On Thu, 29 Aug 2024, Mike Snitzer wrote:
> +
> +bool nfs_uuid_is_local(const uuid_t *uuid, struct net *net, struct auth_domain *dom)
> +{
> + bool is_local = false;
> + nfs_uuid_t *nfs_uuid;
> +
> + rcu_read_lock();
> + nfs_uuid = nfs_uuid_lookup(uuid);
> + if (nfs_uuid) {
> + nfs_uuid->net = maybe_get_net(net);
I know I said it looked wrong to be getting a ref for the domain but not
the net - and it did. But that doesn't mean the fix was to get a ref
for the net and to hold it indefinitely.
This ref is now held until the client happens to notice that localio
doesn't work any more (because nfsd_serv_try_get() fails). So the
shutdown of a net namespace will be delayed indefinitely if the NFS
filesystem isn't being actively used.
I would prefer that there were a way for the net namespace to reach back
into the client and disconnect itself. Probably this would be a
linked-list in struct nfsd_net which linked list_heads in struct
nfs_client. This list would need to be protected by a spinlock -
probably global in nfs_common so client could remove itself and server
could remove all clients after clearing their net pointers.
It is probably best if I explain all of what I am thinking as a patch.
Stay tuned.
NeilBrown
