Re: [RFC 2/3] ext2: Convert ext2 regular file buffered I/O to use iomap

[Christoph Hellwig <hch@infradead.org>]

On Thu, Nov 30, 2023 at 08:54:31AM +0530, Ritesh Harjani wrote:
> So I took a look at this. Here is what I think -
> So this is useful of-course when we have a large folio. Because
> otherwise it's just one block at a time for each folio. This is not a
> problem once FS buffered-io handling code moves to iomap (because we
> can then enable large folio support to it).

Yes.

> However, this would still require us to pass a folio to ->map_blocks
> call to determine the size of the folio (which I am not saying can't be
> done but just stating my observations here).

XFS currently maps based on the underlyig reservation (delalloc extent)
and not the actual map size.   This works because on-disk extents are
allocated as unwritten extents, and only the actual written part is
the converted.  But if you only want to allocate blocks for the part
actually written you actually need to pass in the dirty range and not
just use the whole folio.  This would be the incremental patch to do
that:

http://git.infradead.org/users/hch/xfs.git/commitdiff/0007893015796ef2ba16bb8b98c4c9fb6e9e6752

But unless your block allocator is very cheap doing what XFS does is
probably going to work much better.

> ...ok so here is the PoC for seq counter check for ext2. Next let me
> try to see if we can lift this up from the FS side to iomap - 

This looks good to me from a very superficial view.  Dave is the expert
on this, though.