Re: [PATCH] 2.6.4-rc2-mm1: vm-split-active-lists

[Nikita Danilov <Nikita@Namesys.COM>]

Nick Piggin writes:
 > 
 > 
 > Nikita Danilov wrote:
 > 
 > >Nick Piggin writes:
 > > > 
 > >
 > >[...]
 > >
 > > > 
 > > > By the way, I would be interested to know the rationale behind
 > > > mark_page_accessed as it is without this patch, also what is it doing in
 > > > rmap.c (I know hardly anything actually uses page_test_and_clear_young, but
 > > > still). It seems to me like it only serves to make VM behaviour harder to
 > > > understand, but I'm probably missing something. Andrew?
 > >
 > >With your patch, once a page got into inactive list, its PG_referenced
 > >bit will only be checked by VM scanner when page wanders to the tail of
 > >list. In particular, if is impossible to tell pages that were accessed
 > >only once while on inactive list from ones that were accessed multiple
 > >times. Original mark_page_accessed() moves page to the active list on
 > >the second access, thus making it less eligible for the reclaim.
 > >
 > >
 > 
 > With my patch though, it gives unmapped pages the same treatment as
 > mapped pages. Without my patch, pages getting a lot of mark_page_accessed
 > activity can easily be promoted unfairly past mapped ones which are simply
 > getting activity through the pte.

Another way to put it is that treatment of file system pages is dumbed
down to the level of mapped ones: information about access patterns is
just discarded.

 > 
 > I say just set the bit and let the scanner handle it.

I think that decisions about balancing VM and file system caches should
be done by higher level, rather than by forcing file system to use
low-level mechanisms designed for VM, where only limited information is
provided by hardware. Splitting page queues is a step in a right
direction, as it allows to implement more precise replacement for the
file system cache.

Nikita.