[v3,6/7] mm: make clear_huge_page cache clear only around the fault address

From: Andi Kleen <ak@linux.intel.com>

Clearing a 2MB huge page will typically blow away several levels
of CPU caches. To avoid this only cache clear the 4K area
around the fault address and use a cache avoiding clears
for the rest of the 2MB area.

Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
---
 mm/memory.c |   34 +++++++++++++++++++++++++++++-----
 1 files changed, 29 insertions(+), 5 deletions(-)

Hi Kirill,

On Thu, Aug 16, 2012 at 06:15:53PM +0300, Kirill A. Shutemov wrote:
>  	for (i = 0; i < pages_per_huge_page;
>  	     i++, p = mem_map_next(p, page, i)) {

It may be more optimal to avoid a multiplication/shiftleft before the
add, and to do:

  	for (i = 0, vaddr = haddr; i < pages_per_huge_page;
  	     i++, p = mem_map_next(p, page, i), vaddr += PAGE_SIZE) {

>  		cond_resched();
> -		clear_user_highpage(p, addr + i * PAGE_SIZE);
> +		vaddr = haddr + i*PAGE_SIZE;

Not sure if gcc can optimize it away because of the external calls.

> +		if (!ARCH_HAS_USER_NOCACHE || i == target)
> +			clear_user_highpage(page + i, vaddr);
> +		else
> +			clear_user_highpage_nocache(page + i, vaddr);
>  	}


My only worry overall is if there can be some workload where this may
actually slow down userland if the CPU cache is very large and
userland would access most of the faulted in memory after the first
fault.

So I wouldn't mind to add one more check in addition of
!ARCH_HAS_USER_NOCACHE above to check a runtime sysctl variable. It'll
waste a cacheline yes but I doubt it's measurable compared to the time
it takes to do a >=2M hugepage copy.

Furthermore it would allow people to benchmark its effect without
having to rebuild the kernel themself.

All other patches looks fine to me.

Thanks!
Andrea
--
To unsubscribe from this list: send the line "unsubscribe sparclinux" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Thu, Aug 16, 2012 at 06:16:47PM +0200, Andrea Arcangeli wrote:
> Hi Kirill,
> 
> On Thu, Aug 16, 2012 at 06:15:53PM +0300, Kirill A. Shutemov wrote:
> >  	for (i = 0; i < pages_per_huge_page;
> >  	     i++, p = mem_map_next(p, page, i)) {
> 
> It may be more optimal to avoid a multiplication/shiftleft before the
> add, and to do:
> 
>   	for (i = 0, vaddr = haddr; i < pages_per_huge_page;
>   	     i++, p = mem_map_next(p, page, i), vaddr += PAGE_SIZE) {
> 

Makes sense. I'll update it.

> >  		cond_resched();
> > -		clear_user_highpage(p, addr + i * PAGE_SIZE);
> > +		vaddr = haddr + i*PAGE_SIZE;
> 
> Not sure if gcc can optimize it away because of the external calls.
> 
> > +		if (!ARCH_HAS_USER_NOCACHE || i == target)
> > +			clear_user_highpage(page + i, vaddr);
> > +		else
> > +			clear_user_highpage_nocache(page + i, vaddr);
> >  	}
> 
> 
> My only worry overall is if there can be some workload where this may
> actually slow down userland if the CPU cache is very large and
> userland would access most of the faulted in memory after the first
> fault.
> 
> So I wouldn't mind to add one more check in addition of
> !ARCH_HAS_USER_NOCACHE above to check a runtime sysctl variable. It'll
> waste a cacheline yes but I doubt it's measurable compared to the time
> it takes to do a >=2M hugepage copy.

Hm.. I think with static_key we can avoid cache overhead here. I'll try.
 
> Furthermore it would allow people to benchmark its effect without
> having to rebuild the kernel themself.
> 
> All other patches looks fine to me.

Thanks, for review. Could you take a look at huge zero page patchset? ;)

On Thu, Aug 16, 2012 at 07:43:56PM +0300, Kirill A. Shutemov wrote:
> Hm.. I think with static_key we can avoid cache overhead here. I'll try.

Could you elaborate on the static_key? Is it some sort of self
modifying code?

> Thanks, for review. Could you take a look at huge zero page patchset? ;)

I've noticed that too, nice :). I'm checking some detail on the
wrprotect fault behavior but I'll comment there.
--
To unsubscribe from this list: send the line "unsubscribe sparclinux" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Thu, Aug 16, 2012 at 08:29:44PM +0200, Andrea Arcangeli wrote:
> On Thu, Aug 16, 2012 at 07:43:56PM +0300, Kirill A. Shutemov wrote:
> > Hm.. I think with static_key we can avoid cache overhead here. I'll try.
> 
> Could you elaborate on the static_key? Is it some sort of self
> modifying code?

Runtime code patching. See Documentation/static-keys.txt. We can patch it
on sysctl.

> 
> > Thanks, for review. Could you take a look at huge zero page patchset? ;)
> 
> I've noticed that too, nice :). I'm checking some detail on the
> wrprotect fault behavior but I'll comment there.
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org.  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

On Thu, Aug 16, 2012 at 09:37:25PM +0300, Kirill A. Shutemov wrote:
> On Thu, Aug 16, 2012 at 08:29:44PM +0200, Andrea Arcangeli wrote:
> > On Thu, Aug 16, 2012 at 07:43:56PM +0300, Kirill A. Shutemov wrote:
> > > Hm.. I think with static_key we can avoid cache overhead here. I'll try.
> > 
> > Could you elaborate on the static_key? Is it some sort of self
> > modifying code?
> 
> Runtime code patching. See Documentation/static-keys.txt. We can patch it
> on sysctl.

I guessed it had to be patching the code, thanks for the
pointer. It looks a perfect fit for this one agreed.
--
To unsubscribe from this list: send the line "unsubscribe sparclinux" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html