Re: [PATCH 6/9] spaceman/defrag: workaround kernel xfs_reflink_try_clear_inode_flag()

[Dave Chinner <david@fromorbit.com>]

On Tue, Jul 09, 2024 at 12:10:25PM -0700, Wengang Wang wrote:
> xfs_reflink_try_clear_inode_flag() takes very long in case file has huge number
> of extents and none of the extents are shared.

Got a kernel profile showing how bad it is?

> 
> workaround:
> share the first real extent so that xfs_reflink_try_clear_inode_flag() returns
> quickly to save cpu times and speed up defrag significantly.

That's nasty.

Let's fix the kernel code, not work around it in userspace.

I mean, it would be really easy to store if an extent is shared in
the iext btree record for the extent. If we do an unshare operation,
just do a single "find shared extents" pass on the extent tree and
mark all the extents that are shared as shared.  Then set a flag on
the data fork saying it is tracking shared extents, and so when we
share/unshare extents in that inode from then on, we set/clear that
flag in the iext record. (i.e. it's an in-memory equivalent of the
UNWRITTEN state flag).

Then after the first unshare, checking for nothing being shared is a
walk of the iext btree over the given range, not a refcountbt
walk. That should be much faster.

And we could make it even faster by adding a "shared extents"
counter to the inode fork. i.e. the first scan that sets the flags
also counts the shared extents, and we maintain that as we maintain
the iin memory extent flags....

That makes the cost of xfs_reflink_try_clear_inode_flag() basically
go to zero in these sorts of workloads. IMO, this is a much better
solution to the problem than hacking around it in userspace...

-Dave.
-- 
Dave Chinner
david@fromorbit.com