Re: [PATCH 3/3] x86/bugs: Remove support for Spectre v2 LFENCE "retpolines"

[Josh Poimboeuf]

On Fri, Apr 12, 2024 at 11:10:34AM -0700, Josh Poimboeuf wrote:
> I found several bugs where code assumes that X86_FEATURE_RETPOLINE
> actually means retpolines (imagine that!).  In fact that feature also
> includes the original AMD LFENCE "retpolines", which aren't in fact
> retpolines.
> 
> Really, those "retpolines" should just be removed.  They're already
> considered vulnerable due to the fact that the speculative window after
> the indirect branch can still be long enough to do multiple dependent
> loads.  And recent tooling makes such gadgets easier to find.
> 
> Also, EIBRS_LFENCE tests worse in real-world benchmarks than the actual
> BHI mitigations, so it's both slower and less secure.
> 
> Specifically this removes support for the following cmdline options:
> 
>   - spectre_v2=retpoline,amd
>   - spectre_v2=retpoline,lfence
>   - spectre_v2=eibrs,lfence
> 
> Now when any of those options are used, it will print an error and fall
> back to the defaults (spectre_v2=auto spectre_bhi=on).
> 
> Signed-off-by: Josh Poimboeuf <jpoimboe@xxxxxxxxxx>

Compile fix:

diff --git a/arch/x86/include/asm/disabled-features.h b/arch/x86/include/asm/disabled-features.h
index e1c421282a78..3a1349c0f225 100644
--- a/arch/x86/include/asm/disabled-features.h
+++ b/arch/x86/include/asm/disabled-features.h
@@ -53,7 +53,7 @@
 #ifdef CONFIG_MITIGATION_RETPOLINE
 # define DISABLE_RETPOLINE	0
 #else
-# define DISABLE_RETPOLINE	(1 << (X86_FEATURE_RETPOLINE & 31)
+# define DISABLE_RETPOLINE	(1 << (X86_FEATURE_RETPOLINE & 31))
 #endif
 
 #ifdef CONFIG_MITIGATION_RETHUNK