Re: [RFC] memory unplug v5 [1/6] migration by kernel

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

On Thu, 14 Jun 2007 07:19:19 -0700 (PDT)
Christoph Lameter <clameter@sgi.com> wrote:

> On Thu, 14 Jun 2007, KAMEZAWA Hiroyuki wrote:
> 
> > But...during discussion with you, I found anon_vma is now freed by RCU...
> > 
> > Ugh, then, what I have to do is  rcu_read_lock() -> rcu_read_unlock() while
> > migrating anon ptes. If we can rcu read lock here, we don't need dummy_vma.
> > How about this ?
> 
> Hmmmm... Looks good. Maybe take the RCU lock unconditionally? Is there a 
> problem if we do so? Then the patch becomes very small and it looks 
> cleaner. 
Ok, maybe no problem if "if" is removed.

> 
> Is there an issue with calling try_to_unmap for an unmapped page? We check 
> in try_to_unmap if the pte is valid. If it was unmapped then try_to_unmap 
> will fail anyways.
> 
I met following case.
---
   CPU 0                                          CPU 1

do_swap_page()                                                                        
  -> read_swap_cache_async()               
	-> # alloc new page 
	   # page is added to swapcache
           # page is locked here.
           # added to LRU                      <- we find this page because of PG_lru
           # start asynchrous read I/O         lock_page()
	   # page is unlocked here             we acquire the lock.
  -> lock_page()                                     
     wait....                                  unmap_and_move() is called.
                                               try_to_unmap() is called.
                                               PageAnon() returns 0. beacause the page is not
                                               added to rmap yet. page->mapping is NULL, here.
                                               try_to_unmap_file() is called.
                                               try_to_unmap_file() touches NULL pointer.
--
An unmapped swapcache page, which is just added to LRU, may be accessed via migrate_page().
But page->mapping is NULL yet. 

Hmm, should I add following check instead of page_mapped() ?
--
if (likely(page->mapping))
	try_to_unmap(page,1)
--
Note: Because the page's page_count() is not migratable,(do_swap_page() has one ref.)
      this migration will fail with -EAGAIN.

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>