On Wed 29-07-09 17:04:14, Tao Ma wrote:> Hi Jan,
> I am now reading quota support in ocfs2 and have some qs. Wish you can
> help me clarify it. Thanks.
>
> ocfs2_quota_write:
> In the comment, you said that "we know the transaction is already
> started", so it should be called within a transaction. But in this
> function, we call ocfs2_extend_no_holes and ocfs2_simple_size_update which
> will start another transaction. So we can survive with nested transaction
> or this scenario will never happen?
The comment is right, we absolutely require a transaction being started
when when entering ocfs2_quota_write(), otherwise deadlocks could occur.
JBD supports nested transactions - when journal_start() is called and
the current process already has a transaction open, it just returns the
currently open transaction. So you only have to be sure to initially ask
for enough credits when opening the transaction - and quota code hopefully
gets this right.
Hmm, but looking at the code again, there's one catch - normally,
allocator locks nest outside the transaction and in ocfs2_quota_write()
they are acquired inside the transaction and that can lead to deadlocks.
Aw, that's going to be nasty to fix... I have to rethink the locking
logic...
Thanks for your inquiry :)
Honza
--
Jan Kara <jack at suse.cz>
SUSE Labs, CR