Re: [PATCHv2 net-next 3/5] sctp: add SCTP_EXPOSE_POTENTIALLY_FAILED_STATE sockopt

[Neil Horman <nhorman@tuxdriver.com>]

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@tuxdriver.com> 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@aculab.com> 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.

Neil

> >
> > What if you added a check in get_peer_addr_info to only return -EACCESS
> > if pf_expose is 0 and the application isn't subscribed to the PF event?
> >
> > Neil
> >
> > > >
> > > >         David
> > > >
> > > > -
> > > > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> > > > Registration No: 1397386 (Wales)
> > > >
> > >
>