Re: [PATCH v3] x86/cpu: Use SERIALIZE in sync_core() when available

[Ricardo Neri]

On Thu, Aug 06, 2020 at 12:57:26PM -0700, Dave Hansen wrote:
> On 8/6/20 12:25 PM, Ricardo Neri wrote:
> >  static inline void sync_core(void)
> >  {
> >  	/*
> > -	 * There are quite a few ways to do this.  IRET-to-self is nice
> > +	 * Hardware can do this for us if SERIALIZE is available. Otherwise,
> > +	 * there are quite a few ways to do this.  IRET-to-self is nice
> 
> This seems to imply that things other than SERIALIZE aren't the hardware
> doing this.  All of these methods are equally architecturally
> serializing *and* provided by the hardware.

Indeed, I can see how the wording may imply that.

> 
> I also don't quite get the point of separating the comments from the
> code.  Shouldn't this be:
> 
> 	/*
> 	 * The SERIALIZE instruction is the most straightforward way to
> 	 * do this but it not universally available.
> 	 */

I regard the comment as describing all the considered options to for
architectural serialization. What about if I add SERIALIZE as another
option? I propose to discuss it towards the end of the comment:

	/*
	 * There are quite a few ways to do this.  IRET-to-self is nice
	 * because it works on every CPU, at any CPL (so it's compatible
	 * with paravirtualization), and it never exits to a hypervisor.
	 * The only down sides are that it's a bit slow (it seems to be
	 * a bit more than 2x slower than the fastest options) and that
	 * it unmasks NMIs.  The "push %cs" is needed because, in
	 * paravirtual environments, __KERNEL_CS may not be a valid CS
	 * value when we do IRET directly.
	 *
	 * In case NMI unmasking or performance ever becomes a problem,
	 * the next best option appears to be MOV-to-CR2 and an
	 * unconditional jump.  That sequence also works on all CPUs,
	 * but it will fault at CPL3 (i.e. Xen PV).
	 *
	 * CPUID is the conventional way, but it's nasty: it doesn't
	 * exist on some 486-like CPUs, and it usually exits to a
	 * hypervisor.
	 *
 	 * The SERIALIZE instruction is the most straightforward way to
 	 * do this as it does not clobber registers or exit to a
	 * hypervisor. However, it is not universally available.
 	 *
	 * Like all of Linux's memory ordering operations, this is a
	 * compiler barrier as well.
	 */

What do you think?

> 	if (static_cpu_has(X86_FEATURE_SERIALIZE)) {
> 		asm volatile(__ASM_SERIALIZE ::: "memory");
> 		return;
> 	}
> 
> 	/*
> 	 * For all other processors, IRET-to-self is nice ...
> 	 */
> 	iret_to_self();