Fix deadlock in journal_create

Marcelo,
I sent this initially against 2.4.21-rc6 and it didn't make it in
-- even though it got Stephen's OK. Here's a resend -- I've identified that
the bug still exists and 2.4.22-pre7. The patch didn't need to be changed as it still applies cleanly.

I ran across a deadlock when trying to do a journal_create
at mount time. The problem is that journal_create does a sync_dev which eventually tries to do a get_super which does:

down_read(&s->s_umount);

The problem arises if I call journal_create from my read_super method in which case get_sb_bdev has already done:

down_write(&s->s_umount);

from alloc_super. Replacing the sync_dev call with an fsync_no_super seems to have fixed the deadlock.

It you want to test this out using ext3 (I have verified it with my own filesystem), simply follow these steps:
1) create an ext2 filesystem on a device
2) mount that new partition and make a "journal file" on it (using dd). get the inode number of that file.
3) unmount it and remount it as an ext3 filesystem using the option:
journal=inode_number where inode number is the inode number of the journal file you just created.
This will get ext3's read_super method to call ext3_create_journal which will hang during a journal_create.

A trivial patch to fix this is attached.

	--Mark

--
Mark Fasheh
Software Developer, Oracle Corp
[email protected]

@@ -912,7 +912,7 @@ int journal_create(journal_t *journal)
 		__brelse(bh);
 	}
 
-	sync_dev(journal->j_dev);
+	fsync_no_super(journal->j_dev);
 	jbd_debug(1, "JBD: journal cleared.\n");
 
 	/* OK, fill in the initial static fields in the new superblock */

Hi,

Marcelo, please apply --- the jbd internal journal create code is in practice unused by ext3 these days now that e2fsprogs groks journals, but other filesystems wanting to use that code will need this patch.

Cheers,
Stephen