Re: [PATCH v2 1/3] x86: record SSP at non-guest entry points

[Jan Beulich <jbeulich@suse.com>]

On 09.04.2026 13:22, Andrew Cooper wrote:
> On 09/04/2026 9:13 am, Jan Beulich wrote:
>> On 08.04.2026 18:58, Andrew Cooper wrote:
>>>   It would be an rdssp;push on
>>> one side, and a pop into any register on the other side.  Furthermore,
>>> given that the ssp= doesn't exclude storing it for some user frames,
>>> just store it for all.  It's one push/pop into a hot cacheline, and
>>> makes a substantial reduction in complexity.
>> I'm having significant reservations against that. I use the 0 put there
>> in subsequent patches, to identify absence of that data being available.
> 
> Well, that's not safe then.
> 
> You've already got non-zero values there on entry-from-PV because
> there's no CPL check gating RDSPP the common IDT paths.

In that case we get safe values (as read from the MSR during the privilege
level switch). And wait - no, the latter two patches don't make any such
assumptions, I don't think. I may have mis-remembered, or I may have
remembered how things were at a very early stage. So really the concern is
with RDSSPQ's resource use. This could in principle be as cheap as MOV,
but how do I know? (The latency/throughput data I can find doesn't include
this insn.) Plus for the purely PV entrypoints I don't see why we would
want/need the slightly larger code size that would result.

Jan