freeswitch/libs/freetdm/docs/ftdm-open-issues.txt

77 lines
6.3 KiB
Plaintext

== This is an email sent to stkn responding to his post http://stkn.techmage.de/archives/200 ==
1. FreeTDM open channels immediately.
Yes, this seems to come from the early days of OpenZAP where probably the analog modules were implemented first and was dragged incorrectly to ISDN.
The open API also sets the flag in use so others cannot use it. We should decouple this open and reserve operations. Does not seem very hard to do,
I'll take a look at it on my way back to Toronto this week or earlier if possible, I'll create a git branch for it and send you the url for peer review.
Feel free to do the opposite if you are in a hurry.
2. FreeTDM does not completely close the channel on hangup.
As long as the signaling module calls ftdm_channel_close() the fio->close API will be called where the IO module is responsible
for turning off ec or any other operations to leave the channel unused/idle (not really closed). There was indeed a bug where ftdm_zt module
was not properly closing the channel, if that is what you mean, then it means that was already fixed by you with your audio mode turn off patch.
Now, a more lengthy response ...
When we, Sangoma, took over openzap and renamed it to freetdm, it was not just a name change, many API cleanups were done so API users cannot access
the internal data structures and cannot explicitly move between channel states. Also, some order on the state transitions has been put into place.
We've tried to keep backwards compatibility though, so old modules may not follow all of the conventions. Some unwritten conventions are:
1. Signaling modules are responsible for calling ftdm_channel_close() when done with the channel. In fact, signaling modules should not
call ftdm_channel_done unless they really know what they're doing. ftdm_channel_close will call ftdm_channel_done anyways.
2. When remote end hangs up a call, signaling module should go to TERMINATING state and send SIGEVENT_STOP to the user. The signaling stack
should not acknowledge the hangup until the user implicitly moves us to HANGUP (via ftdm_channel_call_hangup()). There is a core timer
that will ensure we move from TERMINATING to HANGUP if the user does not moves us fast enough (arbitrarely 3000ms), although that is considered
a bug in the user application (at some point we saw that in FreeSWITCH) and we complain loudly when that happens.
3. Signaling modules should leave the channel with NO pending states once they unlock it. This means, whenever you have a message to process
you must lock the channel, check its current state (the user may have already changed the state to something else), process any pending
state set by the user, then process your message and if required move the channel to a new state and process that state before releasing the lock.
There is a chance of a deadlock due to the callbacks we use to notify users of messages if we call those callbacks with the channel lock held and
that is why new signaling modules should use the FTDM_SPAN_USE_SIGNALS_QUEUE to defer the notification to the user when the ftdm_channel_t is not
locked calling ftdm_span_trigger_signals() at the end of their processing loop which delivers the callback notifications to users.
Having said that, there is a major architectural change coming, probably early 2011, and I'd like to have your input.
== Decoupling calls from IO channels ==
We've also found that the concept of a channel and a call are unnecesarily coupled together in FreeTDM. I've never asked Anthony the reasoning
for this, but it seems the API was meant to be used with, or without calls. That is, a user should be able to just 'open' a channel and start
reading and writing raw data without ever launching one of the provided signaling stacks on top, may be just because the user is using their
own signaling stack and just needs freetdm/openzap as an I/O API. In fact, signaling modules work like this (but there are defects there).
The problem is, as you have noticed, the way ftdm_channel_open_xx APIs has been implemented is abusing this concept.
Problems with current implementation.
1. Exclusive channel is the only way we can request channels to the ISDN network. There is no room to change the channel later on easily.
2. Multi-rate calls cannot be done. We need to associate multiple B-channels to a single call.
3. Low-bit-rate calls (not sure what is the proper name for them) where a single B-channel can feed media to multiple calls with a low bit-rate
codec (I know of at least one request from a big company for this to support some Q.SIG stuff for Eurocontrol).
What do we need?
I think there should be only one open() API that wraps calls around fio->open(), the other current open_xx functions that do hunting, should
be eliminated and some extra arguments provided to ftdm_channel_call_place to provide information about the channel where we would like to
place the call or group hunting information and the API would return immediately providing a ftdm_call_t reference pointer that can be used
to further track the status of that call, or in the case where a channel exclusivity was requested and we know for a fact is not available,
just return failure. At that point there is no I/O channel involved yet. Eventually after the signaling stack hunts the channel and
negotiaties its availability with the network, an I/O channel can be attached to that call, or even attach multiple channels (for the case
of multi rate calls?).
This involves heavy changes. The state machine should be moved out of the ftdm_channel_t structure along with probably some other members
and stuffed into a new ftdm_call_t structure. Signaling modules would have to be updated to run their state machines with this call
structure and then at some point attach IO channels to the call. There would be 2 type of read/write APIs, one low level ftdm_channel_read
which receives an ftdm_channel_t as parameter and an ftdm_call_read() which receives an ftdm_call_t as parameter, for the case of multi-rate
calls or low-bit-rate multiplexed calls in the same B-channel this API would return just the media corresponding to that call already
demultiplexed or concatenated, or whatever needs to be done with the media from the multiple low-level IO channels.
This is fairly disruptive, but I believe should be done sooner or later.