Re: Performance Regression in Linux Kernel 5.19

[Borislav Petkov]

A couple more notes after talking to tglx:

So this works as expected. The threat model where the guest needs
to protect itself from malicious userspace is there so if the guest
emulates a CPU which is affected by retbleed and the hypervisor exposes
SPEC_CTRL, then the guest *should* enable IBRS to flush the RSB.

It is a lot nastier if the guest emulates a CPU which is *not* affected
by retbleed but the host uarch is - then the guest will be vulnerable
and it would not even warn about it! So people should be careful what
they do there.

In addition, if the guest trusts its userspace, it might disable IBRS
in order not to suffer the penalty but that's left to the guest owner.
The default setting has to be secure.

HTH.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette