lkml.org 
[lkml]   [2009]   [Feb]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v2] fix mlocked page counter mistmatch
On Wed, Feb 04, 2009 at 07:28:19PM +0900, KOSAKI Motohiro wrote:
> > With '29-rc3-git5', I found,
> >
> > static int try_to_mlock_page(struct page *page, struct vm_area_struct *vma)
> > {
> > int mlocked = 0;
> >
> > if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
> > if (vma->vm_flags & VM_LOCKED) {
> > mlock_vma_page(page);
> > mlocked++; /* really mlocked the page */
> > }
> > up_read(&vma->vm_mm->mmap_sem);
> > }
> > return mlocked;
> > }
> >
> > It still try to downgrade mmap_sem.
> > Do I miss something ?
>
> sorry, I misunderstood your "downgrade". I said linus removed downgrade_write(&mma_sem).
>
> Now, I understand this issue perfectly. I agree you and lee-san's fix is correct.
> Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>

Good. I will send adrew with your ACK agian.

>
>
> and, I think current try_to_mlock_page() is correct. no need change.
> Why?
>
> 1. Generally, mmap_sem holding is necessary when vma->vm_flags accessed.
> that's vma's basic rule.
> 2. However, try_to_unmap_one() doesn't held mamp_sem. but that's ok.
> it often get incorrect result. but caller consider incorrect value safe.
> 3. try_to_mlock_page() need mmap_sem because it obey rule (1).
> 4. in try_to_mlock_page(), if down_read_trylock() is failure,
> we can't move the page to unevictable list. but that's ok.
> the page in evictable list is periodically try to reclaim. and
> be called try_to_unmap().
> try_to_unmap() (and its caller) also move the unevictable page to unevictable list.
> Therefore, in long term view, the page leak is not happend.

Thanks for clarification.
In long term view, you're right.

but My concern is that munlock[all] pathes always hold down of mmap_sem.
After all, down_read_trylock always wil fail for such cases.

So, current task's mlocked pages only can be reclaimed
by background or direct reclaim path if the task don't exit.

I think it can increase reclaim overhead unnecessary
if there are lots of such tasks.

What's your opinion ?

>
> this explanation is enough?
>
> thanks.
>

--
Kinds Regards
MinChan Kim



\
 
 \ /
  Last update: 2009-02-05 00:39    [from the cache]
©2003-2011 Jasper Spaans