Re: [PATCH v2 14/23] NFSv4/flexfiles: Honor ndc_immediate on CB_NOTIFY_DEVICEID CHANGE

["Anna Schumaker" <anna@kernel.org>]

On Fri, Aug 21, 2026, at 12:29 PM, Benjamin Coddington wrote:
> When a CHANGE notification carries ndc_immediate, RFC 8881 Section
> 20.12 says the change is enforced immediately and the client might not
> be able to complete pending I/O.  In addition to un-pinning the stripe's
> device node, mark the old node unavailable.
>
> Marking does not recall the references already handed out.  A write
> whose DS connection is already up keeps using the old node until that
> I/O errors: nfs4_ff_layout_prepare_ds() returns early on a live
> ds_clp, and the unavailable flag is only consulted when a connection
> is being established.  What the mark does change is that a read skips
> the node while another mirror is usable, and that an IOMODE_RW segment
> still pinning it stops counting as fully available -- so I/O the
> server rejects falls back to the MDS rather than being retried against
> a mapping the server has already withdrawn.
>
> The walk can also exchange out a node that already carries the new
> mapping: a re-resolve that completed between the unhash and the walk
> reaching the stripe, raced by this walk or by the walk of a later
> CHANGE for the same deviceid.  Ripping such a node out is harmless (it
> is still hashed, so the next I/O re-pins it from the cache), but it
> must not be marked unavailable.  Stale vintages are distinguishable by
> hashed-ness: every superseded node was unhashed before its walk began
> and is never re-inserted, while a fresh node is inserted before it is
> installed.  Only mark nodes that are no longer hashed.
>
> That test is hlist_unhashed_lockless(): the hook runs under the layout
> inode's i_lock and rcu_read_lock(), but not under nfs4_deviceid_lock,
> which is what serializes the writers of node.pprev -- and __hlist_del()
> stores a neighbour's pprev with WRITE_ONCE(), so removing any other
> entry in the same bucket can write the field this test reads.
>
> Without ndc_immediate, pending I/O drains on the old mapping and only
> new I/O re-resolves, as before.
>
> Assisted-by: Claude:claude-fable-5
> Signed-off-by: Benjamin Coddington <bcodding@hammerspace.com>
> ---
>  fs/nfs/flexfilelayout/flexfilelayout.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
>
> diff --git a/fs/nfs/flexfilelayout/flexfilelayout.c 
> b/fs/nfs/flexfilelayout/flexfilelayout.c
> index 46ca58e8e96b..ebb19a919f9d 100644
> --- a/fs/nfs/flexfilelayout/flexfilelayout.c
> +++ b/fs/nfs/flexfilelayout/flexfilelayout.c
> @@ -2541,6 +2541,14 @@ static void ff_layout_reresolve_deviceid(struct 
> pnfs_layout_hdr *lo,
>  				kfree(put);
>  				continue;
>  			}
> +			/* Only mark stale vintages: a node still hashed was

I don't think I've seen "vintage" used when talking about anything other than
wine before.

> +			 * fetched after the unhash and carries the new
> +			 * mapping.  Lockless -- nfs4_deviceid_lock, which
> +			 * serialises pprev, is not held here.
> +			 */

Nit: this comment is pretty wordy. Do we really need to define what "lockless"
means here?

Anna

> +			if (immediate &&
> +			    hlist_unhashed_lockless(&old->id_node.node))
> +				nfs4_mark_deviceid_unavailable(&old->id_node);
>  			put->dev = &old->id_node;
>  			list_add(&put->node, head);
>  		}
> -- 
> 2.53.0