On Mon, Jun 28, 2021 at 10:59:55PM +0100, Matthew Wilcox wrote:

quoted

quoted

so permit pages without an iop to enter writeback and create an iop
*then*.  Would that solve your problem?

It is the right thing to do, especially when combined with a feature
patch to not bother to create the iomap_page structure on small
block size file systems when the extent covers the whole page.

We don't know the extent layout at the point where *this* patch creates
iomap_pages during writeback.  I imagine we can delay creating one until
we find out what our destination layout will be?

Hmm.  Actually ->page_mkwrite is always is always called on an uptodate
page and we even assert that.  I should have remembered the whole page
fault path better.

So yeah, I think we should take patch 1 from Andreas, then a non-folio
version of your patch as a start.  The next steps then would be in
approximate order:

 1. remove the iomap_page_create in iomap_page_mkwrite_actor as it
    clearly is not needed at that point
 2. don't bother to create an iomap_page in iomap_readpage_actor when
    the iomap spans the whole page
 3. don't create the iomap_page in __iomap_write_begin when the
    page is marked uptodate or the write covers the whole page 

delaying the creation further in iomap_writepage_map will be harder
as the loop around iomap_add_to_ioend is still fundamentally block
based.