On Tue 25-08-20 10:10:20, Theodore Y. Ts'o wrote:
> (Adding the OCFS2 maintainers, since my possibly insane idea proposed
> below would definitely impact them!)
>
> On Tue, Aug 25, 2020 at 01:16:16PM +0100, Christoph Hellwig wrote:
> > On Tue, Aug 25, 2020 at 02:05:54PM +0200, Jan Kara wrote:
> > > Discarding blocks and buffers under a mounted filesystem is hardly
> > > anything admin wants to do. Usually it will confuse the filesystem and
> > > sometimes the loss of buffer_head state (including b_private field) can
> > > even cause crashes like:
> >
> > Doesn't work if the file system uses multiple devices. I think we
> > just really need to split the fs buffer_head address space from the
> > block device one. Everything else is just going to cause a huge mess.
>
> I wonder if we should go a step further, and stop using struct
> buffer_head altogether in jbd2 and ext4 (as well as ocfs2).
What about the cache coherency issues I've pointed out in my reply to
Christoph?
> This would involve moving whatever structure elements from the
> buffer_head struct into journal_head, and manage writeback and reads
> requests directly in jbd2. This would allow us to get detailed write
> errors back, which is currently not possible from the buffer_head
> infrastructure.
>
> The downside is this would be a pretty massive change in terms of LOC,
> since we use struct buffer_head in a *huge* number of places. If
> we're careful, most of it could be handled by a Coccinelle script to
> rename "struct buffer_head" to "struct journal_head". Fortunately, we
> don't actually use that much of the fs/buffer_head functions in
> fs/{ext4,ocfs2}/*.c.
>
> One potentially tricky bit is that ocfs2 hasn't been converted to
> using iomap, so it's still using __blockdev_direct_IO. So it's data
> blocks for DIO would still have to use struct buffer_head (which means
> the Coccinelle script won't really work for fs/ocfs2, without a lot of
> manual rework) --- or ocfs2 would have to switched to use iomap at
> least for DIO support.
>
> What do folks think?
Otherwise yes, this would be doable although pretty invasive as you
mention.
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
