On Tue, May 10, 2011 at 02:51:24PM +0200, Jan Kara wrote:

On Mon 09-05-11 16:03:18, Darrick J. Wong wrote:

quoted

I am still chasing down what exactly is broken in ext3.  data=writeback mode
passes with no failures.  data=ordered, however, does not pass; my current
suspicion is that jbd is calling submit_bh on data buffers but doesn't call
page_mkclean to kick the userspace programs off the page before writing it.

  Yes, ext3 in data=ordered mode writes pages from
journal_commit_transaction() via submit_bh() without clearing page dirty
bits thus page_mkclean() is not called for these pages. Frankly, do you
really want to bother with adding support for ext2 and ext3? People can use
ext4 as a fs driver when they want to start using blk-integrity support.
Especially ext2 patch looks really painful and just from a quick look I can
see code e.g. in fs/ext2/namei.c which isn't handled by your patch yet.

Yeah, I agree that ext2 is ugly and ext3/jbd might be more painful.  Are there
any other code that wants stable pages that's already running with ext3?  In
this months-long discussion I've heard that encryption and raid also like
stable pages during writes.  Have those users been broken this whole time, or
have they been stabilizing pages themselves?

I suppose we can cross the "ext3 fails horribly on DIF" bridge when someone
complains about it.  Possibly we could try to steer them to btrfs.

--D