[RFC] probably invalid accounting in jbd

[Alex Tomas <bzzz@tmi.comex.ru>]

hi!

I think current journal_release_buffer() which is used by ext3 allocator
in -mm tree has a bug. look, it won't decrement credits if concurrent
thread already put buffer on metadata list. but this may ext3_try_to_allocate()
may overflow handle's credist.
so, jbd will hit J_ASSERT_JH(jh, handle->h_buffer_credits > 0) somewhere


void journal_release_buffer (handle_t *handle, struct buffer_head *bh)
{
        .....
        if (jh->b_jlist == BJ_Reserved && jh->b_transaction == transaction &&
            !buffer_jdirty(jh2bh(jh))) {
                JBUFFER_TRACE(jh, "unused: refiling it");
                handle->h_buffer_credits++;
                __journal_refile_buffer(jh);
        }
        ....
}


here is the patch against 2.5.69-mm6:

diff -puNr linux-2.5.69-mm6/fs/jbd/transaction.c edited/fs/jbd/transaction.c
--- linux-2.5.69-mm6/fs/jbd/transaction.c	Fri May 16 18:03:20 2003
+++ b_commited_data-race/fs/jbd/transaction.c	Fri May 23 11:10:21 2003
@@ -1146,10 +1146,10 @@ void journal_release_buffer (handle_t *h
 	if (jh->b_jlist == BJ_Reserved && jh->b_transaction == transaction &&
 	    !buffer_jdirty(jh2bh(jh))) {
 		JBUFFER_TRACE(jh, "unused: refiling it");
-		handle->h_buffer_credits++;
 		__journal_refile_buffer(jh);
 	}
 	spin_unlock(&journal_datalist_lock);
+	handle->h_buffer_credits++;
 
 	JBUFFER_TRACE(jh, "exit");
 	unlock_journal(journal);