Re: [PATCH 17/17] xfs: reduce ilock roundtrips in xfs_qm_vop_dqalloc

["Darrick J. Wong" <djwong@kernel.org>]

On Fri, Oct 17, 2025 at 05:50:43AM +0200, Christoph Hellwig wrote:
> On Thu, Oct 16, 2025 at 08:59:41AM -0700, Darrick J. Wong wrote:
> > > > ...and I guess we no longer detach dquots from live inodes now, so we
> > > > really only need ILOCK_EXCL to prevent multiple threads from trying to
> > > > allocate and attach a new xfs_dquot object to the same inode, right?
> > > 
> > > Yes.
> > 
> > I wonder then, if there /were/ two threads racing to dqattach the same
> > inode, won't the radix_tree_insert return EEXIST, preventing us
> > from exposing two dquot for the same id because xfs_qm_dqget will just
> > loop again?
> 
> I think so.
> 
> > [Though looking at that xfs_qm_dqget -> xfs_qm_dqget_cache_insert ->
> > radix_tree_insert sequence, it looks like we can also restart
> > indefinitely on other errors like ENOMEM.]
> 
> Yes.  I have patches fixing that as part of moving to xarrays instead
> of radix trees.  But I suspect a resizable hash table might actually
> be the better fit, so I didn't look into submitting that quite yet.

<nod> Well if either of those are coming next then this looks ok to me.
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>

--D