Re: [PATCHv3 3/5] mm: pack compound_dtor and compound_order into one word in struct page

[Kirill A. Shutemov]

On Fri, Aug 21, 2015 at 09:13:42AM +0200, Michal Hocko wrote:
> On Thu 20-08-15 16:26:04, Andrew Morton wrote:
> > On Wed, 19 Aug 2015 12:21:44 +0300 "Kirill A. Shutemov" <kirill.shutemov@xxxxxxxxxxxxxxx> wrote:
> > 
> > > The patch halves space occupied by compound_dtor and compound_order in
> > > struct page.
> > > 
> > > For compound_order, it's trivial long -> int/short conversion.
> > > 
> > > For get_compound_page_dtor(), we now use hardcoded table for destructor
> > > lookup and store its index in the struct page instead of direct pointer
> > > to destructor. It shouldn't be a big trouble to maintain the table: we
> > > have only two destructor and NULL currently.
> > > 
> > > This patch free up one word in tail pages for reuse. This is preparation
> > > for the next patch.
> > > 
> > > ...
> > >
> > > @@ -145,8 +143,13 @@ struct page {
> > >  						 */
> > >  		/* First tail page of compound page */
> > >  		struct {
> > > -			compound_page_dtor *compound_dtor;
> > > -			unsigned long compound_order;
> > > +#ifdef CONFIG_64BIT
> > > +			unsigned int compound_dtor;
> > > +			unsigned int compound_order;
> > > +#else
> > > +			unsigned short int compound_dtor;
> > > +			unsigned short int compound_order;
> > > +#endif
> > 
> > Why not use ushort for 64-bit as well?
> 
> Yeah, I have asked the same in the previous round. So I've tried to
> compile with ushort. The resulting code was slightly larger
>    text    data     bss     dec     hex filename
>  476370   90811   44632  611813   955e5 mm/built-in.o.prev
>  476418   90811   44632  611861   95615 mm/built-in.o.after
> 
> E.g. prep_compound_page
> before:
> 4c6b:       c7 47 68 01 00 00 00    movl   $0x1,0x68(%rdi)
> 4c72:       89 77 6c                mov    %esi,0x6c(%rdi)
> after:
> 4c6c:       66 c7 47 68 01 00       movw   $0x1,0x68(%rdi)
> 4c72:       66 89 77 6a             mov    %si,0x6a(%rdi)
> 
> which looks very similar to me but I am not an expert here so it might
> possible that movw is slower.
> 
> __free_pages_ok
> before:
> 63af:       8b 77 6c                mov    0x6c(%rdi),%esi
> after:
> 63b1:       0f b7 77 6a             movzwl 0x6a(%rdi),%esi
> 
> which looks like a worse code to me. Whether this all is measurable or
> worth it I dunno. The ifdef is ugly but maybe the ugliness is a destiny
> for struct page.

I don't care about the ifdef that much. If you guys prefer to drop it I'm
fine with that.

-- 
 Kirill A. Shutemov
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/