Re: [PATCH 3/3] mm/memcg: apply add/del_page to lruvec

[Chen Baozi <baozich@gmail.com>]

On Thu, May 31, 2012 at 03:58:59PM -0700, Hugh Dickins wrote:
> On Thu, 31 May 2012, baozich wrote:
> > > 
> > > In their place, mem_cgroup_page_lruvec() to decide the lruvec,
> > > previously a side-effect of add, and mem_cgroup_update_lru_size()
> > > to maintain the lru_size stats.
> > I have a stupid question. I'm not sure whether there is reduplication
> > to put both "page" and "zone" parameter in mem_cgroup_page_lruvec(),
> > for I noticed that the "struct zone *zone" parameter are usually from 
> > page_zone(page) in most cases. I think that the semantics of this function
> > is to grab the lruvec the page belongs to. So will it be ok if we pass
> > only "page" as the parameter, which I think would be cleaner? Please
> > fix me if I missed something.
> 
> I share your dislike for passing down an "unnecessary" argument,
> but I do think it's justified here.
> 
> If the zone pointer were available simply by page->zone, then yes,
> I'd agree with you that it's probably silly to pass zone separately.
> 
> But page_zone(page) is never as trivial as that, and on some memory
> layouts it can be a lookup that you'd really prefer to avoid repeating.
> 
> In every(?) case where we're using mem_cgroup_page_lruvec(), the zone
> is already known: it's just been used for spin_lock_irq(&zone->lru_lock).
> 
> And when CONFIG_CGROUP_MEM_RES_CTLR is not set, the inline function
> mem_cgroup_page_lruvec() uses only zone, not page at all: I wouldn't
> want to be slowing down that case with another page_zone(page) lookup.
> 
> Also it's somewhat academic (though not for v3.5), in that this function
> goes away in the patches I build upon it; and I expect it also to go away
> in the patches Konstantin would build upon it - mem_cgroup_page_lruvec()
> is a staging point, before we combine memcg/zone lookup with the locking.
> 
> Hugh
I see. Thanks for the detailed explaination.

Baozi