Messages in this thread | | | Date | Fri, 30 Nov 2012 10:03:03 -0800 | From | "H. Peter Anvin" <> | Subject | Re: [PATCH 8/8] x86, cleanups: Simplify sync_core() in the case of no CPUID |
| |
On 11/30/2012 09:01 AM, Linus Torvalds wrote: > On Thu, Nov 29, 2012 at 1:24 PM, H. Peter Anvin <hpa@linux.intel.com> wrote: >> >> Thinking about it some more, there is another reason to not do this, >> which is that we don't want this particular CPUID to be paravirtualized; >> we're after the synchronizing side effect, not the CPUID return value >> itself. >> >> So let's leave it as a primitive; it gets too confusing otherwise. > > Hmm. The virtualization issue brings up another point: do we *really* > want to use cpuid for serialization at all? >
Well, the grand total of serializing instructions are:
INVD, INVEPT, INVLPG, INVVPID, LGDT, LIDT, LLDT, LTR, MOV to CR, MOV to DR, WBINVD, WRMSR, CPUID, IRET, RSM.
It doesn't really leave a lot of wiggle room, and in the microcode case, the use of CPUID level 1 is actually mandated (presumably to get a uniform sequence for validation purposes.)
> From all of the above, the alternatives case is kinda relevant for virt > where we do text_poke_early in a loop for every alternative section > so this could pile up to a bunch of vmexits depending on the emulated > hardware. Might be worth a replacement if it is noticeable in guests.
This is still boot time, and I really doubt it is measurable in the long run. Yes, exists suck, but at least CPUID is generally a quick exit, since all the relevant state is in registers.
-hpa
-- H. Peter Anvin, Intel Open Source Technology Center I work for Intel. I don't speak on their behalf.
| |