On Fri, 2020-04-17 at 12:46 -0400, Olga Kornievskaia wrote:
> On Fri, Apr 17, 2020 at 12:20 PM Trond Myklebust
> <trondmy@xxxxxxxxxxxxxxx> wrote:
> > On Fri, 2020-04-17 at 11:43 -0400, Olga Kornievskaia wrote:
> > > Hi Trond,
> > >
> > > On Fri, Apr 17, 2020 at 11:31 AM Trond Myklebust
> > > <trondmy@xxxxxxxxxxxxxxx> wrote:
> > > > Hi Olga,
> > > >
> > > > On Fri, 2020-04-17 at 11:15 -0400, Olga Kornievskaia wrote:
> > > > > When nconnect is used, SEQUENCE operation currently isn't
> > > > > bound
> > > > > to
> > > > > a particular transport. The problem is created on an idle
> > > > > mount,
> > > > > where SEQUENCE is the only operation being sent and opened
> > > > > TPC
> > > > > connections are slowly being close from the lack of use. If
> > > > > SEQUENCE
> > > > > is not assigned to the main connection, the main connection
> > > > > can
> > > > > be closed and with that so is the back channel bound to that
> > > > > connection.
> > > > >
> > > > > Since the only way client handles callback_path down is by
> > > > > sending
> > > > > BIND_CONN_TO_SESSION requesting to bind both backchannel and
> > > > > fore
> > > > > channel on the connection that was left going, but that
> > > > > connection
> > > > > was already bound to only forechannel. According to the spec,
> > > > > it's
> > > > > not allowed to change channel binding after they are done.
> > > > >
> > > > > The fix is to make sure that a lone SEQUENCE always goes on
> > > > > the
> > > > > main connection, keeping backchannel alive.
> > > > >
> > > > > Fixes: 5a0c257f8 ("NFS: send state management on a single
> > > > > connection")
> > > > > Signed-off-by: Olga Kornievskaia <kolga@xxxxxxxxxx>
> > > > > ---
> > > > > fs/nfs/nfs4proc.c | 2 +-
> > > > > 1 file changed, 1 insertion(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
> > > > > index 99e9f2e..461f85d 100644
> > > > > --- a/fs/nfs/nfs4proc.c
> > > > > +++ b/fs/nfs/nfs4proc.c
> > > > > @@ -8857,7 +8857,7 @@ static struct rpc_task
> > > > > *_nfs41_proc_sequence(struct nfs_client *clp,
> > > > > .rpc_client = clp->cl_rpcclient,
> > > > > .rpc_message = &msg,
> > > > > .callback_ops = &nfs41_sequence_ops,
> > > > > - .flags = RPC_TASK_ASYNC | RPC_TASK_TIMEOUT,
> > > > > + .flags = RPC_TASK_ASYNC | RPC_TASK_TIMEOUT |
> > > > > RPC_TASK_NO_ROUND_ROBIN,
> > > > > };
> > > > > struct rpc_task *ret;
> > > > >
> > > >
> > > > This works only in the case where the client is only sending
> > > > SEQUENCE
> > > > instructions. There are other cases where it could be sending
> > > > out
> > > > other
> > > > operations that also renew the lease, but is doing it very
> > > > infrequently. Won't that also run into the same problem?
> > >
> > > Hm... I see so main channel can still go idle and close, when
> > > infrequent operations are happening on the other connections
> > > before
> > > round-robin-ing to the main connection....
> > >
> > > > Is the fundamental problem here that we're not handling the
> > > > SEQ4_STATUS_CB_PATH_DOWN / SEQ4_STATUS_CB_PATH_DOWN_SESSION
> > > > flags
> > > > correctly or is there something else going on?
> > >
> > > Yes the client doesn't recover properly. But the fix wasn't
> > > trivial
> > > to
> > > me (so I thought my patch was enough but I see it's not). Say
> > > client
> > > shuts down the main connection because it was idle. Now whatever
> > > operations goes on a different connection is going to get
> > > callback
> > > down. The only way the client can create a new backchannel
> > > (according
> > > to the spec) is if it creates a brand new connection and sends
> > > BIND_CONN_TO_SESSION there (all existing connections are already
> > > bound
> > > to fore channel and according to the spec you can't modify the
> > > existing binding). But then we'd need to make sure that it's the
> > > first
> > > one in the list of connections we iterate thru (as i think 1st
> > > marks
> > > the main connection?) as the other operations that supposed to
> > > only
> > > go
> > > on main connection need to know which connection to pick.
> > >
> > > The reason it's not seen against linux is because it doesn't
> > > follow
> > > the spec is doesn't reject attempts to bind a backchannel to an
> > > already existing connection that was only bound for fore channel.
> > >
> >
> > Oh, I see. So the server is replying NFS4ERR_INVAL in order to let
> > the
> > client know that it is trying to change the channel bindings for
> > that
> > connection.
>
> Well server isn't failing because client is asking for FORE_OR_BOTH
> and it's a choice so server is returning FORE. I'm not sure we can
> ask
> the server to fail the request with ERR_INVAL.... (rather I can ask
> but ) rather can we expect the server to do that always?
In RFC5661, Section 18.34.3 I found the following normative text:
Invoking BIND_CONN_TO_SESSION on a connection already associated with
the specified session has no effect, and the server MUST respond with
NFS4_OK, unless the client is demanding changes to the set of
channels the connection is associated with. If so, the server MUST
return NFS4ERR_INVAL.
IOW: it sounds like your server isn't following the spec either. 🙂
> > Hmm... Is there any reason why we can't just add a handler to
> > nfs4_bind_one_conn_to_session_done() that intercepts NFS4ERR_INVAL,
> > and
> > disconnects the xprt before retrying?
> > We should probably add a wrapper to xprt_force_disconnect() in
> > include/linux/sunrpc/clnt.h. Something like the following?
> >
> >
> > static inline void rpc_task_close_connection(struct rpc_task *task)
> > {
> > if (task->tk_xprt)
> > xprt_force_disconnect(task->tk_xprt);
> > }
> >
--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@xxxxxxxxxxxxxxx
