On 04/21/2018 06:35 PM, speck for Borislav Petkov wrote:
> On Sat, Apr 21, 2018 at 06:13:16PM -0400, speck for Jon Masters wrote:
>> I propose keeping either the x86 CPU capability "Memory Disambiguation"
>> (as you said, via the string in the define). But only for what is
>> displayed in /proc/cpuinfo.
> 
> Actually, thinking about this more, we don't need to have any strings
> in /proc/cpuinfo denoting the CPU has the memory disambiguation feature
> because we only want to know whether we can *disable* the MD. And that
> we can do only on a subset of CPUs. IOW, there are other CPU models
> which *might* have MD but where MD cannot be disabled... yet.
> 
> So having MD or whatever string we agree upon in /proc/cpuinfo will be a
> lie.
> 
> IOW, we only need two flags:
> 
> X86_FEATURE_STBUF_BYPASS
> X86_FEATURE_STBUF_BYPASS_MITIGATE
> 
> or whatever those are going to be called and neither of the two should
> be visible in /proc/cpuinfo.
> 
> SSB state will be visible in /sys/devices/system/cpu/vulnerabilities/
> 
> Yap, I think that should work.

All sounds reasonable. You could display "MDD" if you like in
/proc/cpuinfo later but since that's ABI (on some level) not adding it
for now is safe for the reasons given above. Hopefully Linus can let us
know if he likes X86_FEATURE_SPEC_STORE_BYPASS vs the "STBUF" version.

I'm going over Konrad's patches at the moment trying to integrate some
of the suggestions from the past 24 hours and will send him what I come
up with tonight in case that's helpful to him. I've got syncs setup with
IBM p and z and Arm early next week so can convey whatever is agreed.

Jon.

-- 
Computer Architect | Sent from my Fedora powered laptop