Re: [PATCH] MM: use DIV_ROUND_UP() in mm/memory.c

[Rolf Eike Beer <eike-kernel@sf-tec.de>]

Andrew Morton wrote:
> Rolf Eike Beer wrote:

>> --- a/mm/memory.c
>> +++ b/mm/memory.c
>> @@ -1838,12 +1838,11 @@ void unmap_mapping_range(struct address_space
>> *mapping,
>>  {
>>  	struct zap_details details;
>>  	pgoff_t hba = holebegin >> PAGE_SHIFT;
>> -	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
>> +	pgoff_t hlen = DIV_ROUND_UP(holelen, PAGE_SIZE);
>>
>>  	/* Check for overflow. */
>>  	if (sizeof(holelen) > sizeof(hlen)) {
>> -		long long holeend =
>> -			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
>> +		long long holeend = DIV_ROUND_UP(holebegin + holelen, PAGE_SIZE);
>>  		if (holeend & ~(long long)ULONG_MAX)
>>  			hlen = ULONG_MAX - hba + 1;
>>  	}
>> @@ -2592,7 +2591,7 @@ int make_pages_present(unsigned long addr, unsigned
>> long end)
>>  	write = (vma->vm_flags & VM_WRITE) != 0;
>>  	BUG_ON(addr >= end);
>>  	BUG_ON(end > vma->vm_end);
>> -	len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
>> +	len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
>>  	ret = get_user_pages(current, current->mm, addr,
>>  			len, write, 0, NULL, NULL);
>>  	if (ret < 0)
>
> More seriously, on i386:
>
>    text    data     bss     dec     hex filename
>   15509      27      28   15564    3ccc mm/memory.o	(before)
>   15561      27      28   15616    3d00 mm/memory.o	(after)
>
> I'm not sure why - some of the quantities which we're dividing by there are
> 64-bit and perhaps the compiler has decided not to do shifting.
>
> Now I'm worried about all the other DIV_ROUND_UP() conversions we did.  We
> should get in there and work out why it went bad.

It's the first two places that cause the increase in code size. The last one 
is the exact replacement of how DIV_ROUND_UP() is defined so that can hardly 
make a difference.

If the compiler can't find out to do shifting we might want to improve 
DIV_ROUND_UP() to do some tricks with __builtin_constant_p() to do the shift. 
Something like:

#define DIV_ROUND_UP(n, d)			\
(						\
	__builtin_constant_p(d) ? (		\
		is_power_of_2(d) ?		\
		(((n) + (d) - 1) >> ilog2(d)) :	\
		(((n) + (d) - 1) / (d))		\
	) :					\
	(((n) + (d) - 1) / (d))			\
)

With this version mm/memory.o will result in the same size with and without my 
patch so I guess it's doing what it should. If you like it tell me and I'll 
send a patch.

Eike

Attachment: signature.asc
Description: This is a digitally signed message part.