Re: [PATCH] swapin flush cache bug

[NIIBE Yutaka <gniibe@m17n.org>]

Hello Stephen, 

Stephen C. Tweedie wrote:
 > First, don't we want to do a flush_page_to_ram() *before* starting the
 > swap IO?

Well, let me explain the issue.  It is the thing we need to do
flushing *after* I/O.

--------------------------
Problem with virtually indexed physically tagged write-back cache.

(1) Page got swapped out

           Swap out
   [ Page ] ----> [ Disk ]

(2) Page got swapped in asynchronously, possibly by read-ahead

           Swap in
   [ Page ] <---- [ Disk ]
	   K

   The I/O from disk goes through kernel virtual address K.
   We have cache entries indexed by K.

(3) Page fault occurs at user space U

   U ----> [ Page ] -----> [ Disk ]
		   K

   The control goes to do_swap_page, found the page at
   lookup_swap_cache.

   If K and U indexes differently, we have cache alias issues, we need
   to flush the entries indexed by K and let them go to memory.  Or else, 
   user space will see bogus data in Page.
--