Re: 2.6.9-rc2-mm4 ps hang ?

[Andrew Morton]

Hugh Dickins <hugh@xxxxxxxxxxx> wrote:
>
> lock_page inside mmap_sem a ranking bug?  Please recant!

generic_file_buffered_write() can take mmap_sem for reading while holding a
page lock.  In that rare case where the page gets unmapped even though we
manually faulted it in.

Now, that's lock_page->down_read versus down_read->lock_page which I
_think_ is safe, due to down_read semantics.  Even if a third thread is
waiting for a down_write.

Except filemap_nopage() does lock_page too, so we have

	lock_page->down_read->lock_page

as well.

All this does mean that down_write cannot nest either inside or outside
lock_page.

The bigger problem is ext3 and reiser3 transaction start/stop.  It is
equivalent to a down()/up() operation and we get the ranking for that
inconsistent too.  Both wrt lock_page and wrt, I think, down_read(mmap_sem).

generic_file_buffered_write() does, effectively

	lock_page
	->transaction_start
          ->fault
	  ->down_read(mmap_sem)
	    ->lock_page

and over in do_mmap_pgoff() we nest transaction start inside
down_write(mmap_sem):

	do_mmap_pgoff
	->down_write(mmap_sem)
	->generic_file_mmap
	  ->file_accessed
	    ->mark_inode_dirty
	      ->transaction start

It's all a bit of a mess.  Chris Mason and I have discussed it on and off. 
I think Chris has a workload which actually does trigger a deadlock.

Maybe dropping and retaking mmap_sem in generic_file_mmap would be a
sufficient stopgap.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/