Re: [PATCH] x86/retpoline: Fill return stack buffer on vmexit

["Woodhouse, David" <dwmw@amazon.co.uk>]

[-- Attachment #1: Type: text/plain, Size: 1195 bytes --]

On Wed, 2018-01-10 at 15:22 -0800, David Lang wrote:
> I somewhat hate to ask this, but for those of us following at home, what does 
> this add to the overhead?
> 
> I am remembering an estimate from mid last week that put retpoline at replacing 
> a 3 clock 'ret' with 30 clocks of eye-bleed code

Retpoline doesn't replace 'ret'.

It replaces indirect branches (jmp *%rax) of which there aren't quite
as many in the kernel.

The eye-bleed retpoline thunk does actually stop speculation and cause
a pipeline stall. For the RSB stuffing that's not the case; there are
no barriers here.

The actual performance numbers depend on the precise CPU being used,
and I'm not sure anyone has done the microbenchmarks of each *specific*
part for of the mitigations separately. For this *particular* patch...
well, we strive to avoid vmexits anyway, and Intel has spent the last
decade adding more and more tricks to the CPU to help us *avoid*
vmexits. So a little extra overhead on the vmexit is something we can
probably tolerate.

FWIW the IBRS microcode also requires the RSB-stuffing, so it's kind of
orthogonal to the "retpoline is much faster than IBRS" observation.


[-- Attachment #2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5210 bytes --]