Hello all -

One of the efforts that a number of developers in the community here at
Digium have been at work at are cleaning up test failures exposed by
Jenkins [1]. One of these, in particular, has been rather difficult to
resolve - namely, fax/pjsip/directmedia_reinvite_t38 [2]. This e-mail goes
over what has been accomplished, and asks some questions on how we might
try and fix Asterisk under this scenario.

The directmedia_reinvite_t38 test attempts to do the following:
 (1) UAC1 calls UAC2 through Asterisk, with audio as the media. The dial is
performed using the 'g' flag, such that UAC2 will continue on if UAC1 hangs
 (2) UAC1 and UAC2 are configured for direct media. Asterisk sends a
re-INVITE to UAC1 and UAC2 to initiate direct media.
 (3) After responding with a 200 OK to the direct media requests, UAC1
sends a re-INVITE offering T.38.
 (4) Asterisk sends an INVITE with T.38 to UAC2
 (5) UAC2 sends back a 200 OK for T.38; Asterisk sends that to UAC1.
Asterisk switches out of a direct media bridge to a core bridge.
 (6) UAC1 hangs up. Asterisk sends a re-INVITE to UAC2 for audio back to
Asterisk. UAC2 responds with a 200 OK for the audio.
 (7) Asterisk ejects UAC2 back to the dialplan.

It's important to note that this test never should have passed - an update
to the test suite "fixed" the test erroneously passing, which led to us
investigating why the scenario was failing. This test was copied over from
an identical chan_sip test, which passes.

The PJSIP stack has two issues which make life difficult for it in this
(1) The T.38 logic is implemented in res_pjsip_t38. While that is _mostly_
a very good thing - as it keeps all the fax state logic outside of the
channel driver - we are also a layer removed from interactions that occur
in the channel driver. That makes it challenging to influence direct media
checks and other Asterisk/channel interactions.
(2) Being very asynchronous, requests may be serviced that influence T.38
state while other interactions are occurring in the core. Informing the
core of what has occurred can have more race conditions than what occurs in
chan_sip, which is single threaded.

The first bug discovered when the test was investigated was an issue in
step (2). We never actually initiated a direct media re-INVITE. This was
due to res_pjsip_t38 using a frame hook, and not implementing the
.consume_cb callback. That callback allows a framehook to inform the core
(and also the bridging framework) of the types of frames that a framehook
wants to consume. If a framehook needs audio, a direct media bridge will be
explicitly denied, and - by default - the bridging framework assumes that
framehooks will want all frames. Another bug that was discovered occurred
in step (6). When UAC1 sends a BYE request, nothing informed UAC2 that the
fax had ended - instead, it was merely ejected from the bridge. This meant
that it kept its T.38 session going, and Asterisk never sent a re-INVITE to
UAC2. Both of these bugs were fixed by 726ee873a6.

Except, unfortunately, the second bug wasn't really fixed.

726ee873a6 did the "right" thing by intercepting the BYE request sent by
UAC1, and queueing up a control frame of type AST_CONTROL_T38_PARAMETERS
with a new state of AST_T38_TERMINATED. This is supposed to be passed on to
UAC2, informing it that the T.38 fax has ended, and that it should have its
media re-negotiated back to the last known state (audio) but also back to
Asterisk (since we aren't going to be in a bridge any longer).
Unfortunately, this code was insufficient.

A race condition exists in this case. On the one hand, we've just queued up
a frame on UAC1's channel to be passed into the bridge, which should get
tossed onto UAC2's channel. On the other hand, we've just told the bridging
framework to kill UAC1's channel with extreme prejudice, thereby also
terminating the bridge and ejecting UAC2 off into the dialplan. In the
first case, this is an asynchronous, message passing mechanism; in the
second case, the bridging framework inspects the channel to see if it
should be hung up on *every frame* and *immediately* starts the
hangup/shutdown procedure if it knows the channel should die. This is not
asynchronous in any way. As a result, UAC1 may be hung up and the bridge
dissolved before UAC2 ever gets its control frame from UAC1.

There were a couple of solutions to this problem that were tried:
(1) First, I tried to make sure that enqueued control frames were flushed
out of a channel and passed over the bridge when a hangup was detected. In
practice, this was incredibly cumbersome - some control frames should get
tossed, others need to be preserved. What was worse was the sheer number of
places the bridge dissolution can be triggered. While it wasn't hard to
make sure we flushed frames off an ejected channel into a bridge, it was
nigh impossible to ensure that this occurred every single time before the
other channels were ejected. Again, the bridging framework is ridiculously
- perhaps ludicrously - aggressive in tossing channels out of a bridge once
it has decided the bridge should be dissolved.
(2) Second, I tried to make the bridge ejection process asynchronous. This
was done by enqueuing another control frame onto the channel being ejected;
when it leaves, it flushes its control frames into the bridge. When the
'ejection' control frame gets passed into the bridging core, that causes
the bridge to dissolve. This worked well in some scenarios, and it also
guaranteed that the T.38 control frame would be delivered. Unfortunately,
in other cases, it caused all of the channels to hang out in the bridge ...
permanently. Again, there's a lot of edge cases in the bridging code that
deal with channels being kicked out of a bridge, and the bridge
dissolving... and it was more than I could chew on.

The long and short of it is: while Asterisk 12+ has a nice bridging
framework that hides or eliminates a lot of the horrendous
masquerade/transfer code, as well as the 'triple infinite loop' in
features/channel that existed in Asterisk 11-, it is still ridiculously
complex and prone to breaking spectacularly in subtle ways. Not to mention
both (1) and (2) end up being massive changes to the design that are risky
in an LTS (no one likes it when a channel can't be hung up.)

So those ideas were scratched.

The next solution was to try a bridge mixing technology that specifically
managed the T.38 state. This worked ... really well. Incredibly well, in
fact. It avoided all of the previous problems because, unlike external
modules or even certain places in the bridging core, a bridge technology is
guaranteed by the core to be called in a synchronized fashion when any of
the following occurs:
(1) When a bridge technology is chosen
(2) When that technology is started
(3) When that bridge has a channel added
(4) When that bridge has a channel removed
(5) When that technology is stopped
All of which covers the necessary places to know when a channel has hung
up, and gives us a place where we can safely inform the other channels
before the bridging framework starts doing mean things. bridge_t38 was the
result [3]. It managed a bit of T.38 state for the two channels in a core
bridge that were in a T.38 fax, and, when one of them leaves, it informed
the other channel that it should end its T.38 fax.

Problem solved.


Not quite.

After merging [3] in f42d22d3a1, we noticed that the masquerade test [4]
started to fail. That's a really, really bad sign. The masquerade 'super
test' was originally tested to stress test masquerades in Asterisk 1.8 and
11. It constructs a chain of 300 Local channels, then optimizes them all
down to a single pair of 'real' channels. In Asterisk 12+, masquerades were
eliminated in this scenario, but we instead have a series of incredibly
complex Local channel optimization-caused bridge/swaps/merges that kick off
as the Local channels collapse and merge their bridges down to one. It's a
great "canary in the coal mine" test, as when it fails, it almost certainly
means you've introduced a dead lock into one of the more complex operations
in Asterisk - regardless of the versions.

And lo and behold, we had.

Local channels are weird. One of the 'fun things' they do is 'help' T.38
along by passing along a channel query option for T.38 state. This lets us
do ridiculous things like make sure a T.38 fax works across a Local channel
chain (and is covered by the fax/sip/local_channel_t38_queryoption test).
Unfortunately, the bridge_t38 module had to query for T.38 state in its
compatible callback - this allowed it to determine the current state of
T.38 on the channels in the bridge to see if it needed to be activated.
Unfortunately, in a 300 Local channel chain, that means reaching across 300
bridges - simultaneously - locking bridges, bridge_channels, channels,
PVTs, and the entire world in the process. Since the bridge lock was
already held in the compatible callback, this caused a locking inversion
(no surprise there), deadlocking the whole thing.

This is not a trivial locking situation to resolve. Even if we unlock the
bridge, we're still liable to deadlock merely by trying to lock 300 bridges
simultaneously. (There may even be another bug in here, but it is hardly
worth trying to find or fix at this point.) And we can't remove the query
option code in chan_local, as T.38 faxes will no longer work across Local

As an aside, if there's a lesson in all this, it is that synchronous code
in a heavily multi-threaded environment is bad. Message passing may be
harder to write, but it is far easier to maintain.

Anyway, as a result, I've reverted the bridge_t38 module in 75c800eb28.

So what do we do now?

The crux of this problem is that the bridging framework does not have a
standard way of informing a channel when it has joined or - more
importantly - left a bridge. Direct media has its own mechanism managed by
the RTP engine - so it works around this. However, we have a number of
scenarios where "things happen" in a bridge that involves state on a
channel and - right now - we don't have a unified way of handling it. In
addition to T.38, we also have channels being put on hold, DTMF traversing
a channel, and more. Often, the channel driver has this state - but
instead, we have a lot of 'clean up' logic being added to the bridging core
to handle these situations.

As I see it, we really only have two options here:
(1) Add code to the bridging framework to clean up T.38 on a channel when
it leaves. This is kind of annoying, as it will happen on every channel
when it leaves, regardless of whether or not the channel even supports T.38.
(2) Add a new channel technology callback that a bridge can use to inform a
channel driver that it is being ejected from a bridge. This would give us a
single place to put cleanup logic that has to happen in a channel driver
when it is no longer bridged.

I'm not sure those two options will work, exactly, but it's the best
options that I can think of after exhausting lots of other code changes in
the bridging core. If someone has other suggestions, I'd be more than happy
to entertain them.


[1] https://jenkins.asterisk.org/
[3] https://gerrit.asterisk.org/#/c/1761/

Matthew Jordan
Digium, Inc. | Director of Technology
445 Jan Davis Drive NW - Huntsville, AL 35806 - USA
Check us out at: http://digium.com & http://asterisk.org
-- Bandwidth and Colocation Provided by http://www.api-digital.com --

asterisk-dev mailing list
To UNSUBSCRIBE or update options visit:

Reply via email to