On Mon, May 25, 2020 at 9:10 PM Marcelo Ricardo Leitner
[off-list ref] wrote:

On Mon, May 25, 2020 at 04:42:16PM +0800, Xin Long wrote:

quoted

On Sat, May 23, 2020 at 8:04 PM Jonas Falkevik [off-list ref] wrote:

quoted

On Tue, May 19, 2020 at 10:42 PM Marcelo Ricardo Leitner
[off-list ref] wrote:

quoted

On Fri, May 15, 2020 at 10:30:29AM +0200, Jonas Falkevik wrote:

quoted

On Wed, May 13, 2020 at 11:32 PM Marcelo Ricardo Leitner
[off-list ref] wrote:

quoted

On Wed, May 13, 2020 at 10:11:05PM +0200, Jonas Falkevik wrote:

quoted

On Wed, May 13, 2020 at 6:01 PM Marcelo Ricardo Leitner
[off-list ref] wrote:

quoted

On Wed, May 13, 2020 at 04:52:16PM +0200, Jonas Falkevik wrote:

quoted

Do not generate SCTP_ADDR_{MADE_PRIM,ADDED} events for SCTP_FUTURE_ASSOC assocs.

How did you get them?

I think one case is when receiving INIT chunk in sctp_sf_do_5_1B_init().
Here a closed association is created, sctp_make_temp_assoc().
Which is later used when calling sctp_process_init().
In sctp_process_init() one of the first things are to call
sctp_assoc_add_peer()
on the closed / temp assoc.

sctp_assoc_add_peer() are generating the SCTP_ADDR_ADDED event on the socket
for the potentially new association.

I see, thanks. The SCTP_FUTURE_ASSOC means something different. It is
for setting/getting socket options that will be used for new asocs. In
this case, it is just a coincidence that asoc_id is not set (but
initialized to 0) and SCTP_FUTURE_ASSOC is also 0.

yes, you are right, I overlooked that.

quoted

Moreso, if I didn't
miss anything, it would block valid events, such as those from
 sctp_sf_do_5_1D_ce
   sctp_process_init
because sctp_process_init will only call sctp_assoc_set_id() by its
end.

Do we want these events at this stage?
Since the association is a newly established one, have the peer address changed?
Should we enqueue these messages with sm commands instead?
And drop them if we don't have state SCTP_STATE_ESTABLISHED?

quoted

I can't see a good reason for generating any event on temp assocs. So
I'm thinking the checks on this patch should be on whether the asoc is
a temporary one instead. WDYT?

Agree, we shouldn't rely on coincidence.
Either check temp instead or the above mentioned state?

quoted

Then, considering the socket is locked, both code points should be
allocating the IDR earlier. It's expensive, yes (point being, it could
be avoided in case of other failures), but it should be generating
events with the right assoc id. Are you interested in pursuing this
fix as well?

Sure.

If we check temp status instead, we would need to allocate IDR earlier,
as you mention. So that we send the notification with correct assoc id.

But shouldn't the SCTP_COMM_UP, for a newly established association, be the
first notification event sent?
The SCTP_COMM_UP notification is enqueued later in sctp_sf_do_5_1D_ce().

The RFC doesn't mention any specific ordering for them, but it would
make sense. Reading the FreeBSD code now (which I consider a reference
implementation), it doesn't raise these notifications from
INIT_ACK/COOKIE_ECHO at all. The only trigger for SCTP_ADDR_ADDED
event is ASCONF ADD command itself. So these are extra in Linux, and
I'm afraid we got to stick with them.

Considering the error handling it already has, looks like the
reordering is feasible and welcomed. I'm thinking the temp check and
reordering is the best way forward here.

Thoughts? Neil? Xin? The assoc_id change might be considered an UAPI
breakage.

Some order is mentioned in RFC 6458 Chapter 6.1.1.

      SCTP_COMM_UP:  A new association is now ready, and data may be
         exchanged with this peer.  When an association has been
         established successfully, this notification should be the
         first one.

Oh, nice finding.

quoted

If this is true, as SCTP_COMM_UP event is always followed by state changed
to ESTABLISHED. So I'm thinking to NOT make addr events by checking the
state:

@@ -343,6 +343,9 @@ void sctp_ulpevent_nofity_peer_addr_change(struct

sctp_transport *transport,
        struct sockaddr_storage addr;
        struct sctp_ulpevent *event;

+       if (asoc->state < SCTP_STATE_ESTABLISHED)
+               return;
+
        memset(&addr, 0, sizeof(struct sockaddr_storage));
        memcpy(&addr, &transport->ipaddr, transport->af_specific->sockaddr_len);

With the above said, yep. Thanks.

quoted

It's not easy to completely do assoc_id change/event reordering/temp check.
As:

Temp check should be fine, but agree re the others. Anyhow, the above
will be good already. :-)

Hi Jonas,

What do you think? If you agree, can you please continue to go with it
after testing?

Thanks.

quoted

1. sctp_assoc_add_peer() is called in quite a few places where assoc_id is
   not set.
2. it's almost impossible to move SCTP_ADDR_ADDED from sctp_assoc_add_peer()
   after SCTP_COMM_UP.

quoted

I can make a patch with a check on temp and make COMM_UP event first.
Currently the COMM_UP event is enqueued via commands
while the SCTP_ADDR_ADDED event is enqueued directly.

sctp_add_cmd_sf(commands, SCTP_CMD_EVENT_ULP, SCTP_ULPEVENT(ev));
vs.
asoc->stream.si->enqueue_event(&asoc->ulpq, event);

Do you want me to change to use commands instead of enqueing?
Or should we enqueue the COMM_UP event directly?

-Jonas