Re: [webfinger] [apps-discuss] Mail client configuration via WebFinger

["Paul E. Jones" <[e-mail]>]

Marten,

I'll comment on each point:

General:
While I supported the original work a few years ago, it was killed 
partly because some people stood up and said "we already have several 
discover protocols, why do we need another.  And you were presented with 
much the same on this list when people asked why the DNS mechanism is 
insufficient.  You provided some good arguments, of which I think 
individualized configuration is most important.  Let's make sure the new 
draft is scoped to just mail configuration.  (If we want to go further 
with calendar, contacts, etc. then arguably those should be separate 
URLs.  My mail and calendar is not provided by the same provider, for 
example.)

Caching: HTTP allows one to dictate that how long cached data can live. 
I don't think we need to specify it further.  A client will pull the 
config when the user configures the client.  There's no reason to cache 
it beyond that point.  In the off chance it queries a second time, it's 
not a big deal.  It might get pulled from web cache or perhaps go back 
to the server.  It really does not matter, since this should not be 
queried over and over.  Only in the event that a config change is made, 
the client might need to re-query the config data.  I'm not sure if the 
client should automate that process or prompt the user "would you like 
me to re-validate the configuration settings?" Personally, I like the 
latter.

Certificate pinning: Why? One fetches the mail config file.  The client 
will authenticate the certificate. Is this just to catch the very rare 
case where somebody hacks DNS and gets a fake certificate?  That's 
really a broader Internet infrastructure issue that I think we should 
solve generically.  (For example, use of notaries, DNSSEC/DANE, etc. 
Personally, I use DNSSEC/DANE, but somewhere close to zero people make 
use of that data.)

Certificate Auth: I'm not sure what the problem is here, unless people 
want to use enterprise-issued certificates (i.e., those that are not 
from a public CA).  If that is the case, I would not want my mail client 
to insert those into my trusted certificate store.  So, would those be 
one-time uses?  Not very useful.  I'd say the IT department needs to 
install whatever root certs are needed.  A client should authenticate 
the certificates it sees, but we don't need more words about it other 
than to say the TLS connection must be authenticated.

Document structure:  I think keeping this simple is really best.  I 
don't think we need multiple mail providers. If the email address is 
[e-mail], then I think it's reasonable to assume example.com is 
the provider.  Yes, a user could go enter WebFinger information for some 
other provider, but the average person will not want to do that.  That's 
having the user configure the server so it can configure the client.  
Not much gained there.  I would have something that's no more complex 
than something like this:

>{
>  "incoming" :
>  [
>    {
>      "description" : "IMAP",
>      "protocol" : "imap",
>      "servers" :
>      [
>        {
>          "host" : "imap-01.example.com",
>          "port" : 143,
>          "transport" : "starttls"
>        },
>        {
>          "host" : "imap-02.example.com",
>          "port" : 143,
>          "transport" : "starttls"
>        }
>      ]
>    },
>    {
>      "description" : "POP",
>      "protocol" : "pop3",
>      "servers" :
>      [
>        {
>          "host" : "imap-01.example.com",
>          "port" : 110,
>          "transport" : "starttls"
>        },
>        {
>          "host" : "imap-02.example.com",
>          "port" : 110,
>          "transport" : "starttls"
>        }
>      ]
>    }
>  ],
>  "outgoing" :
>  [
>    {
>      "description" : "SMTP",
>      "protocol" : "smtp",
>      "servers" :
>      [
>        {
>          "host" : "smtp-01.example.com",
>          "port" : 587,
>          "transport" : "starttls"
>        },
>        {
>          "host" : "smtp-02.example.com",
>          "port" : 587,
>          "transport" : "starttls"
>        },
>        {
>          "host" : "smtp-01.example.com",
>          "port" : 465,
>          "transport" : "tls"
>        },
>        {
>          "host" : "smtp-02.example.com",
>          "port" : 465,
>          "transport" : "tls"
>        }
>      ]
>    }
>  ]
>}

I use a arrays to offer different protocols (POP3 and IMAP) with 
descriptions to let the user choose.  If there is only one choice (as in 
the case of SMTP), then the user need not be bothered with selection. 
There are multiple servers, each intended to be an alternative listed in 
order of preference.

I would define a document for each service that might be made available 
through WebFinger (email config, calendar config, contacts config).  
And, each of these documents would be discovered via a different URI.  
So, in the JRD returned via the WebFinger query, there might be this a 
"rel" of type "email" with a URI and another of type "contacts", 
"calendar", etc.  (I did not check to see which rel values are used, but 
the names are not important.)

When a user enters his or her email address, the client can discover 
which of those services are available and offer them, allowing the user 
to use or not use each one.  In my case, calendaring is not provided at 
my domain.  So, I would go to my email program and say "add an account" 
and enter a new email address.  Since each set of services is 
individually selectable by the user, it's easy to mix and match email, 
calendaring, contacts, or whatever from different places.  It will 
require entering different account credentials, but I think that should 
be the way it works.

TLS ought to be required, but it could be left to the admin.  WebFinger 
requires it, but I seriously do not see the security concerns with a 
simple mail config file like the above.  But, if people want to secure 
it, they can.  And security should be through whatever mechanisms HTTP 
presents to the user.  I would not want to invent new ones or try to 
solve how to do something hypothetically.  My mail server requires a 
password and I think it's reasonable the HTTP server require the same.  
Perhaps my calendar requires OAuth.  We could put the URI to the 
authorization server as a property in the WebFinger reply, like this:

>{
>  "rel" : "mail-config",
>  "href" : "https://mail-config.example.com/paulej",
>  "properties" : {
>    "urn:ietf:params:oauth:authserver" : 
>"https://oauth.example.com/paulej"
>  }
>}

That might be insufficient for OAuth, since just having the URI to the 
auth server isn't enough for the Auth server, I think, as it would not 
necessarily understand the "client_id" parameter.  However, I'm not a 
guru in OAuth, so perhaps others can suggest improvements here.   The 
intent is for the client to recognize that it needs to do OAuth 
authentication before attempting to query the resource to get the actual 
configuration.  I assume we could use the same OAuth token for both 
accessing the config resource and with the SMTP and IMAP servers. (Those 
could be different credentials, but I would really hope we do not ask 
users to authenticate multiple times for a single service.)

Paul

------ Original Message ------
From: "Marten Gajda" <[e-mail]>
To: "[e-mail]" <[e-mail]>
Sent: 7/13/2016 6:32:01 PM
Subject: Re: [webfinger] [apps-discuss] Mail client configuration via 
WebFinger

>Hi Paul,
>
>the "current draft" is still the one at 
>https://tools.ietf.org/html/draft-daboo-aggregated-service-discovery-03 
>and the issues at GitHub are discussion items for the upcoming draft.
>I wanted to clarify a few points before I start to work on the draft. 
>If you have anything to add to these points or if you have any other 
>concerns, please let me know, so we can address them.
>Unless there is some more interest in the certificate stuff, I'm not 
>going to add this to the draft. We still can add that later on if it 
>turns out to be useful.
>
>Cheers,
>
>Marten
>
>
>Am 11.07.2016 um 23:31 schrieb Paul E. Jones:
>>Marten,
>>
>>Sorry to just be coming back to this after a whole month passed.
>>
>>What current draft?  Did you write one that I missed?  Or are these 
>>requirements for the draft you would like to see?
>>
>>Paul
>>
>>------ Original Message ------
>>From: "Marten Gajda" <[e-mail]>
>>To: "[e-mail]" <[e-mail]>
>>Sent: 6/8/2016 3:35:05 PM
>>Subject: Re: [webfinger] [apps-discuss] Mail client configuration via 
>>WebFinger
>>
>>>All,
>>>
>>>I've created a GitHub repository to track open issues with the 
>>>current draft, see https://github.com/CalConnect/AUTODISCOVERY/issues
>>>You're welcome to contribute to the discussion, either by creating a 
>>>new issue or by commenting on an existing one.
>>>
>>>Thanks,
>>>
>>>Marten
>>>
>>>Am 05.06.2016 um 23:00 schrieb Marten Gajda:
>>>>I think OpenID Connect already covers the discovery of everything 
>>>>you need to do OAuth2. That involves Webfinger, but there is no need 
>>>>to protect this request, because it doesn't contain personal 
>>>>information.
>>>>So we don't need to reinvent the OAuth2 bootstrapping sequence.
>>>>The only minor issue I see is that you may have to ask the user 
>>>>twice to grant access. Once to authorize the client to access the 
>>>>configuration document and another time to authorize the client to 
>>>>access the individual services. The second step is necessary, 
>>>>because the scope tokens of these services are not known when the 
>>>>first authorization request is presented to the user. The problem 
>>>>with that is, there doesn't seem to be a mechanism to broaden scope, 
>>>>without allowing the user to switch to a different account. To get 
>>>>access to the individual services, the client needs to start another 
>>>>authorization code grant. But the user is always free to log out and 
>>>>log in with a different account before granting access.
>>>>
>>>>Marten
>>>>
>>>>Am 03.06.2016 um 20:13 schrieb George Fletcher:
>>>>>Did a quick scan of the draft document. Given that more and more 
>>>>>systems are trying to remove the need for passwords, it seems like 
>>>>>the final solution needs to support 2FA and biometric mechanisms 
>>>>>for "HTTP authentication". I definitely would not want the 
>>>>>webfinger instance releasing my config data without my 
>>>>>"authentication". I suppose OAuth2 could be used to protect the 
>>>>>webfinger APIs though there is a bit of a chicken-and-egg here:)
>>>>>
>>>>>I kind of like the suggestion around returning a 401 with data in 
>>>>>the WWW-Authenticate header on where to get a token to use with the 
>>>>>API. This would allow the client to start and OAuth2 flow with the 
>>>>>Authorization Server specified and that would give the user a clear 
>>>>>indication of what password/credentials are being requested. Once 
>>>>>the client gets the token, it uses it with the webfinger call and 
>>>>>now the service-configuration data is returned because the token is 
>>>>>the authorization for the specified id.
>>>>>
>>>>>Thanks,
>>>>>George
>>>>>
>>>>>On 6/3/16 10:44 AM, Marten Gajda wrote:
>>>>>>Note that the idea behind 
>>>>>>https://tools.ietf.org/html/draft-daboo-aggregated-service-discovery-03 
>>>>>>is to put the service configuration for all services of a provider 
>>>>>>into a single document.
>>>>>>
>>>>>>So you would receive something like this:
>>>>>>
>>>>>>{
>>>>>>     "rel " :  "service-configuration",
>>>>>>     "href " : "https://example.com/service-configuration"
>>>>>>}
>>>>>>
>>>>>>If a user uses the same account identifier at another provider the 
>>>>>>Webfinger request could return their configuration too (given 
>>>>>>there is some mechanism to add it and the provider actually 
>>>>>>supports that).
>>>>>>
>>>>>>{
>>>>>>     "rel " :  "service-configuration",
>>>>>>     "href " : "https://example.com/service-configuration"
>>>>>>},
>>>>>>{
>>>>>>     "rel " :  "service-configuration",
>>>>>>     "href " : "https://acme-calendar.com/service-configuration"
>>>>>>}
>>>>>>
>>>>>>Without that it would be more difficult to setup your account at 
>>>>>>acme-calendar.com with your login "[e-mail]". 
>>>>>>acme-calendar.com would have to issue some kind of user-identifier 
>>>>>>like [e-mail] for auto-configuration purposes, even 
>>>>>>though you don't use it for authentication (because you use 
>>>>>>[e-mail] for authentication). I think that's the idea 
>>>>>>behind the `acct:` URI scheme, but I don't think that you can 
>>>>>>explain to users why they need another user identifier and when to 
>>>>>>use one or the other.
>>>>>>But that also raises the privacy concerns I mentioned earlier. If 
>>>>>>the request is not authenticated, everyone could see that you have 
>>>>>>an account at acme-calendar.com.
>>>>>>
>>>>>>Regarding SSO: There is an RFC that extends SASL based 
>>>>>>authentication to support the token authentication mechanisms as 
>>>>>>used by OAuth1 and OAuth2, see https://tools.ietf.org/html/rfc7628
>>>>>>So SSO already works with IMAP, POP3 and SMTP.
>>>>>>
>>>>>>Cheers
>>>>>>
>>>>>>Marten
>>>>>>
>>>>>>
>>>>>>
>>>>>>Am 03.06.2016 um 15:40 schrieb Paul E. Jones:
>>>>>>>Marten,
>>>>>>>
>>>>>>>
>>>>>>>>So how would the UI work?
>>>>>>>>
>>>>>>>>1) User enters user identifier, most likely an email address, 
>>>>>>>>like mailto:[e-mail] and hits "next"
>>>>>>>>
>>>>>>>>2) Client sends a Webfinger request to 
>>>>>>>>https://exampe.com/.well-known/webfinger?... to see if there is 
>>>>>>>>a configuration document
>>>>>>>>
>>>>>>>>   -> response 404 Not Found
>>>>>>>>    a) Client falls back to "manual setup" or another 
>>>>>>>>auto-configuration mechanism
>>>>>>>>
>>>>>>>>   -> response 401 Unauthorized
>>>>>>>
>>>>>>>One should not get 401 querying the WebFinger information for the 
>>>>>>>user.  What should happen is that the server should return a JSON 
>>>>>>>object that contains a link relation that might look like this:
>>>>>>>{
>>>>>>>     "rel " :  "mail-config",
>>>>>>>     "href " : 
>>>>>>>"https://mailserver.example.com/config?user=paulej%40packetizer.com"
>>>>>>>}
>>>>>>>
>>>>>>>Or something like that.  The mail client should query that URI it 
>>>>>>>is that one that should result in a potential 401.  The JSON 
>>>>>>>format that would come back here would need to be something we 
>>>>>>>define.  It could be based on JRD, but would not have to be.
>>>>>>>
>>>>>>>Otherwise, I think the flow looks right.
>>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>>    b) Client asks for password at "example.com" and goes back to 
>>>>>>>>step 2 (this time with authentication)
>>>>>>>>
>>>>>>>>   -> response 200 Ok
>>>>>>>>    c) client moves on to next step
>>>>>>>>
>>>>>>>>3) (optional) Client presents found services to the user to let 
>>>>>>>>him select the services to set up
>>>>>>>>
>>>>>>>>4) Client runs the setup handler for each selected service
>>>>>>>>
>>>>>>>>
>>>>>>>>I think in general that could work but it raises two questions:
>>>>>>>>
>>>>>>>>1) How to make sure the user really understands that he's 
>>>>>>>>authenticating at example.com in step 2b? If the user tries to 
>>>>>>>>configure calendar sync with "acme-calendar.com" where his login 
>>>>>>>>happens to be mailto:[e-mail] he might not be prepared 
>>>>>>>>to being asked for his example.com password. Maybe I'm just 
>>>>>>>>paranoid or overcautious, but I think that it might easily 
>>>>>>>>happen that the users sends his acme-calendar.com password to 
>>>>>>>>example.com in that case (since Basic authentication is still 
>>>>>>>>the most common mechanism, the client basically sends the 
>>>>>>>>password in plain text).
>>>>>>>
>>>>>>>Yeah, that's a valid concern.  The only thing I can suggest is 
>>>>>>>that the Subject CN from the certificate is presented to the 
>>>>>>>user.  Alternatively, there could be two passwords: one that is 
>>>>>>>the "configuration password" and one that is the email password.  
>>>>>>>However, I don't think two passwords will help us.  If I want to 
>>>>>>>hack somebody and can gain access to their WF config, I can 
>>>>>>>auto-config their email client to point to my own IMAP server and 
>>>>>>>just retrieve the password that way.
>>>>>>>
>>>>>>>So, I think we have to rely on the certificate presented to the 
>>>>>>>mail client and the user will have to know to trust it.  The mail 
>>>>>>>provider can tell the user "when configuring email, ensure that 
>>>>>>>the configuration server is mailconfig.example.com and do not 
>>>>>>>provide a password if that is not the name of the configuration 
>>>>>>>server indicated."
>>>>>>>
>>>>>>>If the mail config information is not protected with a password, 
>>>>>>>we probably would want the customer to verify that the SMTP 
>>>>>>>server information is correct before the password is provided.  
>>>>>>>These would be the steps users would take if performing manual 
>>>>>>>configuration, but the software presents that information to the 
>>>>>>>user without the user having to enter it.
>>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>>2) How does the client know which credentials to use to set up 
>>>>>>>>the individual services in step 4? Should the client ask the 
>>>>>>>>user for the credentials for each service or can it assume that 
>>>>>>>>some of them share the same credentials? Is that something an 
>>>>>>>>auto-configuration document can help with?
>>>>>>>
>>>>>>>It would be nice if there is a config indicator that says "SMTP 
>>>>>>>server and IMAP server passwords are the same", so the client 
>>>>>>>does not have to ask.
>>>>>>>
>>>>>>>If we use a "config password" then we could even have the server 
>>>>>>>tell the client the password for services, which would 
>>>>>>>transparently allow those to be different.  Alternatively, the 
>>>>>>>client will have to ask each about each one.
>>>>>>>
>>>>>>>For calendaring services, then yes: the client would have to ask 
>>>>>>>the user.  It could ask if it's the same or different than the 
>>>>>>>email password or the config password.  For some services, the 
>>>>>>>authentication mechanisms will be entirely different (like Google 
>>>>>>>Calendar).  The mail client will just have to know how to access 
>>>>>>>the service.  For that reason, I'm a little hesitant to suggest 
>>>>>>>including calendaring service config along with the email config 
>>>>>>>data.  We could have each of those services listed in the users' 
>>>>>>>WebFinger document.  For example, I might have this entry in my 
>>>>>>>WF document:
>>>>>>>{
>>>>>>>     "rel" : "calendar"
>>>>>>>     "href" : "urn:whatever:google"
>>>>>>>}
>>>>>>>
>>>>>>>Note I did not provide a URL.  The reason is that this is an 
>>>>>>>entirely different service that has an entirely different access 
>>>>>>>method.  Maybe the client can ask the user "is 
>>>>>>>[e-mail] our user ID for your Google calendar?"  In 
>>>>>>>my case, I don't think it is.  Certainly, it's not my gmail ID.  
>>>>>>>And, I would not want to advertise that to the world, 
>>>>>>>necessarily.  Anyway, I think we should limit the scope of what 
>>>>>>>we try to do to things that are standard OR we define a generic 
>>>>>>>URN that a client will just have to know how to deal with.
>>>>>>>
>>>>>>>
>>>>>>>>Ideally the server supports some kind of SSO mechanism like 
>>>>>>>>OpenID Connect, so you don't need to enter your password 
>>>>>>>>multiple times. But a working auto-configuration is really the 
>>>>>>>>precondition for this, because an OpenID Connect implementations 
>>>>>>>>needs a way to discover the OAuth2 scope tokens to request from 
>>>>>>>>the server and auto-configuration is really the way to do that, 
>>>>>>>>IMO. For this it would be helpful to have some mechanism to 
>>>>>>>>request a broader scope from the user (without allowing him to 
>>>>>>>>switch to a different account), but that's a different topic I 
>>>>>>>>guess.
>>>>>>>
>>>>>>>I definitely like the idea of SSO.  But, I struggle to see how we 
>>>>>>>would use this in practice with mail autoconfig since SMTP, IMAP, 
>>>>>>>and POP servers require a password, anyway.  If we use that as a 
>>>>>>>means to have those passwords provided behind the scenes (as I 
>>>>>>>indicated above), that might be a good argument for using OpenID 
>>>>>>>Connect.  In that way, the ISP can also ensure that passwords are 
>>>>>>>REALLY complex and unknown even to the user.  Not a bad practice, 
>>>>>>>in that we can view those passwords as complex tokens.
>>>>>>>
>>>>>>>Would that kind of use of OpenID Connect to retrieve the password 
>>>>>>>be workable?  (I'll admit I don't have so much experience with 
>>>>>>>OpenID Connect.  I implemented OpenID 2.0, but that's not ideal 
>>>>>>>for what we'd want to accomplish here.  I don't have a good feel 
>>>>>>>for how Connect might make this better.)
>>>>>>>
>>>>>>>Paul
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>_______________________________________________ webfinger mailing 
>>>>>>>list 
>>>>>>>[e-mail]://www.ietf.org/mailman/listinfo/webfinger
>>>>>>
>>>>>>
>>>>>>-- Marten Gajda CEO dmfs GmbH Schandauer Straße 34 01309 Dresden 
>>>>>>GERMANY phone: [phone] email: [e-mail] Managing 
>>>>>>Director: Marten Gajda Registered address: Dresden Registered No.: 
>>>>>>AG Dresden HRB 34881 VAT Reg. No.: DE303248743
>>>>>>
>>>>>>_______________________________________________ webfinger mailing 
>>>>>>list 
>>>>>>[e-mail]://www.ietf.org/mailman/listinfo/webfinger
>>>>>
>>>>
>>>>-- Marten Gajda CEO dmfs GmbH Schandauer Straße 34 01309 Dresden 
>>>>GERMANY phone: [phone] email: [e-mail] Managing 
>>>>Director: Marten Gajda Registered address: Dresden Registered No.: 
>>>>AG Dresden HRB 34881 VAT Reg. No.: DE303248743
>>>>
>>>>_______________________________________________ webfinger mailing 
>>>>list 
>>>>[e-mail]://www.ietf.org/mailman/listinfo/webfinger
>>>
>>>-- Marten Gajda CEO dmfs GmbH Schandauer Straße 34 01309 Dresden 
>>>GERMANY phone: [phone] email: [e-mail] Managing 
>>>Director: Marten Gajda Registered address: Dresden Registered No.: AG 
>>>Dresden HRB 34881 VAT Reg. No.: DE303248743
>>
>>
>>_______________________________________________ webfinger mailing list 
>>[e-mail]://www.ietf.org/mailman/listinfo/webfinger
>
>-- Marten Gajda CEO dmfs GmbH Schandauer Straße 34 01309 Dresden 
>GERMANY phone: [phone] email: [e-mail] Managing 
>Director: Marten Gajda Registered address: Dresden Registered No.: AG 
>Dresden HRB 34881 VAT Reg. No.: DE303248743