Re: Delayed allocation and page_lock vs transaction start ordering

[Jan Kara <jack@suse.cz>]

On Wed 28-05-08 15:13:52, Aneesh Kumar K.V wrote:
> On Wed, May 28, 2008 at 11:33:24AM +0200, Jan Kara wrote:
> > On Tue 27-05-08 20:41:28, Aneesh Kumar K.V wrote:
> > > On Tue, May 27, 2008 at 02:43:12PM +0200, Jan Kara wrote:
> > > > On Mon 26-05-08 23:30:43, Aneesh Kumar K.V wrote:
> > > > 
> > > > > I have got another question now related to page_mkwrite. AFAIU writepage
> > > > > writeout dirty buffer_heads. It also looks at whether the pages are
> > > > > dirty or not. In the page_mkwrite callback both are not true. ie we call
> > > > > set_page_dirty from do_wp_page after calling page_mkwrite. I haven't
> > > > > verified whether the above is correct or not. Just thinking reading the
> > > > > code.
> > > >   Writepage call itself doesn't look at whether the page is dirty or not -
> > > > that flag is already cleared when writepage is called. You are right that
> > > > the page is marked dirty only after page_mkwrite is called - the meaning of
> > > > page_mkwrite() call is roughly "someone wants to do the first write to this
> > > > page via mmap, prepare filesystem for that". But we don't really care
> > > > whether the page is dirty or not - we know it carries correct data (it is
> > > > uptodate) and so we can write it if we want (and need).
> > > > 
> > > 
> > > I am looking at  __block_write_full_page and we have
> > > 
> > > if (!buffer_mapped(bh) && buffer_dirty(bh)) {
> > > 	WARN_ON(bh->b_size != blocksize);
> > > 	err = get_block(inode, block, bh, 1);
> > > 	if (err)
> > > 
> > > ie, we do get_block only if the buffer_head is dirty. So I am bit
> > > doubtful whether we are actually allocating blocks via page_mkwrite.
> >   Good catch, we should mark unmapped buffers dirty before calling writepage.
> > Actually, if the page didn't have any buffers, block_write_full_page() will
> > create them all dirty so that's probably why I didn't hit it in my testing
> > but it's definitely safer to mark them dirty explicitely. Thanks.
> 
> looking at create_empty_buffers we do that only if page is marked as
> dirty. In the case of page_mkwrite the page is also not marked dirty
> when we call the call back right ?
  But in block_write_full_page() we do:
        if (!page_has_buffers(page)) {
                create_empty_buffers(page, blocksize,
                                        (1 << BH_Dirty)|(1 << BH_Uptodate));
        }
  So buffers are created dirty...

> >   It is enough to change ext4_bh_mapped() to something like:
> > static int ext4_bh_prepare_fill(handle_t *handle, struct buffer_head *bh)
> > {
> > 	if (!buffer_mapped(bh)) {
> > 		/*
> > 		 * Mark buffer as dirty so that block_write_full_page()
> > 		 * writes it
> > 		 */
> > 		set_buffer_dirty(bh);
> > 		return 1;
> > 	}
> > 	return 0;
> > }
> > 
> >   Should I send you an updated patch with this change and the changes we spoke
> > about yesterday, or just an incremental changes which you will fold yourself
> > into the big one?
> > 
> 
> This will mark only the first unmapped buffer_head as dirty. What about
> the rest of the buffer_heads in the page that are unmapped ?
  Oops, I forgot that walk_page_buffers() stops after the first non-zero
return. So we have to split the function - keep ext4_bh_mapped() and
add one more traversal with in case there is some unmapped buffer:

static int ext4_bh_prepare_fill(handle_t *handle, struct buffer_head *bh)
{
	if (!buffer_mapped(bh)) {
		/*
		 * Mark buffer as dirty so that block_write_full_page()
		 * writes it
		 */
		set_buffer_dirty(bh);
	}
	return 0;
}

> I am looking at pushing the ext4_page_mkwrite before rest of  the
> changes. That is needed to handle ENOSPC when mmap write to files with
> holes.
  I see. OK.

								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR