Re: [PATCH 5/6] x86/mm/kaiser: Optimize RESTORE_CR3

[Peter Zijlstra]

On Wed, Nov 29, 2017 at 09:02:12PM +0100, Borislav Petkov wrote:
> On Wed, Nov 29, 2017 at 11:33:06AM +0100, Peter Zijlstra wrote:
> > +.macro RESTORE_CR3 scratch_reg:req save_reg:req
> >  	STATIC_JUMP_IF_FALSE .Lend_\@, kaiser_enabled_key, def=1
> > +
> > +	/* ASID bit 11 is for user */
> > +	bt	$11, \save_reg
> 
> <---- newline here.

Seems weird to me, the bt and jnc are a pair.

> > +	/*
> > +	 * KERNEL pages can always resume with NOFLUSH as we do
> > +	 * explicit flushes.
> > +	 */
> > +	jnc	.Lnoflush_\@
> > +
> > +	/*
> > +	 * Check if there's a pending flush for the user ASID we're
> > +	 * about to set.
> > +	 */
> > +	movq	\save_reg, \scratch_reg
> > +	andq	$(0x7FF), \scratch_reg
> > +	bt	\scratch_reg, PER_CPU_VAR(user_asid_flush_mask)
> > +	jnc	.Lnoflush_\@
> > +
> > +	btr	\scratch_reg, PER_CPU_VAR(user_asid_flush_mask)
> > +	jmp	.Ldo_\@
> 
> Can you save yourself one of the BT-insns?
> 
> 	...
> 	andq	$(0x7FF), \scratch_reg
> 	btr     \scratch_reg, PER_CPU_VAR(user_asid_flush_mask)
> 	jnc	.Lnoflush_\@
> 	jmp     .Ldo_\@
> 	...
> 
> or am I missing a case?

BTR is an unconditional write and will modify the line and cause a
write-back later. The common case is the bit not set, so BT, which is a
pure read, avoids all that overhead.