The netfs copy-to-cache that is used by Ceph with local caching sets up a
new request to write data just read to the cache.  The request is started
and then left to look after itself whilst the app continues.  The request
gets notified by the backing fs upon completion of the async DIO write, but
then tries to wake up the app because NETFS_RREQ_OFFLOAD_COLLECTION isn't
set - but the app isn't waiting there, and so the request just hangs.

Fix this by setting NETFS_RREQ_OFFLOAD_COLLECTION which causes the
notification from the backing filesystem to put the collection onto a work
queue instead.

Fixes: e2d46f2ec332 ("netfs: Change the read result collector to only use one work item")
Reported-by: Max Kellermann <max.kellermann@ionos.com>
Link: https://lore.kernel.org/r/CAKPOu+8z_ijTLHdiCYGU_Uk7yYD=shxyGLwfe-L7AV3DhebS3w@mail.gmail.com/
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Paulo Alcantara <pc@manguebit.org>
cc: Viacheslav Dubeyko <slava@dubeyko.com>
cc: Alex Markuze <amarkuze@redhat.com>
cc: Ilya Dryomov <idryomov@gmail.com>
cc: netfs@lists.linux.dev
cc: ceph-devel@vger.kernel.org
cc: linux-fsdevel@vger.kernel.org
cc: stable@vger.kernel.org
---
 fs/netfs/read_pgpriv2.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/fs/netfs/read_pgpriv2.c b/fs/netfs/read_pgpriv2.c
index 5bbe906a551d..080d2a6a51d9 100644
--- a/fs/netfs/read_pgpriv2.c
+++ b/fs/netfs/read_pgpriv2.c
@@ -110,6 +110,7 @@ static struct netfs_io_request *netfs_pgpriv2_begin_copy_to_cache(
 	if (!creq->io_streams[1].avail)
 		goto cancel_put;
 
+	__set_bit(NETFS_RREQ_OFFLOAD_COLLECTION, &creq->flags);
 	trace_netfs_write(creq, netfs_write_trace_copy_to_cache);
 	netfs_stat(&netfs_n_wh_copy_to_cache);
 	rreq->copy_to_cache = creq;

On Fri, Jul 11, 2025 at 5:10 PM David Howells <dhowells@redhat.com> wrote:
>
> The netfs copy-to-cache that is used by Ceph with local caching sets up a
> new request to write data just read to the cache.  The request is started
> and then left to look after itself whilst the app continues.  The request
> gets notified by the backing fs upon completion of the async DIO write, but
> then tries to wake up the app because NETFS_RREQ_OFFLOAD_COLLECTION isn't
> set - but the app isn't waiting there, and so the request just hangs.
>
> Fix this by setting NETFS_RREQ_OFFLOAD_COLLECTION which causes the
> notification from the backing filesystem to put the collection onto a work
> queue instead.

Thanks David, you can add me as Tested-by if you want.

I can't test the other patch for the next two weeks (vacation). When
I'm back, I'll install both fixes on some heavily loaded production
machines - our clusters always shake out the worst in every piece of
code they run!