Re: [netconf] [yang-doctors] Yangdoctors last call review of draft-ietf-netconf-subscribed-notifications-21

["Eric Voit (evoit)" <[e-mail]>]

> From: Per Hedeland, January 25, 2019 4:32 PM
> To: Eric Voit (evoit) <[e-mail]>; Martin Bjorklund <[e-mail]>
> Cc: [e-mail]; [e-mail];
> [e-mail]
> Subject: Re: [netconf] [yang-doctors] Yangdoctors last call review of draft-ietf-
> netconf-subscribed-notifications-21
> 
> On 2019-01-25 22:15, Eric Voit (evoit) wrote:
> >> From: Per Hedeland, January 25, 2019 1:20 PM
> >>
> >> On 2019-01-25 16:41, Eric Voit (evoit) wrote:
> >>>> From: Martin Bjorklund, January 25, 2019 10:30 AM
> >>>>
> >>>> "Eric Voit (evoit)" <[e-mail]> wrote:
> >>>>>> From: Martin Bjorklund, January 25, 2019 3:40 AM
> >>>>>>
> >>>>>> "Eric Voit (evoit)" <[e-mail]> wrote:
> >>>>>>>> "Eric Voit (evoit)" <[e-mail]> wrote:
> >>>>>>>>>> "Eric Voit (evoit)" <[e-mail]> wrote:
> >>>>>>>>>>>> From: Martin Bjorklund, January 24, 2019 9:40 AM
> >>>>>>>>>>>>
> >>>>>>>>>>>> "Eric Voit (evoit)" <[e-mail]> wrote:
> >>>>>>>>>>>>>> From: Martin Bjorklund, January 24, 2019 8:17 AM
> >>>>>>
> >>>>>> [...]
> >>>>>>
> >>>>>>>>>>>>>> Thinking some more, what is supposed to happen if the
> >>>>>>>>>>>>>> client on the same session sends first an
> >>>>>>>>>>>>>> establish-subscription with dscp 42, and then another
> >>>>>>>>>>>>>> establish-
> >>>>>> subscription with dscp 10?
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> This would be allowed.
> >>>>>>>>>>>>
> >>>>>>>>>>>> On linux at least this is a sockopt, i.e., the option
> >>>>>>>>>>>> applies to the socket, which means all packets on the
> >>>>>>>>>>>> session.  So how is this supposed to be implemented if
> >>>>>>>>>>>> different messages on the session should have different dscp
> values?
> >>>>>>>>>>>> Or is the
> >>>>>>>>>>>> idea that you send the msg, flush all data from ssh/tls to
> >>>>>>>>>>>> tcp, then flush the tcp buffers (not that easy...)?
> >>>>>>>>>>>>
> >>>>>>>>>>>> Even if there's just one establish-subscription with a dscp
> >>>>>>>>>>>> value, since it applies to the session it means that all
> >>>>>>>>>>>> normal rpcs on this session will get the same dscp value.
> >>>>>>>>>>>> It is not clear that this is the intention?
> >>>>>>>>
> >>>>>>>> Did you miss these questions?
> >>>>>>>
> >>>>>>> With NETCONF, one DSCP will apply to single TCP session.  So
> >>>>>>> there is no
> >>>>>> issue.
> >>>>>>
> >>>>>> The problem is if a client sends two establish-subscriptions with
> >>>>>> different values for dscp on the same session.  Or even if the
> >>>>>> client sends one
> >>>>>> establish-
> >>>>>> subscription with som dscp value, then the dscp will be applied
> >>>>>> to all packets from other rpcs as well.
> >>>>>>
> >>>>>> ... and even worse, with SSH channels you can have multiple
> >>>>>> NETCONF session on the same TCP session, which means that the
> >>>>>> dscp value applies to all packets in all NETCONF sessions sharing
> >>>>>> the same TCP session.
> >>>>>>
> >>>>>> I don't know what the right thing to do is.  Probably first agree
> >>>>>> that this is in fact a problem, then maybe simply document this
> >>>>>> somehow.
> >>>>>
> >>>>> I hadn't thought about multiple NETCONF SSH channels.  I agree
> >>>>> something to guide developers is needed here.  The easiest fix is
> >>>>> to recommend not supporting feature "dscp" with NETCONF.  (It
> >>>>> shouldn't be absolutely prohibited as DSCP related RFC's such as
> >>>>> RFC7657 embed text like "a single DSCP should be used for all
> >>>>> packets in a TCP
> >>>>> connection".)
> >>>>>
> >>>>> Here is the text I suggest is put into
> >>>>> draft-ietf-netconf-netconf-event-notifications, Section 5:
> >>>>>
> >>>>> "The feature "dscp" SHOULD NOT be supported over NETCONF.  This
> >>>>> will avoid the potential for out-of-order packet delivery of the
> >>>>> set of all traffic within the TCP session."
> >>>>
> >>>> I think this is a bit too limiting.  For configured subscriptions
> >>>> it can be perfectly fine, since the server can ensure that there's
> >>>> a single TCP session to the receiver in this case.  (I know we
> >>>> don't support
> >> configured subscriptions at the moment).
> >>>>
> >>>> Perhaps we can say in the SN document:
> >>>>
> >>>>     If a server that supports the "dscp" feature cannot guarantee that
> >>>>     the only packets sent on an underlying transport session are from
> >>>>     the subscription, then it should reject the subscription with a
> >>>>     "dscp-unavailable" error.  This will avoid the potential for
> >>>>     out-of-order packet delivery of the set of all traffic within the
> >>>>     TCP session.
> >>>
> >>> This works for me.    I will put this in Section 2.3, directly after the
> sentence:
> >>>
> >>>      If the publisher supports the "dscp" feature, then a subscription
> >>>      with a "dscp" leaf MUST result in a corresponding [RFC2474] DSCP
> >>>      marking being placed within the IP header of any resulting
> >>>      notification messages and subscription state change notifications.
> >>>
> >>>    unless I hear any objections on this thread.
> >>
> >> I'm not sure why the "avoid out-of-order" rationale was presented.
> >> There is no order "of the set of all traffic within the TCP session"
> >> to be out-of. Regardless of DSCP, SSH peers will send data on the
> >> channels (NETCONF sessions) in an order that depends, among other
> >> things, on the rate of consumption of that data. They will preserve
> >> order on any given channel, but will not consider order with respect
> >> to other channels (arguably, not having to do that is the whole point of
> having flow control at the SSH level).
> >
> > Hi Per,
> >
> > What you say is absolutely true.  You likely know this, but Martin asked to
> include the out-of-order rationale because if a TCP session carries traffic
> consisting of a set of DSCP values, then the network elements between the
> publisher and receiver might reorder the packets.
> 
> Hm, from the e-mail thread, it seems that the issue of reordering was
> introduced by you, and Martin's text just copied your text regarding that issue,
> while relaxing the requirement from completely forbidding the use of the DSCP
> parameter to disallowing it where it could have undesired effects.
> 
> > While the order might be restored at the receiver (per your point), this might
> also be seen as loss.  The result to the application is that higher priority dscp
> packets will be seen no faster those with lower priority.  Again you likely know
> this, but there is more on this general topic in places like RFC-7657, Section 5.2.
> >
> > To cover what Martin wants, and reduce the non-normative text, how about
> the following?...
> >
> > "Where TCP is used, a publisher which supports the "dscp" feature SHOULD
> ensure that a subscription's notification messages are returned within a single
> TCP transport session where all traffic shares the subscription's requested
> "dscp" leaf value.   Where this cannot be guaranteed, any "establish
> subscription" RPC request SHOULD be rejected with a "dscp-unavailable"
> error."
> 
> This seems fine to me. The point, as far as I can see, is that a DSCP parameter
> must not be applied to a session (or subscription) that didn't "ask" for it. This is
> enough of a rationale IMHO, no need to go into the details of what the result
> may or may not be as far as reordering of packets goes.

Thanks Per, this works for me.

Eric

> --Per
> 
> > Eric
> >
> >> I suggest simply dropping the "This will avoid ..." sentence.
> >>
> >> --Per
> >>
> >>> Eric
> >>>
> >>>> I think such text belongs to the SN document, since it may apply to
> >>>> other transports than just NETCONF.
> >>>>
> >>>>
> >>>> /martin
> >>>>
> >>>>
> >>>>
> >>>>> With this text, a RESTCONF publisher can support the feature dscp.
> >>>>> And if someone attempts to use dscp with a NETCONF subscription to
> >>>>> the same publisher, the error "dscp-unavailable" can be sent.
> >>>>>
> >>>>> Eric
> >>>>>
> >>>>>
> >>>>> ...
> >>>>>> /martin
> >>>>>
> >>>
> >>> _______________________________________________
> >>> netconf mailing list
> >>> [e-mail]
> >>> https://www.ietf.org/mailman/listinfo/netconf
> >>>
> >