Re: [PATCH v2 3/4] x86, mwaitt: introduce mwaix delay with a configurable timer

[Borislav Petkov <bp@suse.de>]

On Fri, Jun 12, 2015 at 04:15:23PM -0700, Andy Lutomirski wrote:
> > How about this:
> >
> >         /*
> >          * This should be a memory location in a cache line which is
> >          * unlikely to be touched by other processors.  The actual
> >          * content is immaterial as it is not actually modified in any way.
> >          */
> >         mwait_ptr = &current_thread_info()->flags;
> >
> > and then
> >
> >         __monitor(mwait_ptr, 0, 0);
> >
> > We already do this in mwait_play_dead().
> >
> > However, am I even correct in assuming that ->flags won't really be
> > touched as we're doing delay() and nothing pokes into current anyway?
> 
> We poke flags remotely, but not frequently enough for this to be a
> problem.  However, I don't know that touching current in udelay is
> okay.
> 
> How about some read-mostly percpu variable, such as cpu_tss?

Yeah, those look much safer since they're static and are
____cacheline_aligned_in_smp, which is exactly what we want.

I guess we can do

	__monitorx(this_cpu_ptr(&cpu_tss), 0, 0);

with a nice comment ontop why we're doing it.

asm looks ok too:

	movq	$cpu_tss, %rax	#, tcp_ptr__
	add %gs:this_cpu_off(%rip), %rax	# this_cpu_off, tcp_ptr__
	xorl	%edx, %edx	# tmp248
	movq	%rdx, %rcx	# tmp248, tmp248
#APP
# 22 "./arch/x86/include/asm/mwait.h" 1
	.byte 0x0f, 0x01, 0xc8;
# 0 "" 2
#NO_APP

Thanks.

-- 
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--