[asterisk-dev] PJSIP Device Feature Key Synchronization architecture
asterisk at phreaknet.org
asterisk at phreaknet.org
Wed Nov 9 09:37:55 CST 2022
On 10/21/2022 7:53 AM, Joshua C. Colp wrote:
> On Fri, Oct 21, 2022 at 8:29 AM <asterisk at phreaknet.org
> <mailto:asterisk at phreaknet.org>> wrote:
>
> On 10/21/2022 5:34 AM, Joshua C. Colp wrote:
> > On Thu, Oct 20, 2022 at 8:23 PM <asterisk at phreaknet.org
> <mailto:asterisk at phreaknet.org>
> > <mailto:asterisk at phreaknet.org <mailto:asterisk at phreaknet.org>>>
> wrote:
> >
> > Hi, all,
> >
> > Something I have been working on recently is adding
> support to
> > PJSIP for device feature key synchronization (the
> as-feature-event
> > Broadworks spec that many common IP phones, e.g. Polycom
> support) to
> > control server-side features from endpoints. It's using the
> PJSIP
> > pub/sub capabilities; I had to add the ability to execute a
> custom
> > module callback when a SUBSCRIBE is refreshed, but with that
> > addition,
> > it works as it's supposed to.
> >
> >
> > You should further elaborate on all of the server-side features you
> > expect to implement, if it extends beyond DND.
>
> The other one is Call Forwarding (Always/Busy/No Answer, with
> number of
> rings)
>
> >
> > I wanted to solicit some input on what an ideal way of
> triggering
> > updates should be. Currently we have hints, which actually work
> > reasonably all right for Do Not Disturb, which is simply a
> boolean
> > on/off, easily represented with a hint and custom device
> state for
> > DND.
> > The PJSIP module emits an AMI event, the user can process it and
> > change
> > the device state if needed, which will trigger a NOTIFY to
> go out
> > to the
> > endpoint.
> >
> >
> > What user? An outside AMI application? An internal consumer in
> Asterisk?
>
> The administrator of the Asterisk system, who can add AMI logic to
> receive the event and then do something with it.
> The reason this is needed is the phone isn't turning DND on directly,
> for example. It's merely a request. The server can decide to not
> allow
> it, for example if that phone isn't allowed to toggle DND. The server
> will process it, and send it the updated status. This usually
> reflects
> what the phone wanted, but not necessarily.
> (I've elaborated on this more below)
>
>
> Okay, that's not the administrator of an Asterisk system. That's a
> developer using Asterisk who may be the administrator. Those are two
> separate things. As soon as you bring AMI into the mix then it's
> making a solution using Asterisk. That's not a wrong thing, but it's
> important to be clear in the audience. Many average users of Asterisk
> doesn't know or care about AMI, 'nor do they code for it. They may use
> additional solutions that utilize AMI but they themselves haven't the
> foggiest of the details.
>
>
> >
> > This sort of came up about 12 years ago[1]. The actual SIP stuff
> > is not
> > complicated; it's the user interface to it that requires more
> > thought.
> > For call forwarding, there are more moving pieces and abusing
> > hints/custom device state for that is super clunky. You can't
> > communicate the call forwarding target, # of rings, etc. in
> a device
> > state, so additional hints are then needed for that. It
> works but
> > it's
> > super clunky and I don't think this is a great pipeline.
> >
> >
> > Okay, so this covers call forwarding as well.
> Yes.
> >
> >
> > I'm wondering if people have thoughts on what an ideal mechanism
> > would
> > be for users, once they process a request to enable/disable
> a feature
> > from the phone, to communicate that to the PJSIP module. The
> problem
> > with abusing hints, especially for call forwarding, is it's
> not a
> > good
> > way to communicate details into the module. One option
> perhaps is to
> > have dialplan extensions, setup in a manner similar to use with
> > EVAL_EXTEN, where it returns the current value needed, as any
> > relevant
> > function, DB, ODBC, CURL, custom function, etc. could be used to
> > retrieve the current feature value. The clunky part is more
> > signaling to
> > the PJSIP module that it needs to send the phone the updated
> > status (by
> > checking those extensions, for example). The device state
> callback
> > happens to be convenient for this kind of signaling but not
> really
> > appropriate here. It would be better to push the info into
> the module
> > directly rather than the signaling it and making it retrieve the
> > updated
> > data in some arbitrary way.
> >
> > So with this in mind, I'm currently leaning in the direction
> of a
> > dialplan function/AMI action that could be used to set the
> > appropriate
> > info for a subscription, which would trigger the NOTIFY, and
> then
> > nothing would actually need to be added to the dialplan at all
> > (inasmuch
> > as hints and things of that nature). One d I think starting
> purely
> > from that perspective
> >
> > isadvantage of this is that
> > for every single update, unlike callbacks, we have to
> traverse the
> > entire list of subscriptions (though maybe that's not a big
> deal).
> > The
> > bigger problem is this is push only and the PJSIP module still
> > needs to
> > be able to "pull" feature statuses on demand, which is where the
> > hint/lookup model is useful. A potential middle ground
> solution is
> > use
> > the dialplan function/AMI action to push only, but cache all of
> > this in
> > AstDB (as subscriptions themselves are), so that we can
> retrieve the
> > latest/most current value at any point if needed. Then we don't
> > need to
> > be concerned at all with how the user is managing state as
> that is
> > fully
> > decoupled, although obviously this would lead to a little
> > duplication/redundancy. Any thoughts?
> >
> >
> > You've thrown a lot of lower level implementation details at us
> and to
> > be quite honest it's overwhelming. There's no full user facing
> > examples of how it would all work, beyond the bits and pieces in
> your
> > text that we'd then have to deduce and after reading a few times it
> > doesn't feel very friendly. To start off with: Is this a
> developer API
> > and interface, or is this also meant for the common everyday
> user? I
> > would hope it's for the common everyday user, in which case it
> should
> > be approached from that perspective first with implementation
> details
> > following.
>
> Yes, it's for the common every day end user. A callback mechanism if
> used would be more of a development one but that would be more a
> means
> to an end.
>
> Here's an example of what I have in my dialplan right now, in the
> subscribe_context for the endpoint:
> exten => dnd_Polycom5,hint,Custom:${EXTEN}
> exten => callfwd_Polycom5,hint,Custom:${EXTEN}
> exten => callfwdalways_Polycom5,1,${FOOBAR(callforward,2135)}
> exten => callfwdbusy_Polycom5,1,${FOOBAR(callforwardbusy,2135)}
> exten =>
> callfwdnoanswer_Polycom5,1,${FOOBAR(callforwardnoanswer,2135)}
>
> The user gets the AMI event, processes it with whatever processing is
> needed (e.g. checking that the DND feature is available for that
> line,
> setting it in AstDB, ODBC, or whatever is the source of truth for
> feature statuses), and then updates the relevant hints above.
>
> The module is currently hardcoded to use these extensions in the
> subscribe context: the prefix + the endpoint name. Obviously
> that's also
> inflexible.
> Right now, the first 2 extensions the user will set to signal the
> module
> to send an updated NOTIFY. The first hint by virtue of being binary
> contains the DND status itself, and the 3 bottom extensions are
> needed
> to retrieve the call forwarding numbers from the source of truth for
> these features. (Here, FOOBAR is a custom - but any arbitrary -
> dialplan
> function I have that retrieves the status).
>
> I bring this up only to show the current implementation and how
> hacky it
> is; I don't like this at all or think it's appropriate (except for
> DND,
> possibly). It was more a proof of concept of testing the
> underlying SIP
> technology.
>
> A better implementation might look like:
>
> User gets the AMI event and processes it as usual, and then simply
> does
> Set(PJSIP_DEVICE_FEATURE_STATE(PJSIP/Polycom1,donotdisturb)=enabled)
> or
> Set(PJSIP_DEVICE_FEATURE_STATE(PJSIP/Polycom1,callforwardingnoanswer)=8675309,4)
>
> (forward on no answer to 8675309 after 4 rings).
>
> Internally, PJSIP_DEVICE_FEATURE_STATE could also persist input to
> AstDB
> so it's available to the module.
>
> > As an everyday user I'd expect not to have to deal with AMI or
> complex
> > dialplan. I'd expect to be able to set and get the information from
> > the dialplan using dialplan functions (or a single function) so I
> > could use that in the dialplan, and have it "just work" with my
> phones
> > that support the feature. I'd expect it to persist across Asterisk
> > restarts. For example ${EXTENSION_DND(alice)} for retrieving DND
> > status of Alice, and if Asterisk restarted then that should stay
> the same.
>
> That's a good point. The problem is that means that Asterisk
> internally
> is the source of truth of DND, and that may not necessarily be what
> people want. For instance, that wouldn't meet my own requirements.
> The
> way that Broadworks works is the phone is merely requesting a certain
> disposition, but the server isn't under any obligation to carry it
> out.
> So I think there needs to be some mechanism for the user to be
> involved
> in that pipeline, to be able to deny something that a phone wants.
> Maybe
> the user doesn't have DND, maybe certain phones aren't allowed to
> toggle, whatever. Lots of people store their feature states in MySQL
> databases and use them for Asterisk clusters. Some systems might have
> specific requirements for that. So users should have flexibility to
> reject it. On some systems, maybe the same DND status is used for
> several lines and Asterisk internally would not have any idea
> about this.
> There are lots of different scenarios that are beyond what I think
> Asterisk itself should handle, hence the "two-part" process described
> above: the user (system admin) can do whatever needs to be done, and
> then just tell PJSIP what the new state is. PJSIP doesn't need to
> know
> or care about where feature states are actually stored or what the
> logic
> is or how they are mapped to endpoints.
>
> I do see your point though and I think it would be nice to have a
> "simple, default mode" where Asterisk will internally just "approve
> everything" that the phone wants, and users can use that if that
> suits
> their needs, but fundamentally I think users should be able to be
> involved in the decision pipeline if they want/need to. I'm not
> yet sure
> what that would look like: maybe a pjsip.conf option to emit AMI
> events
> rather than auto-handling them? And then the
> PJSIP_DEVICE_FEATURE_STATE
> function would have to be used to tell PJSIP what to do, and Asterisk
> itself would not be the source of truth for feature statuses in this
> case (but it would cache them as described above) (though in this
> case
> reading the PJSIP_DEVICE_FEATURE_STATE could still return the cached
> disposition)
>
>
> Okay, YOUR usage is based on an API interface for developers to allow
> external source of truth and logic. That's what it is. My response did
> not approach it from that perspective.
>
> I don't have thoughts on that aspect currently. It requires more time.
The new architecture proposed (not using hints) has been working pretty
well for a few weeks. I'm using it in the more "developer" mode but the
other mode exists as well. The last hurdle is being able to send a
multipart XML response when necessary, since if more than one feature
needs to be sent to the device, the Broadworks spec outlines a multipart
response devices will accept.
Right now I'm using a body_generator module in order to generate the
actual content for the NOTIFY, similar to how res_pjsip_exten_state and
res_pjsip_mwi do things. However the body generator only gets a pointer
to an ast_str, so there's no way for it to do anything PJSIP related
like multipart. I was thinking that one elegant option would be to have
a separate body generator for each of the features, and then
res_pjsip_pubsub would handle creating the multipart body.
There is some stuff in res_pjsip_pubsub currently to handle multipart
subscriptions, for subscriptions that have "children". The individual
features are not "child subscriptions" in the right sense, but if they
were, then it might work within the constraints of the pubsub interface.
However, I know you suggested it would be preferable to handle any
multipart stuff in the module itself using the PJSIP multipart APIs, as
opposed to changing anything in res_pjsip_pubsub.
Looking at the existing support for multipart in res_pjsip_pubsub, it
seems this is limited to subscriptions as discussed in RFC 4662, which
require a "Supported: eventlist" header in them, even if you define a
resource list in pjsip.conf. The Broadworks device feature sync
subscriptions do not include such a header in the SUBSCRIBE. So, the
subscription in question seems to be a single, childless subscription,
that nonetheless needs to send a multipart response, potentially.
So, a couple questions, assuming my understanding of the above is correct:
* Since it doesn't appear that res_pjsip_pubsub currently supports
multipart in the way that's needed here, do you think there's any
approach we could take that retains being able to use body generator
modules? I don't see how we can, unless res_pjsip_pubsub itself
supported it. The support it has is close, but not really what it
would need to be. Would it be reasonable to extend such
functionality to support this kind of subscription? i.e. not require
"Supported: eventlist" in order to be considered a resource list?
* Alternately, if we didn't want to modify res_pjsip_pubsub to support
this unconventional usage, perhaps the module itself could directly
add some "dummy" child subscriptions to itself, one for each
feature, so that one body generator could then be used for each of
them. ast_sip_subscription is an opaque structure though, so this
would require breaking some abstraction and using the internals of
ast_sip_subscription in the module, in order to be able to add
children to it.
* Assuming we don't want to do anything with pubsub children, would it
be fine/preferred to ditch the body generator altogether and do
everything in a single module? It just feels a bit clunkier this
way, since it'll likely duplicate a lot of what res_pjsip_pubsub is
doing internally for building NOTIFYs, in fact this would likely
require bypassing much of what res_pjsip_pubsub does in order to
generate the entire response itself.
* Is it even worth trying to support multipart responses for non
resource list subscriptions? Would sending up to 4 NOTIFYs instead
of a single multipart one be an "acceptable kludge", given the
requirements of res_pjsip_pubsub?
All of the approaches listed seem a bit hacky to me, though in different
ways. Wondering if you think a particular approach is more promising
than the others...
Thanks!
NA
More information about the asterisk-dev
mailing list