Re: [PATCH 15/16] btrfs: refactor the zoned device handling in cow_file_range

[Christoph Hellwig <hch@lst.de>]

On Tue, Jun 06, 2023 at 02:20:50PM +0000, Naohiro Aota wrote:
> > +			/*
> > +			 * For zoned devices, let the caller retry after writing
> > +			 * out the already allocated regions or waiting for a
> > +			 * zone to finish if no allocation was possible at all.
> > +			 *
> > +			 * Else convert to -ENOSPC since the caller cannot
> > +			 * retry.
> > +			 */
> > +			if (btrfs_is_zoned(fs_info)) {
> > +				if (start == orig_start)
> > +					return -EAGAIN;
> > +				*done_offset = start - 1;
> 
> This will hit a null pointer dereference when it is called from
> submit_uncompressed_range(). Well, that means we need to implement the
> partial writing also in submit_uncompressed_range().

I think we need to do that anyway, don't we?  As far as I can tell
submit_uncompressed_range is simply broken right now on zoned devices
if cow_file_range fails to allocate all blocks.

> BTW, we also need to handle -EAGAIN from btrfs_reserve_extent() in
> btrfs_new_extent_direct() as it currently returns -EAGAIN to the userland.

Indeed.

> >  			ASSERT(ret == 0);
> >  			return 0;
> >  		}
> > +		if (ret == -EAGAIN) {
> > +			ASSERT(btrfs_is_zoned(inode->root->fs_info));
> 
> Is this necessary? run_delalloc_zoned() should be called only for zoned
> mode anyway.

True.