On Mon 02-05-11 10:41:55, Christoph Hellwig wrote:

On Mon, May 02, 2011 at 04:20:55PM +0200, Jan Kara wrote:

quoted

  Hmm, but what prevents the following race?

  Thread 1					Thread 2
..
xfs_trans_alloc()
  xfs_wait_for_freeze(mp, SB_FREEZE_TRANS);
						freeze_super()

                                                  sb->s_frozen = SB_FREEZE_TRANS;

quoted

						  ...
						  xfs_fs_freeze()
						    ...
						    xfs_quiesce_attr()

						      waits for all active
						      transactions

quoted

						    ...

   xfs_trans_alloc
     -> blocks in xfs_wait_for_freeze

  But why should it block when xfs_wait_for_freeze() gets called before
freeze_super() gets called? The other thread calls freeze_super() just
after xfs_wait_for_freeze() in thread 1 and before _xfs_trans_alloc() gets
called.  Or am I missing some serialization somewhere?

     (thus doesn't get to _xfs_trans_alloc)

quoted

  _xfs_trans_alloc()
    atomic_inc(&mp->m_active_trans);
    ... goes on modifying the filesystem

  It seems to be a similar problem as in ext4 - the atomic_inc() and
vfs_check_frozen() are in the wrong order...

I can't see the problem in this scheme.  Note that we want
_xfs_trans_alloc to be able to create a transaction for
xfs_fs_log_dummy, so that we can write the dummy log record after
freezing out all other transactions, so that one is special cased
and doesn't do the xfs_wait_for_freeze.

  OK.

								Honza
-- 
Jan Kara [off-list ref]
SUSE Labs, CR