From: Neil Horman <nhorman@xxxxxxxxxxxxx> > Sent: 14 October 2019 13:42 > To: Xin Long <lucien.xin@xxxxxxxxx> > Cc: David Laight <David.Laight@xxxxxxxxxx>; network dev <netdev@xxxxxxxxxxxxxxx>; linux-sctp@xxxxxxxxxxxxxxx; Marcelo > Ricardo Leitner <marcelo.leitner@xxxxxxxxx>; davem@xxxxxxxxxxxxx > Subject: Re: [PATCHv2 net-next 3/5] sctp: add SCTP_EXPOSE_POTENTIALLY_FAILED_STATE sockopt > > On Mon, Oct 14, 2019 at 04:36:34PM +0800, Xin Long wrote: > > On Thu, Oct 10, 2019 at 12:18 AM Neil Horman <nhorman@xxxxxxxxxxxxx> wrote: > > > > > > On Tue, Oct 08, 2019 at 11:28:32PM +0800, Xin Long wrote: > > > > On Tue, Oct 8, 2019 at 9:02 PM David Laight <David.Laight@xxxxxxxxxx> wrote: > > > > > > > > > > From: Xin Long > > > > > > Sent: 08 October 2019 12:25 > > > > > > > > > > > > This is a sockopt defined in section 7.3 of rfc7829: "Exposing > > > > > > the Potentially Failed Path State", by which users can change > > > > > > pf_expose per sock and asoc. > > > > > > > > > > If I read these patches correctly the default for this sockopt in 'enabled'. > > > > > Doesn't this mean that old application binaries will receive notifications > > > > > that they aren't expecting? > > > > > > > > > > I'd have thought that applications would be required to enable it. > > > > If we do that, sctp_getsockopt_peer_addr_info() in patch 2/5 breaks. > > > > > > > I don't think we can safely do either of these things. Older > > > applications still need to behave as they did prior to the introduction > > > of this notification, and we shouldn't allow unexpected notifications to > > > be sent. > > Hi, Neil > > > > I think about again, and also talked with QE, we think to get unexpected > > notifications shouldn't be a problem for user's applications. > > > On principle, I disagree. Regardless of what the RFC does, we shouldn't > send notifications that an application aren't subscribed to. Just > because QE doesn't think it should be a problem (and for their uses it > may well not be an issue), we can't make that general assumption. > > > RFC actually keeps adding new notifications, and a user shouldn't expect > > the specific notifications coming in some exact orders. They should just > > ignore it and wait until the ones they expect. I don't think some users > > would abort its application when getting an unexpected notification. > > > To make that assertion is to discount the purpose of the SCTP_EVENTS > sockopt entirely. the SCTP_EVENTS option is a whitelist operation, so > they expect to get what they subscribe to, and no more. > > > We should NACK patchset v3 and go with v2. What do you think? > > > No, we need to go with an option that maintains backwards compatibility > without relying on the assumption that applications will just ignore > events they didn't subscribe to. Davids example is a case in point. Although I don't enable the SCTP_PEER_ADDR_CHANGE indications. But rfc 6458 doesn't say that the list might be extended. Aren't there 3 separate items here: 1) The SCTP protocol changes (to better handle primary path failure). 2) The SCTP_GET_PEER_ADDR_INFO sockopt. 3) The MSG_NOTIFICATION indication for SCTP_ADDR_POTENTIALLY_FAILED. Looking at RFC 7829 section 7.3. 7.3 defines SCTP_EXPOSE_POTENTIALLY_FAILED_STATE. For compatibility this must default to 'disabled'. This is even true if the application has set the SCTP_PEER_ADDR_THLDS. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
