Re: [tip:x86/mm] x86/mm: Avoid superflous leave_mm() in the TLB flush path

Re: [tip:x86/mm] x86/mm: Avoid superflous leave_mm() in the TLB flush path

[Linus Torvalds]

On Mon, Dec 5, 2011 at 9:56 AM, tip-bot for Shaohua Li
<shaohua.li@intel.com> wrote:
>
> x86/mm: Avoid superflous leave_mm() in the TLB flush path
>
> If just one page VA tlb is required to be flushed and current
> task is in lazy TLB state, doing leave_mm() is superfluous
> because it flushes the whole TLB.  This can reduce some TLB
> miss.

Are you sure this is a good idea?

We may be *much* better off leaving the VM and avoiding future IPI's
than trying to flush pages one by one over and over again.

IPI's are expensive.

I don't think you thought this through, and if you *did* think it
through and actually have numbers that it's cheaper to stay with the
MM, I think you need to show those numbers.

                     Linus

Re: [tip:x86/mm] x86/mm: Avoid superflous leave_mm() in the TLB flush path

[Shaohua Li]

On Tue, 2011-12-06 at 07:26 +0800, Linus Torvalds wrote:
> On Mon, Dec 5, 2011 at 9:56 AM, tip-bot for Shaohua Li
> <shaohua.li@intel.com> wrote:
> >
> > x86/mm: Avoid superflous leave_mm() in the TLB flush path
> >
> > If just one page VA tlb is required to be flushed and current
> > task is in lazy TLB state, doing leave_mm() is superfluous
> > because it flushes the whole TLB.  This can reduce some TLB
> > miss.
> 
> Are you sure this is a good idea?
> 
> We may be *much* better off leaving the VM and avoiding future IPI's
> than trying to flush pages one by one over and over again.
> 
> IPI's are expensive.
> 
> I don't think you thought this through, and if you *did* think it
> through and actually have numbers that it's cheaper to stay with the
> MM, I think you need to show those numbers.
good point, I didn't think of it. It's unlikely we just flush one tlb
page. please drop the patch.

Thanks,
Shaohua

Re: [tip:x86/mm] x86/mm: Avoid superflous leave_mm() in the TLB flush path

[Ingo Molnar]

* Shaohua Li <shaohua.li@intel.com> wrote:

> On Tue, 2011-12-06 at 07:26 +0800, Linus Torvalds wrote:
> > On Mon, Dec 5, 2011 at 9:56 AM, tip-bot for Shaohua Li
> > <shaohua.li@intel.com> wrote:
> > >
> > > x86/mm: Avoid superflous leave_mm() in the TLB flush path
> > >
> > > If just one page VA tlb is required to be flushed and current
> > > task is in lazy TLB state, doing leave_mm() is superfluous
> > > because it flushes the whole TLB.  This can reduce some TLB
> > > miss.
> > 
> > Are you sure this is a good idea?
> > 
> > We may be *much* better off leaving the VM and avoiding future IPI's
> > than trying to flush pages one by one over and over again.
> > 
> > IPI's are expensive.
> > 
> > I don't think you thought this through, and if you *did* think it
> > through and actually have numbers that it's cheaper to stay with the
> > MM, I think you need to show those numbers.
>
> good point, I didn't think of it. It's unlikely we just flush 
> one tlb page. please drop the patch.

Dropped, thanks guys.

	Ingo