Re: [BUG 2.6.31-rc1] HIGHMEM64G causes hang in PCI init on 32-bit x86

[Yinghai Lu <yinghai@kernel.org>]

Linus Torvalds wrote:
> 
> On Tue, 30 Jun 2009, Yinghai Lu wrote:
> ?
>> +#define round_up(x, y) ({ __typeof__(x) __mask = (y)-1; \
>> +			 ((x)+__mask) & ~__mask; })
>> +#define round_down(x, y) ({ __typeof__(x) __mask = (y)-1; (x) & ~__mask; })
> 
> Yes, except we might as well simplify it. Do it without the statement 
> expressions, using just a single 'y'. Like this:
> 
>   #define __round_mask(x,y) ((__typeof__(x))((y)-1))
>   #define round_up(x,y) (((x) | __round_mask(x,y))+1)
>   #define round_down(x,y) ((x) & ~__round_mask(x,y))
> 
> (Yeah, it uses 'x' twice, but the second one is for 'typeof', which 
> doesn't actually cause the value to be evaluated, so it's ok).
> 
> Now those 'round_xyz()' operations will always return a value of a type 
> that is the same as the type of 'x' except it's gone through the normal C 
> integer promotion rules (ie if 'x' is a smaller type than 'int', then it 
> will be promoted to 'int').
> 
> Not very well tested, but it _looks_ correct, and uses Peter's trick, and 
> willlet the compiler notice that
> 
> 	round_up(x,y)-1
> 
> is the same thing as
> 
> 	x | (y-1)
> 
> which it _could_ have perhaps figured out before, but now it's way more 
> obvious.

how about x = 0, y = 0x100?

YH