On Fri, 2018-01-05 at 09:28 -0800, Linus Torvalds wrote:
> 
> Yes, I would suggest against expecting altinstructions to have
> relocation information. They are generated in a different place, so..
> 
> That said, I honestly like the inline version (the one that is in the
> google paper first) of the retpoline more than the out-of-line one.
> And that one shouldn't have any relocation issues, because all the
> offsets are relative.

Note that the *only* issue with the relocation is that it pushes me to
use X86_FEATURE_NO_RETPOLINE for my feature instead of
X86_FEATURE_RETPOLINE as might be natural. And actually there's a
motivation to do that anyway, because of the way three-way alternatives
interact.

With the existing negative flag I can do 

 ALTERNATIVE_2(retpoline, K8: lfence+jmp; NO_RETPOLINE: jmp)

But if I invert it, I think I need two feature flags to get the same functionality — X86_FEATURE_RETPOLINE and X86_FEATURE_RETPOLINE_AMD:

 ALTERNATIVE_2(jmp, RETPOLINE: retpoline, RETPOLINE_AMD: lfence+jmp)

So I was completely prepared to live with the slightly unnatural
inverse logic of the feature flag. But since you asked...

> We want to use that one for the entry stub anyway, can't we just
> standardize on that one for all our assembly?
> 
> If the *compiler* uses the out-of-line version, that's a separate
> thing. But for our asm cases, let's just make it all be the inline
> case, ok?

OK.... it starts off looking a bit like this. You're right; with the
caveats above it will let me invert the logic to X86_FEATURE_RETPOLINE
because the alternatives mechanism no longer needs to adjust any part
of the retpoline code path when it's in 'altinstr'.

And it does let me use a simple NOSPEC_JMP in the entry trampoline
instead of open-coding it again, which is nice.

But the first pass of it, below, is fugly as hell. I'll take another
look at *using* the ALTERNATIVE_2 macro instead of reimplementing it
for NOSPEC_CALL, but I strongly suspect that's just going to land me
with a fairly unreadable __stringify(jmp;call;lfence;jmp;call;mov;ret)
monstrosity all on a single line. Assembler macros are... brittle.

diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S
index 76f94bbacaec..8f7e1129f493 100644
--- a/arch/x86/entry/entry_64.S
+++ b/arch/x86/entry/entry_64.S
@@ -188,17 +188,7 @@ ENTRY(entry_SYSCALL_64_trampoline)
 	 */
 	pushq	%rdi
 	movq	$entry_SYSCALL_64_stage2, %rdi
-	/*
-	 * Open-code the retpoline from retpoline.S, because we can't
-	 * just jump to it directly.
-	 */
-	ALTERNATIVE "call 2f", "jmp *%rdi", X86_FEATURE_NO_RETPOLINE
-1:
-	lfence
-	jmp	1b
-2:
-	mov	%rdi, (%rsp)
-	ret
+	NOSPEC_JMP rdi
 END(entry_SYSCALL_64_trampoline)
 
 	.popsection
diff --git a/arch/x86/include/asm/nospec-branch.h b/arch/x86/include/asm/nospec-branch.h
index eced0dfaddc9..1c8312ff186a 100644
--- a/arch/x86/include/asm/nospec-branch.h
+++ b/arch/x86/include/asm/nospec-branch.h
@@ -14,15 +14,54 @@
  */
 .macro NOSPEC_JMP reg:req
 #ifdef CONFIG_RETPOLINE
-	ALTERNATIVE __stringify(jmp __x86.indirect_thunk.\reg), __stringify(jmp *%\reg), X86_FEATURE_NO_RETPOLINE
+	ALTERNATIVE_2 "call 1112f", __stringify(lfence;jmp *%\reg), X86_FEATURE_K8, __stringify(jmp *%\reg), X86_FEATURE_NO_RETPOLINE
+1111:
+	lfence
+	jmp	1111b
+1112:
+	mov	%\reg, (%_ASM_SP)
+	ret
 #else
-	jmp *%\reg
+	jmp	*%\reg
 #endif
 .endm
 
+/*
+ * Even __stringify() on the arguments doesn't really make it nice to use
+ * the existing ALTERNATIVE_2 macro here. So open-code our own version...
+ */
 .macro NOSPEC_CALL reg:req
 #ifdef CONFIG_RETPOLINE
-	ALTERNATIVE __stringify(call __x86.indirect_thunk.\reg), __stringify(call *%\reg), X86_FEATURE_NO_RETPOLINE
+140:
+	jmp	1113f
+1110:
+	call	1112f
+1111:
+	lfence
+	jmp	1111b
+1112:
+	mov	%\reg, (%_ASM_SP)
+	ret
+1113:
+	call	1110b
+141:
+	.skip -((alt_max_short(new_len1, new_len2) - (old_len)) > 0) * \
+		(alt_max_short(new_len1, new_len2) - (old_len)),0x90
+142:
+
+	.pushsection .altinstructions,"a"
+	altinstruction_entry 140b,143f,X86_FEATURE_K8,142b-140b,144f-143f,142b-141b
+	altinstruction_entry 140b,144f,X86_FEATURE_NO_RETPOLINE,142b-140b,145f-144f,142b-141b
+	.popsection
+
+	.pushsection .altinstr_replacement,"ax"
+143:
+	lfence
+	call	*%\reg
+144:
+	call	*%\reg
+145:
+	.popsection
 #else
 	call *%\reg
 #endif
diff --git a/arch/x86/lib/retpoline.S b/arch/x86/lib/retpoline.S
index 2a4b1f09eb84..5c15e4307da5 100644
--- a/arch/x86/lib/retpoline.S
+++ b/arch/x86/lib/retpoline.S
@@ -6,19 +6,14 @@
 #include <asm/cpufeatures.h>
 #include <asm/alternative-asm.h>
 #include <asm/export.h>
+#include <asm/nospec-branch.h>
 
 .macro THUNK sp reg
 	.section .text.__x86.indirect_thunk.\reg
 
 ENTRY(__x86.indirect_thunk.\reg)
 	CFI_STARTPROC
-	ALTERNATIVE_2 "call 2f", __stringify(lfence;jmp *%\reg), X86_FEATURE_K8, __stringify(jmp *%\reg), X86_FEATURE_NO_RETPOLINE
-1:
-	lfence
-	jmp	1b
-2:
-	mov	%\reg, (%\sp)
-	ret
+	NOSPEC_JMP \reg
 	CFI_ENDPROC
 ENDPROC(__x86.indirect_thunk.\reg)
 EXPORT_SYMBOL(__x86.indirect_thunk.\reg)