Re: locking rules for ->dirty_inode()

[Nikita Danilov <Nikita@Namesys.COM>]

Andrew Morton writes:
 > Nikita Danilov wrote:
 > > 
 > > Hello,
 > > 
 > > Documentation/filesystems/Locking states that all super operations may
 > > block, but __set_page_dirty_buffers() calls
 > > 
 > >    __mark_inode_dirty()->s_op->dirty_inode()
 > > 
 > > under mapping->private_lock spin lock. This seems strange, because file
 > > systems' ->dirty_inode() assume that they are allowed to block. For
 > > example, ext3_dirty_inode() allocates memory in
 > > 
 > >    ext3_journal_start()->journal_start()->new_handle()->...
 > > 
 > 
 > OK, thanks.
 > 
 > mapping->private_lock is taken there to pin page->buffers()
 > (Can't lock the page because set_page_dirty is called under
 > page_table_lock, and other locks).
 > 
 > I'm sure we can just move the spin_unlock up to above the
 > TestSetPageDirty(), but I need to zenuflect for a while over
 > why I did it that way.
 > 
 > It's necessary to expose buffer-dirtiness and page-dirtiness
 > to the rest of the world in the correct order.  If we set the
 > page dirty and then the buffers, there is a window in which writeback
 > could find the dirty page, try to write it, discover clean buffers
 > and mark the page clean.  We would end up with a !PageDirty page,
 > on mapping->clean_pages, with dirty buffers.  It would never be
 > written.
 > 
 > Yup.  We can move that spin_unlock up ten lines.

Actually, I came over this while trying to describe lock ordering in
reiser4 after I just started integrating other kernel locks there. I
wonder, has somebody already done this, writing up kernel lock
hierarchy, that is?

Nikita.