On Tue, 2025-09-02 at 20:10 +0800, Li Lingfeng wrote:
> Hi,
>
> 在 2025/9/2 18:21, Jeff Layton 写道:
> > On Tue, 2025-09-02 at 10:22 +0800, Li Lingfeng wrote:
> > > When file access conflicts occur between clients, the server recalls
> > > delegations. If the client holding delegation fails to return it after
> > > a recall, nfs4_laundromat adds the delegation to cl_revoked list.
> > > This causes subsequent SEQUENCE operations to set the
> > > SEQ4_STATUS_RECALLABLE_STATE_REVOKED flag, forcing the client to
> > > validate all delegations and return the revoked one.
> > >
> > > However, if the client fails to return the delegation due to a timeout
> > > after receiving the recall or a server bug, the delegation remains in the
> > > server's cl_revoked list. The client marks it revoked and won't find it
> > > upon detecting SEQ4_STATUS_RECALLABLE_STATE_REVOKED. This leads to a loop:
> > > the server persistently sets SEQ4_STATUS_RECALLABLE_STATE_REVOKED, and the
> > > client repeatedly tests all delegations, severely impacting performance
> > > when numerous delegations exist.
> > >
> > It is a performance impact, but I don't get the "loop" here. Are you
> > saying that this problem compounds itself? That testing all delegations
> > causes others to be revoked?
> The delegation will be removed from server->delegations in client after
> NFSPROC4_CLNT_DELEGRETURN is performed.
> nfs4_delegreturn_done
> nfs_delegation_mark_returned
> nfs_detach_delegation
> nfs_detach_delegation_locked
> list_del_rcu // remove delegation from server->delegations
>
> From the client's perspective, the delegation has been returned, but on
> the server side, it is left in the cl_revoked list.[1].
>
> Subsequently, every sequence from the client will be flagged with
> SEQ4_STATUS_RECALLABLE_STATE_REVOKED as long as cl_revoked remains
> non-empty.
> nfsd4_sequence
> seq->status_flags |= SEQ4_STATUS_RECALLABLE_STATE_REVOKED
>
> When the client detects SEQ4_STATUS_RECALLABLE_STATE_REVOKED while
> processing a sequence result, it sets NFS_DELEGATION_TEST_EXPIRED for all
> delegations and wakes up the state manager for handling.
> nfs41_sequence_done
> nfs41_sequence_process
> nfs41_handle_sequence_flag_errors
> nfs41_handle_recallable_state_revoked
> nfs_test_expired_all_delegations
> nfs_mark_test_expired_all_delegations
> nfs_delegation_mark_test_expired_server
> // set NFS_DELEGATION_TEST_EXPIRED for delegations in
> server->delegations
> nfs4_schedule_state_manager
>
> The state manager tests all delegations except the one that was returned,
> as it is no longer in server->delegations.
> nfs4_state_manager
> nfs4_begin_drain_session
> nfs_reap_expired_delegations
> nfs_server_reap_expired_delegations
> // test delegations in server->delegations
>
> There may be a loop:
> 1) send a sequence(client)
> 2) return SEQ4_STATUS_RECALLABLE_STATE_REVOKED(server)
> 3) set NFS_DELEGATION_TEST_EXPIRED for all delegations(client)
> 4) test all delegations by state manager(client)
> 5) send another sequence(client)
>
> The state manager's traversal of delegations occurs between
> nfs4_begin_drain_session and nfs4_end_drain_session. Non-privileged requests
> will be blocked because the NFS4_SLOT_TBL_DRAINING flag is set. If there are
> many delegations to traverse, this blocking time can be relatively long.
> > > Since abnormal delegations are removed from flc_lease via nfs4_laundromat
> > > --> revoke_delegation --> destroy_unhashed_deleg -->
> > > nfs4_unlock_deleg_lease --> kernel_setlease, and do not block new open
> > > requests indefinitely, retaining such a delegation on the server is
> > > unnecessary.
> > >
> > > Reported-by: Zhang Jian <zhangjian496@xxxxxxxxxx>
> > > Fixes: 3bd64a5ba171 ("nfsd4: implement SEQ4_STATUS_RECALLABLE_STATE_REVOKED")
> > > Closes: https://lore.kernel.org/all/ff8debe9-6877-4cf7-ba29-fc98eae0ffa0@xxxxxxxxxx/
> > > Signed-off-by: Li Lingfeng <lilingfeng3@xxxxxxxxxx>
> > > ---
> > > fs/nfsd/nfs4state.c | 11 +++++++++++
> > > 1 file changed, 11 insertions(+)
> > >
> > > diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
> > > index 88c347957da5..aa65a685dbb9 100644
> > > --- a/fs/nfsd/nfs4state.c
> > > +++ b/fs/nfsd/nfs4state.c
> > > @@ -4326,6 +4326,8 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
> > > int buflen;
> > > struct net *net = SVC_NET(rqstp);
> > > struct nfsd_net *nn = net_generic(net, nfsd_net_id);
> > > + struct list_head *pos, *next;
> > > + struct nfs4_delegation *dp;
> > >
> > > if (resp->opcnt != 1)
> > > return nfserr_sequence_pos;
> > > @@ -4470,6 +4472,15 @@ nfsd4_sequence(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
> > > default:
> > > seq->status_flags = 0;
> > > }
> > > + if (!list_empty(&clp->cl_revoked)) {
> > > + list_for_each_safe(pos, next, &clp->cl_revoked) {
> > > + dp = list_entry(pos, struct nfs4_delegation, dl_recall_lru);
> > > + if (dp->dl_time < (ktime_get_boottime_seconds() - 2 * nn->nfsd4_lease)) {
> > > + list_del_init(&dp->dl_recall_lru);
> > > + nfs4_put_stid(&dp->dl_stid);
> > > + }
> > > + }
FYI: this list is protected by the clp->cl_lock. You need to hold it to
do this list walk.
> > > + }
> > > if (!list_empty(&clp->cl_revoked))
> > > seq->status_flags |= SEQ4_STATUS_RECALLABLE_STATE_REVOKED;
> > > if (atomic_read(&clp->cl_admin_revoked))
> > This seems like a violation of the spec. AIUI, the server is required
> > to hang onto a record of the delegation until the client does the
> > TEST_STATEID/FREE_STATEID dance to remove it. Just discarding them like
> > this seems wrong.
> Our expected outcome was that the client would release the abnormal
> delegation via TEST_STATEID/FREE_STATEID upon detecting its invalidity.
> However, this problematic delegation is no longer present in the
> client's server->delegations list—whether due to client-side timeouts or
> the server-side bug [1].
> >
> > Should we instead just administratively evict the client since it's
> > clearly not behaving right in this case?
> Thanks for the suggestion. While administratively evicting the client would
> certainly resolve the immediate delegation issue, I'm concerned that
> approach
> might be a bit heavy-handed.
> The problematic behavior seems isolated to a single delegation. Meanwhile,
> the client itself likely has numerous other open files and active state on
> the server. Forcing a complete client reconnect would tear down all that
> state, which could cause significant application disruption and be perceived
> as a service outage from the client's perspective.
>
> [1]
> https://lore.kernel.org/all/de669327-c93a-49e5-a53b-bda9e67d34a2@xxxxxxxxxx/
>
> Thanks,
> Lingfeng
Ok, I get the problem, but I still disagree with the solution. I don't
think we can just time these things out. Ideally we'd close the race
window, but the sc_status field is protected by the global state_lock
and I don't think we want to take it in revoke_delegation.
The best solution I can see is to have destroy_delegation()
unconditionally set SC_STATUS_CLOSED, and then you can do the list walk
above, but checking for that flag instead of testing for a timeout.
I'm still not thrilled with this solution though. It makes SEQUENCE a
bit more heavyweight than I'd like. I'm starting to think that we need
to rework the overall delegation locking, but that's an ugly problem to
tackle.
--
Jeff Layton <jlayton@xxxxxxxxxx>
