the note well might have been out of date

7710bis

we need a citation for conneg (7231 maybe)

question: when the DHCP infrastructure isn't under their control, and
there was some reluctance to even attempt changing the DHCP
configuration; is there another option?

conclusion: the API has some questions about the use of a link
relation for the discovery of the API endpoint, we should take the
suggestion to the list, but there was general enthusiasm for the idea

Lorenzo noted that this is something that would allow you to
distinguish a 302 from a captive portal 302.

API

content type - no comments on the change
authentication - emile what should the name be, if it is not presented
to the user
tommy - this is just regular validation
lorenzo - the URL is not that important, we don't need to present it
to the user - confidentiality might be more important
martin had a comment that the stated reasons for using TLS aren't
quite right, this is more about removing uncertainty about the
identity of the name
dkg - we validate the name, and that allows the user to see the
identity of the server, which might be important
warren - this is often outsourced to a company with which the user has
no expectation
lorenzo - this is from an entirely untrusted source; can we restrict
API and user interaction to be the same hostname?
tommy - maybe we can present this.
lorenzo - not likely
tommy - we lose flexibility in deployment by restricting names
pierre+tommy - it's good to know that your captor and the API are in cahoots
darshak - hostname restrictions would make deployment very difficult
martin - we aren't going to prevent phishing, because there are no
established expectations at the start of this process
lorenzo - why can't we make these the same thing?  API and page could
be the same
martin - what do you believe that you are gaining?  with redirects you
can end up anywhere
dkg - if we think that this will improve access, the user might not
see the page in future connections, whatever your client does is keyed
to something; if it is the invisible API document we use, then we lose
something
kyle - we're not addressing this for just wifi, as a reminder
lorenzo - if we have a way to reconnect to the portal, then we might
need to think more about this
conclusion - we will take this (new) topic to the list

is there a process by which we would define new keys
warren - (bad idea)
lorenzo - do we need a registry?  what rules?
lorenzo - yes, we need an extension process

warren - extra edgecase - incorrect API listed, either it is down or
the information is wrong
tommy - we fall back
warren - after how long?
tommy - some time, delaying the interaction
kyle - what if the portal is a nuisance?
tommy - we might allow the connection anyway
lorenzo - if the API is advertised on a non-captive network, that's
equivalent to being logged in
tommy - yes, and also maybe useful.  the landing page might be useful
lorenzo - this might provide operators incentive for the purposes of
communicating to users
warren - this might be abused - even if that is what they already do
Eric Vyncke - yes, you can possibly show bad stuff
erik - do we need to do this?
lorenzo - yes

(darshak takes over presentation)

if the UE gets to the portal HTML page, a link relation could be used
to get back the API

warren - normal network, and then I get to a bad page with this link
relation, what then?
darshak - this is only active when you know you are captive
lorenzo - the answer is the first request - the sacrificial request
that the UE makes
warren - you might want to restrict this to the first few seconds
lorenzo - a sacrificial probe is necessary, and we would do it on that
only; in our code, that is entirely separate from the code that
ultimately displays the page
eric kinnear - this won't work if the network is suddenly captive
after not being captive
lorenzo - choice (c) creates more work; I tried to argue for (b), but
that doesn't seem to fly, so that leaves (a) "captive-api"
martin - unless you are specifically looking for this information don't use it
erik - do you need an API at the OS that you could feed this information to?
tommy - creating the link relation creates the possibility of misuse
if people miss the applicability statement
tommy - should we suggest that the probe use the Accept header we use
for the API?
warren - maybe ?  that would be a clear signal that this is a probe,
not sure if that is good or not
darshak - we probably still need more advice about deployment guidance
if we decide to do anything like this

architecture

lorenzo - a signal before closure might be used to plead for
attention; would network operators abuse this?
kyle - a snooze button
tommy - expiry bytes or time in the API don't necessarily mean that
you would prompt; maybe the OS would show a notification, which the
user could choose to ignore; the timer could be a system-level
counter; the signaling protocol could allow the network to continually
change its mind; not sure about the protocol value - how long before
closure should the notification come?
kyle - pestering is possible already, though not to that degree; the
signaling protocol might be more reliable and harder to mess up
warren - portals can already do this and some do - a prompt to look at
the API is valuable; a user will only be annoyed if the timer is low
enough to trigger the OS to do something
lorenzo - there might not be an incentive for the network to have a UE
go to the API; eyeballs are value, but the API represents cost for the
provider; it's OK for us to contact the API and update our clock; this
should always be a hint to go to the API, nothing more; that would
remove the bad incentives
dkg - think about the power structures that we are putting in place
here; no signaling protocol means that the network operator has to
make a commitment, giving network operators more power; it would be
better if the network operator has to stick to their commitments
kyle - what is the cost of visiting the API
dkg - visiting the API is a cost in itself
pierre - this looks insecure; this can be very lightweight; and it
doesn't need to be before closure; we don't need a URI in this message
warren - there is some incentive to play games maybe; the information
needs to be scoped to that network
lorenzo - there is no real cost to the UE to pull the JSON, so it's
probably OK (there is probably more cost on the server)
tommy - once connected, the UE might present information about the
negotiation, which could be used to validate the negotiated terms; an
advance signal might be complicated in some of the deployment
scenarios
lorenzo - again, why before the closure
kyle - what about when a device becomes captive
lorenzo - a portal operator said that the only reason they give away
free wifi is that they get to show a page to people, and they are very
annoyed at the UE not showing that page properly
warren - destination unreachable is a bad user experience; but having
the UE configured with a time might be better
erik - the system would provide the time warning, not necessarily the
ICMP message; why not use destination unreachable always, but instead
recommend that an OS should provide a way for applications to cause
the OS to reevaluate the status of the network
pierre - destination unreachable with a new type might help, but we
couldn't mandate it; recommend that the UE doesn't have to connect
until it believes that the negotiated term/bytes expires; maybe set a
minimum duration
lorenzo - I can hit the API every 5s; if that creates an incentive to
interrupt users, that's OK; no one is going to ask us to hit the API
unless they can use that to trigger that interruption: we can't have
nice things; the only solution is to present confirmation after
connection and expect that the contract is kept; we need to build this
with an adversarial mindset

darshak - logged in is not necessary a binary state - what if the
network wants to block netflix
lorenzo - we won't get consensus on this point
martin - we agreed not to do this
darshak - we should just address the issue in the text
warren - what about plane wifi where you can get to the videos that
are on the local server?  wording might be difficult
lorenzo - though this might be reality, this problem is hard to get
consensus on; we can only agree on a binary state, and any more
graduated access we can't agree on
tommy - we should not engage in the raging debate, but we should not
make these networks worse as a result of our design; e.g., an ICMP
message from netflix on that network shouldn't result in a new flow;
the API is a bootstrap only, we're not trying to change the nature of
the networks, which is right
lorenzo - this can look like an outage for affected services in the
partial connectivity cases, which is ok if that is expected; we can't
make this situation better
erik - we have a URL, the OS might provide a path to get to that, but that's it
kyle - maybe this could be clearer signals to apps
warren - free might be bad and you might want to renegotiate
lorenzo - we can provide that sort of affordance, easily, without
affecting this question
pierre - nothing automatic for this case
lorenzo - selective blocking can produce an awful user experience

conclusion - signal means contact the API, no more; UE can choose to
do that, or not
conclusion - document what we might do networks that selectively block
destinations; that is, nothing specific, though we believe that this
will not make the situation worse for networks that do this, we expect
those networks to communicate with users using the portal web page

backwards compatibility section - delete it

accept header field - remove that language from the architecture doc

signal authenticity
tommy - the requirements are tied to what we want from the signal; the
signal we just discussed doesn't really need any more authenticity
than what we already have; a stronger authentication requirement
implies greater weight on the signal; it creates the wrong impression
pierre - speaking in general terms, we should talk about the spoofing
properties, for instance a very easily spoofable message might be a
problem; maybe an API could whitelist source addresses for these
signals

lorenzo - could we get consensus on this doc if we removed all the
signaling text?
pierre - signaling has some value
warren - we might want to include a note in the intro
martin - we could reduce the detail in the document instead
darshak - it might be good to have some detail

conclusion - we will reduce the level of detail; kyle to look at this

pierre - volunteers an idea - we just say that the enforcement point
can send destination unreachable?
tommy - silent drops are not good, we could say that, don't blackhole packets
lorenzo - suggests that invalid TLS certs are a better idea than blackholes

lorenzo - the value we provide here is in describing a protocol that
helps clarify the contract between different devices- we can't tell
operators how to build what they have already built
warren - destination unreachable can be sticky, I had to reboot my
machine to fix a problem with a destination that was marked as
unreachable, even after I had paid for access

PvD

lorenzo - what if this URL changes?  we should address that in the architecture
tommy - this is useful work; PvD is more targetted; PvD might
supercede the other ways of discovering the API

Eric V - use null for the key.
pierre - be consistent with the proposed 7710 change (urn:ietf:params:...)
dkg - there are now N places to get this information; we're building a
swamp that is going to have failure modes that are completely
impenetrable
lorenzo - I agree, this isn't good; maybe we should only do 7710 and
the proposed HTTP header first; we could wait until PvD is mainstream
(tommy seems to agree)
lorenzo - you might be able to containerize the 7710 option in here
without creating a new path
dkg - if I get two different URIs on the same link and those have the
same value; what does it mean to register with that once?  two PvDs
with the same API endpoint, which did you register with?
pierre - the one associated with your source address
dkg - please document this
tommy - this needs a little more work; we need to specify this in the
API document: once you have interacted with the portal, you need to
check in with the API
lorenzo - login for one wouldn't necessarily mean login for both, you
need to use the network associated with the PvD; UX is *hard*
erik - PvD might be useful in extended networks (e.g., tethering)
darshak - we still need to know which wins
lorenzo - if this is part of an implicit PvD, then it won't matter
pierre - in that case, use an RA
(notes capture failed at this point)
darshak - we need a rule for who wins

conclusion - keep this individual until it matures a little

kyle - what about architecture, which references PvD
tommy - mention it in the abstract and define the role, but avoid the specifics
adam - an informational citation is fine
kyle - we should try tommy's suggestion