Re: [PATCH v8 2/2] xfs: add support for FALLOC_FL_WRITE_ZEROES

[Pankaj Raghav <pankaj.raghav@linux.dev>]

>> +	/*
>> +	 * Allocate written, zeroed extents across the range.  xfs_alloc_file_space()
>> +	 * rounds outward to block granularity:
>> +	 *  - holes (the punched interior and any unallocated edge block) are
>> +	 *    allocated and zeroed;
>> +	 *  - unwritten extents (including unwritten edge blocks) are converted to
>> +	 *    written and zeroed;
>> +	 *  - Already written edge blocks are skipped. The out-of-range bytes of
>> +	 *    a written edge block keep their data (offset_rd -> offset and
>> +	 *    end -> end_rd); their in-range bytes (offset -> offset_ru and
>> +	 *    end_ru -> end were already zeroed by xfs_free_file_space().
>> +	 */
>> +	return xfs_alloc_file_space(ip, offset, len,
>> +			XFS_ALLOC_FILE_SPACE_WRITE_ZEROES);
> 
> ...and now we can just do an accelerated "write zeroes to disk" which is
> conveniently always within EOF now.  I /think/ this looks ok to me now,
> though I'm curious how extensively the new fallocate mode has been
> tested with fsx and unaligned file ranges?  And rt volumes with rt
> extent size > 1 fsblock.
> 

I am running into an issue with rtvol.

Ran generic/363 (fsx) on rt, rextsize=2 (4k block / 8k rt extent). First I ran into this:

xfs_bmap_rtalloc: ASSERT(xfs_extlen_to_rtxmod == 0). WRITE_ZEROES uses
XFS_BMAPI_CONVERT, and xfs_bmap_extsize_align() bails early on convert, so the
rt-unaligned range becomes a sub-rt-extent allocation. I fixed it by rounding the alloc range
out to whole rt extents, like xfs_free_file_space() does for rt extents.

diff --git a/fs/xfs/xfs_bmap_util.c b/fs/xfs/xfs_bmap_util.c
index 855602cb35e8..e52ad4c25b66 100644
--- a/fs/xfs/xfs_bmap_util.c
+++ b/fs/xfs/xfs_bmap_util.c
@@ -710,6 +710,13 @@ xfs_alloc_file_space(
        imapp = &imaps[0];
        startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
        endoffset_fsb = XFS_B_TO_FSB(mp, offset + count);
+
+       if (mode == XFS_ALLOC_FILE_SPACE_WRITE_ZEROES &&
+           xfs_inode_has_bigrtalloc(ip)) {
+               startoffset_fsb = xfs_fileoff_rounddown_rtx(mp, startoffset_fsb);
+               endoffset_fsb = xfs_fileoff_roundup_rtx(mp, endoffset_fsb);
+       }
+
        allocatesize_fsb = endoffset_fsb - startoffset_fsb;

        /*

With that "fixed", I ran into WARN_ON_ONCE(folio_pos > i_size) in iomap_zero_iter,
via xfs_falloc_setsize -> xfs_setattr_size -> xfs_zero_range. The
rt-aligned alloc now leaves written zeroed blocks past a non-rt-aligned
EOF. So WRITE_ZEROES on rtvol breaks "no written blocks past EOF".

I am not sure what would be the most clean way of solving this issue for rtvol.

Let me know your thoughts!

@Zhang yi: did you test this on bigalloc configurations? Maybe you run into a similar problem
when cluster size (extsize for XFS) > block size?

--
Pankaj