Re: [RFC][PATCH 2/3] mm: slab: move around slab ->freelist for cmpxchg

[Dave Hansen <dave@sr71.net>]

On 12/12/2013 09:46 AM, Christoph Lameter wrote:
> On Wed, 11 Dec 2013, Dave Hansen wrote:
>> The write-argument to cmpxchg_double() must be 16-byte aligned.
>> We used to align 'struct page' itself in order to guarantee this,
>> but that wastes 8-bytes per page.  Instead, we take 8-bytes
>> internal to the page before page->counters and move freelist
>> between there and the existing 8-bytes after counters.  That way,
>> no matter how 'stuct page' itself is aligned, we can ensure that
>> we have a 16-byte area with which to to this cmpxchg.
> 
> Well this adds additional branching to the fast paths.

I don't think it *HAS* to inherently.  The reason here is really that we
swap the _order_ of the arguments to the cmpxchg() since their order in
memory changes.  Essentially, we do:

| flags | freelist  | counters |          |
| flags |           | counters | freelist |

I did this so I wouldn't have to make a helper for ->counters.  But, if
we also move counters around, we can do:

| flags | counters | freelist |          |
| flags |          | counters | freelist |

I believe we can do that all with plain pointer arithmetic and masks so
that it won't cost any branches.

--
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>