Re: [-mm][PATCH 2/4] Setup the memrlimit controller (v5)

[Andrea Righi <righi.andrea@gmail.com>]

Andrew Morton wrote:
>> At least we could add something like:
>>
>> #ifdef CONFIG_32BIT
>> #define PAGE_ALIGN64(addr) (((((addr)+PAGE_SIZE-1))>>PAGE_SHIFT)<<PAGE_SHIFT)
>> #else
>> #define PAGE_ALIGN64(addr) PAGE_ALIGN(addr)
>> #endif
>>
>> But IMHO the single PAGE_ALIGN64() implementation is more clear.
> 
> No, we should just fix PAGE_ALIGN.  It should work correctly when
> passed a long-long.  Otherwse it's just a timebomb.
> 
> This:
> 
> #define PAGE_ALIGN(addr) ({				\
> 	typeof(addr) __size = PAGE_SIZE;		\
> 	typeof(addr) __mask = PAGE_MASK;		\
> 	(addr + __size - 1) & __mask;			\
> })
> 
> (with a suitable comment) does what we want.  I didn't check to see
> whether this causes the compiler to generate larger code, but it
> shouldn't.
> 

No, it doesn't work. The problem seems to be in the PAGE_MASK definition
(from include/asm-x86/page.h for example):

/* PAGE_SHIFT determines the page size */
#define PAGE_SHIFT      12
#define PAGE_SIZE       (_AC(1,UL) << PAGE_SHIFT)
#define PAGE_MASK       (~(PAGE_SIZE-1))

The "~" is performed on a 32-bit value, so everything in "and" with
PAGE_MASK greater than 4GB will be truncated to the 32-bit boundary.

What do you think about the following?

#define PAGE_SIZE64 (1ULL << PAGE_SHIFT)
#define PAGE_MASK64 (~(PAGE_SIZE64 - 1))

#define PAGE_ALIGN(addr) ({					\
	typeof(addr) __size = PAGE_SIZE;			\
	typeof(addr) __ret = (addr) + __size - 1;		\
	__ret > -1UL ? __ret & PAGE_MASK64 : __ret & PAGE_MASK;	\
})

-Andrea