Re: [RFC][PATCH] memcg remove css_get/put per pages

[KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

On Wed, 9 Jun 2010 09:47:34 +0900
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:

> > Looks nice, Kamezawa-San could you please confirm the source of
> > raw_spin_lock_irqsave and trylock from /proc/lock_stat?
> >  
> Sure. But above result can be got when lockdep etc..are off.
> (it increase lock overhead)
> 
> But yes, new _raw_spin_lock seems strange.
> 
Here.
==
         ------------------------------
          &(&mm->page_table_lock)->rlock       20812995          [<ffffffff81124019>] handle_mm_fault+0x7a9/0x9b0
          &(&mm->page_table_lock)->rlock              9          [<ffffffff81120c5b>] __pte_alloc+0x4b/0xf0
          &(&mm->page_table_lock)->rlock              4          [<ffffffff8112c70d>] anon_vma_prepare+0xad/0x180
          &(&mm->page_table_lock)->rlock          83395          [<ffffffff811204b4>] unmap_vmas+0x3c4/0xa60
          ------------------------------
          &(&mm->page_table_lock)->rlock              7          [<ffffffff81120c5b>] __pte_alloc+0x4b/0xf0
          &(&mm->page_table_lock)->rlock       20812987          [<ffffffff81124019>] handle_mm_fault+0x7a9/0x9b0
          &(&mm->page_table_lock)->rlock              2          [<ffffffff8112c70d>] anon_vma_prepare+0xad/0x180
          &(&mm->page_table_lock)->rlock          83408          [<ffffffff811204b4>] unmap_vmas+0x3c4/0xa60

                &(&p->alloc_lock)->rlock:       6304532        6308276           0.14        1772.97     7098177.74       23165904       23222238           0.00        1980.76    12445023.62
                ------------------------
                &(&p->alloc_lock)->rlock        6308277          [<ffffffff81153e17>] __mem_cgroup_try_charge+0x327/0x590
                ------------------------
                &(&p->alloc_lock)->rlock        6308277          [<ffffffff81153e17>] __mem_cgroup_try_charge+0x327/0x590



==

Then, new raw_spin_lock is task_lock(). This is because task_lock(mm->owner) makes
cacheline ping pong ;(

So, this is not very good patch for multi-threaded programs, Sigh...
I'll consider how I can get safe access without locks again..

Thanks,
-Kame

--
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>