Re: [PATCH] x86/entry/64: Disallow RDPID in paranoid entry if KVM is enabled

[Paolo Bonzini]

On 21/08/20 10:16, Borislav Petkov wrote:
> On Fri, Aug 21, 2020 at 10:09:01AM +0200, Paolo Bonzini wrote:
>> One more MSR *is* a big deal: KVM's vmentry+vmexit cost is around 1000
>> cycles, adding 100 clock cycles for 2 WRMSRs is a 10% increase.
> 
> The kernel uses TSC_AUX so we can't reserve it to KVM either.

KVM only uses TSC_AUX while in kernel space, because the kernel hadn't
used it until now.  That's for a good reason:

* if possible, __this_cpu_read(cpu_number) is always faster.

* The kernel can just block preemption at its will and has no need for
the atomic rdtsc+vgetcpu provided by RDTSCP.

So far, the kernel had always used LSL instead of RDPID when
__this_cpu_read was not available.  In one place, RDTSCP is used as an
ordered rdtsc but it discards the TSC_AUX value.  RDPID is also used in
the vDSO but it isn't kernel space.

Hence the assumption that KVM makes (and has made ever since TSC_AUX was
introduced.  What is the difference in speed between LSL and RDPID?  I
don't have a machine that has RDPID to test it, unfortunately.

Paolo