Re: bug in migrate page

[Shaohua Li <shaohua.li@intel.com>]

On Wed, 2007-08-22 at 11:52 +0900, KAMEZAWA Hiroyuki wrote:
> On Wed, 22 Aug 2007 10:08:09 +0800
> Shaohua Li <shaohua.li@intel.com> wrote:
> 
> > commit dc386d4d1e98bb39fb967ee156cd456c802fc692 adds rcu_read_lock, but
> > some routines in the lock range might sleep (like lock_buffer,
> > aops->writepage), I saw a 'sleep in atomic' warning. It appears the
> > patch has several versions before. Doing rcu_read_lock in PageAnon
> > sounds break the case of PageAnon(page) && PageSwapCache(page),
> > as .writepage might be called. The dummy anon patch maybe is ok.
> > 
> 
> Thank you for catching.
> 
> Maybe you're correct.
> 
> BTW, in PageAnon(page) && PageSwapCache(page) case, I can't find when
> .writepage is called. Could you explain ?
> 
> In my understanding,
> 
> rcu_read_lock()
> 	-> try_to_unmap()
> 	-> move_to_new_page()
> 		-> migrate_page() // swap has .migratepage member.
> 			-> migrate_page_move_mapping().
> 			-> migrate_page_copy().
> 		-> remove_migration_ptes().
> 
> 
> At quick glance, above path has no writepage() ops.
> just replace swap's radix tree entry.
I missed swap has .migratepage and thought fallback_migrate_page is
used, then I thought doing rcu lock in PageAnon case is ok.

Thanks,
Shaohua