> On 20. May 2021, at 04:05, Xin Long <lucien.xin@xxxxxxxxx> wrote: > > On Wed, May 19, 2021 at 6:24 PM <mleitner@xxxxxxxxxx> wrote: >> >> On Tue, May 18, 2021 at 09:19:21PM +0200, Michael Tuexen wrote: >>> >>> >>>> On 18. May 2021, at 20:33, Xin Long <lucien.xin@xxxxxxxxx> wrote: >>>> >>>> On Tue, May 18, 2021 at 1:38 PM Michael Tuexen <tuexen@xxxxxxxxxxx> wrote: >>>>> >>>>>> On 18. May 2021, at 18:43, Xin Long <lucien.xin@xxxxxxxxx> wrote: >>>>>> >>>>>> Hi, Michael, >>>>>> >>>>>> We're implementing RFC8899 (PLPMTUD) on Linux SCTP recently, >>>>>> and to make this be controlled by setsockopt with >>>>>> SCTP_PEER_ADDR_PARAMS, as in >>>>>> >>>>>> https://datatracker.ietf.org/doc/html/rfc6458#section-8.1.12: >>>>>> >>>>>> we need another two flags to add for spp_flags: >>>>>> >>>>>> SPP_PLPMTUD_ENABLE >>>>>> SPP_PLPMTUD_DISABLE >>>>>> >>>>>> Do you think it makes sense? if yes, does the RFC6458 need to update? >>>>>> if not, do you have a better suggestion for it? >>>>> It is great new that you want to implement RFC 8899. I plan to do the >>>>> same for the FreeBSD stack. >>>>> >>>>> In my view, RFC 8899 is the right way to implement PMTU discovery. >>>>> So I will just use the SPP_PMTUD_ENABLE and SPP_PMTUD_DISABLE. I don't >>>>> think that the user needs to control which method is used. >>>>> I you want to support multiple versions, I would make that >>>>> controllable via a sysctl variable. But I think for FreeBSD, support >>>>> for RFC 8899 will be the only way of doing PMTU discovery. There >>>>> might be multiple choices for details like how to do the searching, >>>>> how long to wait for some events. These will be controllable via >>>>> sysctl. >>>>> >>>>> So in my view, there is no need to extend the socket API. What do you think? >>>> OK, that makes sense to me. >>>> >>>> Another thing I want to know your opinion on is: do you think the HB >>>> should be created >>>> separately for PLPMTUD probe, instead of reusing the old HB that >>>> checks the link connectivity? >>> Yes. I think testing for connectivity is conceptually different >>> from testing a particular PMTU. When testing for PMTU, I think >>> about sending probe packets. Not that they consist of a HB chunk >>> bundled with a PAD chunk. >>>> As the HB for PLPMTUD probe might get lost, which we don't want to >>>> affect the link's >>>> connectivity. >>> Yes, I agree completely. >> >> With this, Xin, seems we should have a separate timer for the >> PROBE_TIMER, other than the heartbeat one. >> >> Otherwise, converging the two logics into one single timer is not >> worth the hassle for saving a timer. For example, we would have to >> have it fire on the active transport but to send only the probe. >> Also, considering they can and (AFAIU the RFC) should have different >> expire timeouts from time to time. >> >> With a separate timer, we won't have issues converging the >> user-selectable heartbeat interval to the recommended 600s >> PMTU_RAISE_TIMER, for example. >> >> Maybe I am missing something. But it seems the hassle for reusing the >> timer here is just not worth it. Thoughts? > You're right, when I was doing testing, I had to set the hb_interval to 5s. > the common value of hb_interval really doesn't fit in the probe_interval. > I will try adding a new timer, but at the same time a probe_interval > by sysctl/sockopt may be needed, or we can use a value equal to > (hb_interval / n) for it? Yeah, this is why I did not write a socket API section for RFC 8899. I think when implementing it, one will find some parameters which can or should be controlled by the application. So I guess there will be more than this one. Timo (CCed) is working on some algorithmic explorations, since RFC 8899, gives you some freedom. Right now, his work is done in a simulation environment, but once that is finished and he came to some conclusions, we will implement this. So what about: * Try to implement PLPMTUD and figure out what you need as user controllable parameters. * Bring this up for discussion and we can agree on them and also on how to control them via socket options. * Then you get the code into the Linux tree. That way we would know what parameters are really needed in an implementation and we would have a common interface. Does that make sense to you and sounds like an acceptable plan? Best regards Michael > >> >> Best, >> Marcelo >> >>> >>> Best regards >>> Michael >>>> >>>>> >>>>> Best regards >>>>> Michael >>>>>> >>>>>> Thanks. >>>>> >>> >>
