Re: [PATCH] Clear reference bit although page isn't mapped.

[Rik van Riel <riel@redhat.com>]

On 12/01/2009 09:55 PM, KOSAKI Motohiro wrote:
>> btw, current shrink_active_list() have unnecessary page_mapping_inuse() call.
>> it prevent to drop page reference bit from unmapped cache page. it mean
>> we protect unmapped cache page than mapped page. it is strange.
>>      
> How about this?
>
> ---------------------------------
> SplitLRU VM replacement algorithm assume shrink_active_list() clear
> the page's reference bit. but unnecessary page_mapping_inuse() test
> prevent it.
>
> This patch remove it.
>    
Shrink_page_list ignores the referenced bit on pages
that are !page_mapping_inuse().

                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
                                         referenced && 
page_mapping_inuse(page)
&& !(vm_flags & VM_LOCKED))
                         goto activate_locked;

The reason we leave the referenced bit on unmapped
pages is that we want the next reference to a deactivated
page cache page to move that page back to the active
list.  We do not want to require that such a page gets
accessed twice before being reactivated while on the
inactive list, because (1) we know it was a frequently
accessed page already and (2) ongoing streaming IO
might evict it from the inactive list before it gets accessed
twice.

Arguably, we should just replace the page_mapping_inuse()
in both places with page_mapped() to simplify things.