Re: [PATCH] net: sctp: recover a tranport when an ack comes

[Neil Horman]

On Fri, Nov 15, 2013 at 09:00:58AM -0500, Vlad Yasevich wrote:
> On 11/15/2013 07:30 AM, Neil Horman wrote:
> >On Thu, Nov 14, 2013 at 09:34:55PM -0500, Vlad Yasevich wrote:
> >>On 11/14/2013 03:40 PM, Chang Xiangzhong wrote:
> >>>Expected Behavior:
> >>>When hearing an ack from a tranport/path, set its state to normal/on if it's
> >>>in abnormal(__partial_failure__ or inactive) state.
> >>>
> >>>state machine of tranport->state
> >>>Whenever a T3_RTX timer expires, then transport->error_count++.
> >>>When (association->pf_retrans < transport->error_count < tranport->pathmaxrtx)
> >>>	transport->state = SCTP_PF //partial failure
> >>>
> >>>When a heartbeat-ack comes or conventional ack acknowledged its availability,
> >>>	transport->state = SCTP_ON
> >>>
> >>>Signed-off-by: Chang Xiangzhong <changxiangzhong@xxxxxxxxx>
> >>>Fixes: 5aa93bcf66f ("sctp: Implement quick failover draft from tsvwg")
> >>
> >>I don't think this is right.  The spec states:
> >>  8.  ACKs for retransmissions do not transition a PF destination back
> >>        to Active state, since a sender cannot disambiguate whether the
> >>        ack was for the original transmission or the retransmission(s).
> >>
> >>
> >>Now, the proper way to this would would be modify
> >>sctp_assoc_control_transport() to transition the transport state to
> >>ACTIVE if it was PF transport that was chosen to send data.
> >>
> >>-vlad
> >>
> >I agree, this patch doesn't agree with the spec, the only time we transition
> >from PF to ACTIVE should be on receipt of ack of new data.
> 
> You mean HB ACK, right?  The 02 spec that see on the ietf site doesn't
> mention anything about transition on SACKs.  Also, there is no way to
> tell right now if the ack is for new or retransmitted data.  We could
> mark chunks that are retransmitted though.
> 
Yes, sorry I wasn't clear, I was speaknig about HB Acks.

> > I'm not even sure if
> >we should allow PF transports to be selected to send new data.  Currently a
> >potentially failed transport will get ignored when specified, and the stack will
> >use the active path in its place.  Only if all transports are PF will a PF
> >transport be chosen.
> 
> Not even that :(.  If all transports are PF, we are going to camp on
> the primary path instead of choosing a PF transport with the lowest
> error count.
> 
Yes, we don't do the smallest error count check, though I have to wonder if
thats really worthwhile.  If all your transports are PF, you're a step away from
a connection reset anyway.
Neil

> -vlad
> 
> >Neil
> >
> >>>---
> >>>  net/sctp/outqueue.c |    1 +
> >>>  1 file changed, 1 insertion(+)
> >>>
> >>>diff --git a/net/sctp/outqueue.c b/net/sctp/outqueue.c
> >>>index 94df758..2557fa5 100644
> >>>--- a/net/sctp/outqueue.c
> >>>+++ b/net/sctp/outqueue.c
> >>>@@ -1517,6 +1517,7 @@ static void sctp_check_transmitted(struct sctp_outq *q,
> >>>  			 * active if it is not so marked.
> >>>  			 */
> >>>  			if ((transport->state == SCTP_INACTIVE ||
> >>>+			     transport->state == SCTP_PF ||
> >>>  			     transport->state == SCTP_UNCONFIRMED) &&
> >>>  			    sctp_cmp_addr_exact(&transport->ipaddr, saddr)) {
> >>>  				sctp_assoc_control_transport(
> >>>
> >>
> >>
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-sctp" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/