Re: [PATCH 9/12] memcg allocate all page_cgroup at boot

[KAMEZAWA Hiroyuki]

On Fri, 26 Sep 2008 14:54:22 +0900
Daisuke Nishimura <nishimura@xxxxxxxxxxxxxxxxx> wrote:
> > Sorry, my brain seems to be sleeping.. above page_mapped() check doesn't
> > help this situation. Maybe this page_mapped() check is not necessary
> > because it's of no use.
> > 
> > I think this kind of problem will not be fixed until we handle SwapCache.
> > 
> I've not fully understood yet what [12/12] does, but if we handle
> swapcache properly, [12/12] would become unnecessary?
> 

Try to illustrate what is trouble more precisely.


in do_swap_page(), page is charged when SwapCache lookup ends.

Here, 
     - charged when page is not mapped.
     - not charged when page is mapped.
set_pte() etc...are done under appropriate lock.

On the other side, when a task exits, zap_pte_range() is called.
It calls page_remove_rmap(). 

Case A) Following is race.

            Thread A                     Thread B

     do_swap_page()                      zap_pte_range()
	(1)try charge (mapcount=1)
                                         (2) page_remove_rmap()
					     (3) uncharge page. 
	(4) map it


Then,
 at (1),  mapcount=1 and this page is not charged.
 at (2),  page_remove_rmap() is called and mapcount goes down to Zero.
          uncharge(3) is called.
 at (4),  at the end of do_swap_page(), page->mapcount=1 but not charged.

Case B) In another scenario.

            Thread A                     Thread B

     do_swap_page()                      zap_pte_range()
	(1)try charge (mapcount=1)
                                         (2) page_remove_rmap()
	(3) map it
                                         (4) uncharge is called.

In (4), uncharge is capped but mapcount can go up to 1.

protocol 12/12 is for case (A).
After 12/12, double-check page_mapped() under lock_page_cgroup() will be fix to
case (B).

Huu, I don't like swap-cache ;)
Anyway, we'll have to handle swap cache later.

Thanks,
-Kame






















--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/