On Mon, Mar 12, 2007 at 04:32:55PM +0900, KAMEZAWA Hiroyuki wrote:
quoted text
> 
> > When I try to put_page(original) (because I don't want it for this
> > moment) I get to bad_page path and need to reboot...
> > 
> > What am I missing? Am I using migrate_pages correctly?
> > 
> 
> ==
> migrate_pages()					    | refcnt = 2,
> 	-> unmap_and_move()			    | refcnt = 2,
> 		-> try_to_unmap()		    | refcnt = 1
> 		-> move_to_lru()		    | refcnt = 1
> 			-> lru_cache_add()  	    | refcnt = 1
> 				-> page_cache_get() | refcnt = 2
> 				-> pagevec_add()    | refcnt = 2 
> 			-> page_cache_put()         | refcnt 1
> ==

Thanks for clarification. 

quoted text
> see pagevec codes and release_pages() in swap.c
> per-cpu pagevec caches freed pages for a while. 
> It will be pushed back to the free list when pagevec goes full.

I am not sure about that. When I looked inside lru_cache_add, it will
call __pagevec_lru_add if per-cpu pagevec goes empty and this function
will add all pages to the zone specific inactive list.

I want to prevent that because I need to use original page for other
purposes and so it can't be in inactive list. So I have created little
change to unmap_and_move function so that original page is added back tu
LRU only if required. migrate_pages uses variant with put_lru parameter
and I am using __migrate_pfn_page which uses put_lru=0. See the attached 
patch.
What do you think about that. Is this way correct?

(Please add me to CC, because I am not list member)

Best regards
-- 
Michal Hocko