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

[Peter Zijlstra]

On Wed, Jun 09, 2021 at 02:23:28PM +0200, Lukasz Majczak wrote:
> śr., 9 cze 2021 o 09:20 Peter Zijlstra <peterz@xxxxxxxxxxxxx> napisał(a):
> >
> > On Wed, Jun 09, 2021 at 09:11:18AM +0200, Lukasz Majczak wrote:
> >
> > > I'm sorry I was on vacation last week - do you still need the requested debugs?
> >
> > If the patch here:
> >
> >   https://lkml.kernel.org/r/YL3q1qFO9QIRL/BA@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
> >
> > does not fix things for you (don't think it actually will), then yes,
> > please send me the information requested.
> 
> Ok, it didn't help. Peter, Josh I have sent you a private email with
> requested information.

OK, I think I've found it. Check this one:

 5d5:   0f 85 00 00 00 00       jne    5db <cpuidle_reflect+0x22>       5d7: R_X86_64_PLT32     __x86_indirect_thunk_r11-0x4


+Relocation section '.rela.altinstructions' at offset 0 contains 14 entries:
+    Offset             Info             Type               Symbol's Value  Symbol's Name + Addend

+0000000000000018  0000000200000002 R_X86_64_PC32          0000000000000000 .text + 5d5
+000000000000001c  0000009200000002 R_X86_64_PC32          0000000000000000 __x86_indirect_alt_call_r11 + 0

Apparently we get conditional branches to retpoline thunks and objtool
completely messes that up. I'm betting this also explains the problems
Nathan is having.

*groan*,.. not sure what to do about this, except return to having
objtool generate code, which everybody hated on. For now I'll make it
skip the conditional branches.

I wonder if the compiler will also generate conditional tail calls, and
what that does with static_call... now I have to check all that.

---

diff --git a/tools/objtool/arch/x86/decode.c b/tools/objtool/arch/x86/decode.c
index 24295d39713b..523aa4157f80 100644
--- a/tools/objtool/arch/x86/decode.c
+++ b/tools/objtool/arch/x86/decode.c
@@ -747,6 +747,10 @@ int arch_rewrite_retpolines(struct objtool_file *file)
 
 	list_for_each_entry(insn, &file->retpoline_call_list, call_node) {
 
+		if (insn->type != INSN_JUMP_DYNAMIC &&
+		    insn->type != INSN_CALL_DYNAMIC)
+			continue;
+
 		if (!strcmp(insn->sec->name, ".text.__x86.indirect_thunk"))
 			continue;