Re: [PATCH] x86/bugs: Qualify RETBLEED_INTEL_MSG

[Pawan Gupta <pawan.kumar.gupta@linux.intel.com>]

On Tue, Oct 07, 2025 at 05:14:29PM -0700, Josh Poimboeuf wrote:
> On Tue, Oct 07, 2025 at 04:08:21PM -0700, Pawan Gupta wrote:
> > On Wed, Oct 08, 2025 at 12:12:29AM +0200, Borislav Petkov wrote:
> > > On Tue, Oct 07, 2025 at 11:22:57AM -0700, Pawan Gupta wrote:
> > > > IMO, rather than targeting the mitigation enabling code it might make more
> > > > sense to compile out the actual mitigations scattered accross the kernel.
> > > > This may also improve performance by reducing the code footprint, and can
> > > > also help getting a cleaner disassembly.
> 
> Isn't that what CONFIG_CPU_MITIGATIONS=n already does today?

Not entirely, looking at entry_64.o for example, I see 243 NOPs with
CONFIG_CPU_MITIGATIONS=n. I suspect a lot of them are due to ALTERNATIVEs
for mitigations. In the same file, clear_bhb_loop() is not compliled out
and adds ~160 bytes.

Another example is process.c which still contains below functions:

0000000000000000 <speculation_ctrl_update_tif>:
00000000000008b0 <speculative_store_bypass_ht_init>:
0000000000000980 <speculation_ctrl_update>:
0000000000000c30 <speculation_ctrl_update_current>:

> > > Probably... however, it needs to be done smartly because sprinkling ifdeffery
> > > and turning what is an already unreadable mess into a bigger abomination,
> > > won't fly. Perhaps split out the mitigations glue into separate compilation
> > > units and build-disable them... we'll see.
> > 
> > Ya, that would be better.
> 
> On a sort of related note, it's confusing that there are two completely
> different classes of MITIGATION options which get conflated:
> 
> 1) compile the kernel with support for certain mitigations:
> 
>     MITIGATION_RETPOLINE
>     MITIGATION_RETHUNK
>     MITIGATION_PAGE_TABLE_ISOLATION
>     MITIGATION_UNRET_ENTRY
>     MITIGATION_CALL_DEPTH_TRACKING
>     MITIGATION_IBPB_ENTRY
>     MITIGATION_IBRS_ENTRY
>     MITIGATION_SRSO
>     MITIGATION_SLS
> 
> 2) enable bug-specific runtime defaults:
> 
>     MITIGATION_GDS
>     MITIGATION_RFDS
>     MITIGATION_SPECTRE_BHI
>     MITIGATION_MDS
>     MITIGATION_TAA
>     MITIGATION_MMIO_STALE_DATA
>     MITIGATION_L1TF
>     MITIGATION_RETBLEED
>     MITIGATION_SPECTRE_V1
>     MITIGATION_SPECTRE_V2
>     MITIGATION_SRBDS
>     MITIGATION_SSB
>     MITIGATION_ITS
>     MITIGATION_TSA
>     MITIGATION_VMSCAPE
> 
> In general, #1 uglify the kernel.  And some #2 depend on #1.
> 
> IMO, we should rename the #1 options.  For example:
> 
>   MITIGATION_RETPOLINE -> X86_UGLY_INDIRECT_THUNKS
>   MITIGATION_RETHUNK   -> X86_UGLY_RETURN_THUNKS
> 
>   etc...
> 
> Then one only needs to grep their .config file for UGLY to understand
> why their disassembly is so inscrutable ;-)

:-) Agree, this is ugly. One way to handle this could be two levels of
Kconfig options like below:

* CONFIG_MITIGATION_FOO			- Compiles out the mitigation
  * CONFIG_MITIGATION_FOO_DEFAULT	- Sets the bootup default
  select X86_UGLY_INDIRECT_THUNKS

Then supporting mitigation like INDIRECT_THUNKS would only be enabled if
there is a dependent mitigation enabled.