Re: [PATCHv2 net-next] sctp: sctp should change socket state when shutdown is received

On Tue, Jun 7, 2016 at 8:08 PM, Marcelo Ricardo Leitner
[off-list ref] wrote:

On Tue, Jun 07, 2016 at 07:03:55PM +0800, Xin Long wrote:

quoted

On Sat, Jun 4, 2016 at 8:22 PM, Marcelo Ricardo Leitner
[off-list ref] wrote:

quoted

Return error? Please don't. Adam Endrodi asked in May (linux-sctp@) a way to
return the addresses used on such attempts and currently this address
returned by accept() is the only one we can get.

[1] I've checked Adam's email, what he asked  was different case,  in
his case, the assoc
closed *after* sctp_accept, that's why he can catch the SCTP_COMM_UP event on
accept_sk but asoc is freed already.

It doesn't have to be after sctp_accept because we will queue
SCTP_COMM_UP as reactions to COOKIE_ECHO and COOKIE_ACK, and then such
event will be moved to the new socket upon sctp_accept() call (which
calls sctp_sock_migrate()).

I meant in his case, assoc close must be *after* sctp_accept.
and he said he couldn't get it as asoc was gone.

cause if assoc close is *before* sctp_accept return, assoc won't
be freed [2], he can still  get the addrs from asoc in SCTP_COMM_UP event

quoted

[2] if assoc close *before* sctp_accept return, I mean asoc close
during sctp_accept schedule out.
listen sk will be the assoc's parent sk, in sctp_cmd_delete_tcb
(SCTP_CMD_DELETE_TCB):

        if (sctp_style(sk, TCP) && sctp_sstate(sk, LISTENING) &&
            (!asoc->temp) && (sk->sk_shutdown != SHUTDOWN_MASK))
                return;

assoc can't be freed. and the event skb will be appended to
listen_sk->sk_receive_queue
when sctp_accept schedule back, it will transfer to
accept_sk->sk_receive_queue. we can
still get the closed assoc information. until we call sctp_close:
                        if (sctp_state(asoc, CLOSED)) {
                                sctp_association_free(asoc);
                                continue;
                        }

and your suggestion to improve for his case [1] is to not schedule
SCTP_CMD_DELETE_TCB
(only for tcp style). right ?

Yes

quoted

if so, in the case [2] that assoc close *before* sctp_accept return,
we should also update
the newsk->sk_state, like:

--- a/net/sctp/socket.c
+++ b/net/sctp/socket.c

@@ -7565,10 +7565,12 @@ static void sctp_sock_migrate(struct sock

*oldsk, struct sock *newsk,
        /* If the association on the newsk is already closed before accept()
         * is called, set RCV_SHUTDOWN flag.
         */
-       if (sctp_state(assoc, CLOSED) && sctp_style(newsk, TCP))
+       if (sctp_state(assoc, CLOSED) && sctp_style(newsk, TCP)) {
+               newsk->sk_state = SCTP_SS_CLOSING;
                newsk->sk_shutdown |= RCV_SHUTDOWN;
+       } else
+               newsk->sk_state = SCTP_SS_ESTABLISHED;

-       newsk->sk_state = SCTP_SS_ESTABLISHED;

so that this two cases have the similar process, and wait for
sctp_close to clean assoc.
what do you think ?

I don't agree with before/after thing, but I agree with the solution. :)

ok, I will post v3. :D

  Marcelo