Re: [PATCH 27/45] xfs: implement buffered writes to zoned RT devices

[Christoph Hellwig <hch@lst.de>]

On Wed, Feb 19, 2025 at 01:47:27PM -0800, Darrick J. Wong wrote:
> I don't want to go adding opencoded logic loops all over the place, that
> would be pretty horrid.  But what if xfs_zone_alloc_ctx were instead a
> general freecounter reservation context?  Then we could hide all this
> "try to reserve space, push a garbage collector once if we can't, and
> try again once" logic into an xfs_reserve_space() function, and pass
> that reservation through the iomap functions to ->iomap_begin.
> 
> But a subtlety here is that the code under iomap_file_buffered_write
> might not actually need to create any delalloc reservations, in which
> case a worst case reservation could fail with ENOSPC even though we
> don't actually need to allocate a single byte.

I think the idea of per-reserving space before starting transactions
is generally a good idea, and I'd be happy to look into reworking the
conventional code towards that.  But I'd rather not do that as part
of this series if I can avoid it.

> Having said that, this is like the 5th time I've read through this patch
> and I don't see anything obviously wrong now, so my last question is:
> Zoned files cannot have preallocations so that's why we don't do this
> for FALLOC_FL_ALLOCATE_RANGE, right?

Exactly.  Should I add this to the commit log?  It's mentioned mostly
in the cover letter at the moment.