On 08/07/2024 02:44, Dave Chinner wrote:
On Sat, Jul 06, 2024 at 09:56:09AM +0200, Christoph Hellwig wrote:
On Fri, Jul 05, 2024 at 04:24:45PM +0000, John Garry wrote:
- if (xfs_inode_has_bigrtalloc(ip))
+
+ /* Only try to free beyond the allocation unit that crosses EOF */
+ if (xfs_inode_has_forcealign(ip))
+ end_fsb = roundup_64(end_fsb, ip->i_extsize);
+ else if (xfs_inode_has_bigrtalloc(ip))
end_fsb = xfs_rtb_roundup_rtx(mp, end_fsb);
Shouldn't we have a common helper to align things the right way?
Yes, that's what I keep saying.
Such a change was introduced in
https://lore.kernel.org/linux-xfs/20240501235310.GP360919@frogsfrogsfrogs/
and, as you can see, Darrick was less than happy with it. That is why I
kept this method which removed recently added RT code.
Darrick, can we find a better method to factor this code out, like below?
The common way to do this is:
align = xfs_inode_alloc_unitsize(ip);
if (align > mp->m_blocksize)
end_fsb = roundup_64(end_fsb, align);
Wrapping that into a helper might be appropriate, though we'd need
wrappers for aligning both the start (down) and end (up).
To make this work, the xfs_inode_alloc_unitsize() code needs to grow
a forcealign check. That overrides the RT rextsize value (force
align on RT should work the same as it does on data devs) and needs
to look like this:
unsigned int blocks = 1;
+ if (xfs_inode_has_forcealign(ip)
+ blocks = ip->i_extsize;
- if (XFS_IS_REALTIME_INODE(ip))
+ else if (XFS_IS_REALTIME_INODE(ip))
blocks = ip->i_mount->m_sb.sb_rextsize;
That's in 09/13
return XFS_FSB_TO_B(ip->i_mount, blocks);
But more importantly shouldn't this also cover hole punching if we
really want force aligned boundaries?
so doesn't the xfs_file_fallocate(PUNCH_HOLES) ->
xfs_flush_unmap_range() -> rounding with xfs_inode_alloc_unitsize() do
the required job?
Yes, that's what I keep saying. There is no difference in the
alignment behaviour needed for "xfs_inode_has_bigrtalloc" and
"xfs_inode_has_forcealign" except for the source of the allocation
alignment value.
