Re: [PATCH v5 1/2] iomap: fix zero padding data issue in concurrent append writes

[Brian Foster <bfoster@redhat.com>]

On Wed, Dec 04, 2024 at 05:06:00PM +0800, Long Li wrote:
> On Tue, Dec 03, 2024 at 09:54:41AM -0500, Brian Foster wrote:
> > Not sure I see how this is a serialization dependency given that
> > writeback completion also samples i_size. But no matter, it seems a
> > reasonable implementation to me to make the submission path consistent
> > in handling eof.
> > 
> > I wonder if this could just use end_pos returned from
> > iomap_writepage_handle_eof()?
> > 
> > Brian
> > 
> 
> It seems reasonable to me, but end_pos is block-size granular. We need
> to pass in a more precise byte-granular end. 

Well Ok, but _handle_eof() doesn't actually use the value itself so from
that standpoint I see no reason it couldn't at least return the
unaligned end pos. From there, it looks like we do still want the
rounded up value for the various ifs state management calls.

I can see a couple ways of doing that.. one is just align the value in
the caller and use the two variants appropriately. Since the ifs_
helpers all seem to be in byte units, another option could be to
sanitize the helpers to the appropriate start/end rounding internally.

Either of those probably warrant a separate prep patch or two rather
than being squashed in with the i_size fix, but otherwise I'm not seeing
much of a roadblock here. Am I missing something?

Brian

> 
> Thanks,
> Long Li
>