On 5/26/20 4:19 AM, Borislav Petkov wrote:

On Tue, May 19, 2020 at 10:16:37PM -0700, Sean Christopherson wrote:

quoted

The whole cache on-demand approach seems like overkill.  The number of CPUID
leaves that are invoked after boot with any regularity can probably be counted
on one hand.   IIRC glibc invokes CPUID to gather TLB/cache info, XCR0-based
features, and one or two other leafs.  A statically sized global array that's
arbitrarily index a la x86_capability would be just as simple and more
performant.  It would also allow fancier things like emulating CPUID 0xD in
the guest if you want to go down that road.

And before we do any of that "caching" or whatnot, I'd like to see
numbers justifying its existence. Because if it is only a couple of
CPUID invocations and the boot delay is immeasurable, then it's not
worth the effort.

I added some rudimentary stats code to see how many times there was a 
CPUID cache hit on a 64-vCPU guest during a kernel build (make clean 
followed by make with -j 64):

   SEV-ES CPUID cache statistics
     0x00000000/0x00000000: 220,384
     0x00000007/0x00000000: 213,306
     0x80000000/0x00000000: 1,054,642
     0x80000001/0x00000000: 213,306
     0x80000005/0x00000000: 210,334
     0x80000006/0x00000000: 420,668
     0x80000007/0x00000000: 210,334
     0x80000008/0x00000000: 420,684

     2,963,658 cache hits

So it is significant in quantity, but I'm not sure what the overall 
performance difference is. If I can find some more time I'll try to 
compare the kernel builds with and without the caching to see if it is 
noticeable.

Thanks,
Tom