+++ Josh Poimboeuf [04/06/18 08:16 -0500]:
On Mon, Jun 04, 2018 at 03:01:31PM +0200, Jessica Yu wrote:
+++ Jessica Yu [04/06/18 11:54 +0200]:
> +++ Jessica Yu [04/06/18 10:05 +0200]:
> > +++ Josh Poimboeuf [02/06/18 12:32 -0500]:
> > > Hi Jessica,
> > >
> > > I found a bug:
> > >
> > > [root@f25 ~]# modprobe livepatch-sample
> > > [root@f25 ~]# grep ' u ' /proc/kallsyms
> > > ffffffff81161080 u klp_enable_patch [livepatch_sample]
> > > ffffffff81a01800 u __fentry__ [livepatch_sample]
> > > ffffffff81161250 u klp_unregister_patch [livepatch_sample]
> > > ffffffff81161870 u klp_register_patch [livepatch_sample]
> > > ffffffff8131f0b0 u seq_printf [livepatch_sample]
> > >
> > > Notice that livepatch modules' undefined symbols are showing up in
> > > /proc/kallsyms. This can confuse klp_find_object_symbol() which can
> > > cause subtle bugs in livepatch.
> > >
> > > I stared at the module kallsyms code for a bit, but I don't see the bug.
> > > Maybe it has something to do with how we save the symbol table in
> > > copy_module_elf(). Any ideas?
> >
> > Hi Josh!
> >
> > This is because we preserve the entire symbol table for livepatch
> > modules, including the SHN_UNDEF symbols. IIRC, this is so that we can
> > still apply relocations properly with apply_relocate_add() after a
> > to-be-patched object is loaded. Normally we don't save these SHN_UNDEF
> > symbols for modules so they do not appear in /proc/kallsyms.
>
> Hm, if having the full symtab in kallsyms is causing trouble, one
> possibility would be to just have the module kallsyms code simply
> skip/ignore undef symbols. That's what we technically do for normal
> modules anyway (we normally cut undef syms out of the symtab). Haven't
> tested this idea but does that sound like it'd help?
See if the following patch (untested) helps. It does not fix the
/proc/kallsyms lookup, that requires a separate patch. But it should
exclude the undef symbols from module_kallsyms_on_each_symbol() and
thus also from klp_find_object_symbol().
That seems like it would work. But wouldn't it be more robust if we
don't store the SHN_UNDEF symbols to start with? Really it's only the
SHN_LIVEPATCH symbols that we need to keep, right?
Hm, didn't we have a reason for preserving the original symbol tables?
It's not that we're storing the undef symbols on purpose, it's just
that in the case of a livepatch module we keep *all* the symbols.
IIRC, the original reasoning was that we needed to keep the original
symtab indices in place, otherwise apply_relocate() (which uses the
original symbol index encoded in the relocation) won't work.
Selectively copying only SHN_LIVEPATCH symbols when building the
module's symbol table will unfortunately mess the symbol indices up :/
Under these restrictions, it just seemed easier to make kallsyms
ignore undef symbols.
What do you think about the following (untested)?
diff --git a/kernel/module.c b/kernel/module.c
index c9bea7f2b43e..78ec9de856e3 100644
--- a/kernel/module.c
+++ b/kernel/module.c
@@ -2586,6 +2586,9 @@ static bool is_core_symbol(const Elf_Sym *src, const
Elf_Shdr *sechdrs,
{
const Elf_Shdr *sec;
+ if (src->st_shndx == SHN_LIVEPATCH)
+ return true;
+
if (src->st_shndx == SHN_UNDEF
|| src->st_shndx >= shnum
|| !src->st_name)
@@ -2632,9 +2635,9 @@ static void layout_symtab(struct module *mod, struct
load_info *info)
/* Compute total space required for the core symbols' strtab. */
for (ndst = i = 0; i < nsrc; i++) {
- if (i == 0 || is_livepatch_module(mod) ||
- is_core_symbol(src+i, info->sechdrs, info->hdr->e_shnum,
- info->index.pcpu)) {
+ if (i == 0 || is_core_symbol(src+i, info->sechdrs,
+ info->hdr->e_shnum,
+ info->index.pcpu)) {
strtab_size += strlen(&info->strtab[src[i].st_name])+1;
ndst++;
}
@@ -2691,9 +2694,9 @@ static void add_kallsyms(struct module *mod, const struct
load_info *info)
mod->core_kallsyms.strtab = s = mod->core_layout.base + info->stroffs;
src = mod->kallsyms->symtab;
for (ndst = i = 0; i < mod->kallsyms->num_symtab; i++) {
- if (i == 0 || is_livepatch_module(mod) ||
- is_core_symbol(src+i, info->sechdrs, info->hdr->e_shnum,
- info->index.pcpu)) {
+ if (i == 0 || is_core_symbol(src+i, info->sechdrs,
+ info->hdr->e_shnum,
+ info->index.pcpu)) {
dst[ndst] = src[i];
dst[ndst++].st_name = s - mod->core_kallsyms.strtab;
s += strlcpy(s, &mod->kallsyms->strtab[src[i].st_name],
