entry has:

st_name -- strtab offset for the name string
st_value -- where this symbol lives
st_size -- size of symbol in bytes
st_shndx -- section index to interpret the @st_value above
st_info -- type+bind
st_other -- visibility

The thing is, we're adding UNDEF symbols, for the linker to resolve.
UNDEF has:

st_value := 0
st_size := 0
st_shndx := 0
st_info := GLOBAL + NOTYPE
st_other := 0

Per that, sh_shndx isn't >= SHN_LORESERVE, and I figured we all good.


Is the problem that .symtab_shndx is expected to contain the exact same
number of entries as .symtab? And I'm adding to .symtab and not to
.symtab_shndx, hence getting them out of sync?

Yes. http://www.sco.com/developers/gabi/latest/ch4.sheader.html says
"Each value corresponds one to one with a symbol table entry and appear in the same order as those entries."

Let me try adding 0s to .symtab_shndx. See if that makes readelf
happier.

That does indeed seem to do the trick. Bit daft if you ask me, anybody
reading that file ought to have a handy bucket of 0s available, but
whatever.

Does the representation use the section index directly? (sym->sym.st_shndx)
This can be fragile when the number of sections changes..., e.g. elf_add_section

So in llvm-objcopy's representation, the section index is represented as
the section object.

struct Symbol {
...
SectionBase *DefinedIn = nullptr;
...
};

In the writer stage, sections are assigned 32-bit indexes and the writer
knows that an SHN_XINDEX for a symbol is needed if the index is >= 0xff00.

---
tools/objtool/elf.c | 25 ++++++++++++++++++++++++-
1 file changed, 24 insertions(+), 1 deletion(-)

diff --git a/tools/objtool/elf.c b/tools/objtool/elf.c
index 743c2e9d0f56..41bca1d13d8e 100644
--- a/tools/objtool/elf.c
+++ b/tools/objtool/elf.c
@@ -717,7 +717,7 @@ static int elf_add_string(struct elf *elf, struct section *strtab, char *str)

struct symbol *elf_create_undef_symbol(struct elf *elf, const char *name)
{
- struct section *symtab;
+ struct section *symtab, *symtab_shndx;
struct symbol *sym;
Elf_Data *data;
Elf_Scn *s;
@@ -769,6 +769,29 @@ struct symbol *elf_create_undef_symbol(struct elf *elf, const char *name)
symtab->len += data->d_size;
symtab->changed = true;

+ symtab_shndx = find_section_by_name(elf, ".symtab_shndx");
+ if (symtab_shndx) {
+ s = elf_getscn(elf->elf, symtab_shndx->idx);
+ if (!s) {
+ WARN_ELF("elf_getscn");
+ return NULL;
+ }
+
+ data = elf_newdata(s);
+ if (!data) {
+ WARN_ELF("elf_newdata");
+ return NULL;
+ }
+
+ data->d_buf = &sym->sym.st_size; /* conveniently 0 */
+ data->d_size = sizeof(Elf32_Word);
+ data->d_align = 4;
+ data->d_type = ELF_T_WORD;
+
+ symtab_shndx->len += 4;
+ symtab_shndx->changed = true;
+ }
+
sym->sec = find_section_by_index(elf, 0);

elf_add_symbol(elf, sym);