On Wed, Oct 23, 2019 at 01:48:35PM +0200, Peter Zijlstra wrote:
> On Mon, Oct 21, 2019 at 04:14:02PM +0200, Peter Zijlstra wrote:
> > On Mon, Oct 21, 2019 at 08:53:12AM -0500, Josh Poimboeuf wrote:
>
> > > Doesn't livepatch code also need to be modified? We have:
> >
> > Urgh bah.. I was too focussed on the other klp borkage :/ But yes,
> > arm64/ftrace and klp are the only two users of that function (outside of
> > module.c) and Mark was already writing a patch for arm64.
> >
> > Means these last two patches need to wait a little until we've fixed
> > those.
>
> So the other KLP issue:
>
> <mbenes> peterz: ad klp, apply_relocate_add() and text_poke()... what
> about apply_alternatives() and apply_paravirt()? They call
> text_poke_early(), which was ok with module_disable/enable_ro(), but
> now it is not, I suppose. See arch_klp_init_object_loaded() in
> arch/x86/kernel/livepatch.c.
>
> <peterz> mbenes: hurm, I don't see why we would need to do
> apply_alternatives() / apply_paravirt() later, why can't we do that
> the moment we load the module
>
> <peterz> mbenes: that is, those things _never_ change after boot
>
> <mbenes> peterz: as jpoimboe explained. See commit
> d4c3e6e1b193497da3a2ce495fb1db0243e41e37 for detailed explanation.
>
> Now sadly that commit missed all the useful information, luckily I could
> find the patch in my LKML folder, more sad, that thread still didn't
> contain the actual useful information, for that I was directed to
> github:
>
> https://github.com/dynup/kpatch/issues/580
>
> Now, someone is owning me a beer for having to look at github for this.
>
> That finally explained that what happens is that the RELA was trying to
> fix up the paravirt indirect call to 'local_irq_disable', which
> apply_paravirt() will have overwritten with 'CLI; NOP'. This then
> obviously goes *bang*.
>
> This then raises a number of questions:
>
> 1) why is that RELA (that obviously does not depend on any module)
> applied so late?
>
> 2) why can't we unconditionally skip RELA's to paravirt sites?
>
> 3) Is there ever a possible module-dependent RELA to a paravirt /
> alternative site?
>
>
> Now, for 1), I would propose '.klp.rela.${mod}' sections only contain
> RELAs that depend on symbols in ${mod} (or modules in general). We can
> fix up RELAs that depend on core kernel early without problems. Let them
> be in the normal .rela sections and be fixed up on loading the
> patch-module as per usual.
>
> This should also deal with 2, paravirt should always have RELAs into the
> core kernel.
>
> Then for 3) we only have alternatives left, and I _think_ it unlikely to
> be the case, but I'll have to have a hard look at that.
>
> Hmmm ?
Something like so on top, I suppose.
---
--- a/arch/x86/kernel/module.c
+++ b/arch/x86/kernel/module.c
@@ -131,7 +131,8 @@ static int __apply_relocate_add(Elf64_Sh
unsigned int symindex,
unsigned int relsec,
struct module *me,
- void *(*write)(void *addr, const void *val, size_t size))
+ void *(*write)(void *addr, const void *val, size_t size),
+ bool klp)
{
unsigned int i;
Elf64_Rela *rel = (void *)sechdrs[relsec].sh_addr;
@@ -157,6 +158,14 @@ static int __apply_relocate_add(Elf64_Sh
val = sym->st_value + rel[i].r_addend;
+ /*
+ * .klp.rela.* sections should only contain module
+ * related RELAs. All core-kernel RELAs should be in
+ * normal .rela.* sections and be applied when loading
+ * the patch module itself.
+ */
+ WARN_ON_ONCE(klp && core_kernel_text(val));
+
switch (ELF64_R_TYPE(rel[i].r_info)) {
case R_X86_64_NONE:
break;
@@ -223,7 +232,7 @@ int apply_relocate_add(Elf64_Shdr *sechd
unsigned int relsec,
struct module *me)
{
- return __apply_relocate_add(sechdrs, strtab, symindex, relsec, me,
memcpy);
+ return __apply_relocate_add(sechdrs, strtab, symindex, relsec, me,
memcpy, false);
}
int klp_apply_relocate_add(Elf64_Shdr *sechdrs,
@@ -234,7 +243,7 @@ int klp_apply_relocate_add(Elf64_Shdr *s
{
int ret;
- ret = __apply_relocate_add(sechdrs, strtab, symindex, relsec, me,
text_poke);
+ ret = __apply_relocate_add(sechdrs, strtab, symindex, relsec, me,
text_poke, true);
if (!ret)
text_poke_sync();
