Re: [PATCH] rxrpc_find_service_conn_rcu: use read_seqbegin() rather than read_seqbegin_or_lock()

[Oleg Nesterov]

On 11/16, David Howells wrote:
>
> Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
>
> > So do you agree that
> >
> > 	- the usage of read_seqbegin_or_lock/need_seqretry in
> > 	  this code makes no sense because read_seqlock_excl()
> > 	  is not possible
>
> Not exactly.  I think it should take a lock on the second pass.

OK, then how about the patch below?

Again, I'd prefer to change the semantics/prototype of need_seqretry()
to enforce the locking on the 2nd pass "automatically", but a) this
needs more discussion and b) I can't do this before I update the users
which use read_seqbegin_or_lock/need_seqretry incorrectly. So lets
discuss this later.

Oleg.

--- a/net/rxrpc/conn_service.c
+++ b/net/rxrpc/conn_service.c
@@ -25,7 +25,7 @@ struct rxrpc_connection *rxrpc_find_service_conn_rcu(struct rxrpc_peer *peer,
 	struct rxrpc_conn_proto k;
 	struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
 	struct rb_node *p;
-	unsigned int seq = 0;
+	unsigned int seq = 1;
 
 	k.epoch	= sp->hdr.epoch;
 	k.cid	= sp->hdr.cid & RXRPC_CIDMASK;
@@ -35,6 +35,7 @@ struct rxrpc_connection *rxrpc_find_service_conn_rcu(struct rxrpc_peer *peer,
 		 * under just the RCU read lock, so we have to check for
 		 * changes.
 		 */
+		seq++; /* 2 on the 1st/lockless path, otherwise odd */
 		read_seqbegin_or_lock(&peer->service_conn_lock, &seq);
 
 		p = rcu_dereference_raw(peer->service_conns.rb_node);