[asterisk-dev] [Code Review] Fix transcode_via_sln and increase utility of PLC

Mon May 10 13:49:59 CDT 2010

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviewboard.asterisk.org/r/622/
-----------------------------------------------------------

(Updated 2010-05-10 13:49:59.618633)

Review request for Asterisk Developers.

Changes
-------

Two main changes:

1. Only call plc_rx when we do not execute a plc_fillin. After looking more closely at how PLC is accomplished in translate.c, I found that this is how things operate.
2. Call plc_rx using the frame's data instead of the plc samples buffer.

I still see the memcpy memory overlap error from valgrind, but I have now confirmed that the same issue exists with the old PLC implementation, so I'm not introducing any new issues with this code change.

Summary
-------

The problem here is a bit complex, so try to bear with me...

It was noticed by a Digium customer that generic PLC (as configured in
codecs.conf) did not appear to actually be having any sort of benefit when
packet loss was introduced on an RTP stream. I reproduced this issue myself
by streaming a file across an RTP stream and dropping approx. 5% of the
RTP packets. I saw no real difference between when PLC was enabled or disabled
when using wireshark to analyze the RTP streams.

After analyzing what was going on, it became clear that one of the problems
faced was that when running my tests, the translation paths were being set
up in such a way that PLC could not possibly work as expected. To illustrate,
if packets are lost on channel A's read stream, then we expect that PLC will
be applied to channel B's write stream. The problem is that generic PLC can
only be done when there is a translation path that moves from some codec to
SLINEAR. When I would run my tests, I found that every single time, read
and write translation paths would be set up on channel A instead of channel
B. There appeared to be no real way to predict which channel the translation
paths would be set up on.

This is where Kevin swooped in to let me know about the transcode_via_sln
option in asterisk.conf. It is supposed to work by placing a read translation
path on both channels from the channel's rawreadformat to SLINEAR. It also
will place a write translation path on both channels from SLINEAR to the
channel's rawwriteformat. Using this option allows one to predictably set up
translation paths on all channels. There are two problems with this, though.
First and foremost, the transcode_via_sln option did not appear to be working
properly when I was placing a SIP call between two endpoints which did not
share any common formats. Second, even if this option were to work, for PLC
to be applied, there had to be a write translation path that would go from
some format to SLINEAR. It would not work properly if the starting format
of translation was SLINEAR.

The one-line change presented in this review request in chan_sip.c fixed the
first issue for me. The problem was that in sip_request_call, the
jointcapability of the outbound channel was being set to the format passed to
sip_request_call. This is nativeformats of the inbound channel. Because of this,
when ast_channel_make_compatible was called by app_dial, both channels already
had compatibly read and write formats. Thus, no translation path was set up at
the time. My change is to set the jointcapability of the sip_pvt created during
sip_request_call to the intersection of the inbound channel's nativeformats and
the configured peer capability that we determined during the earlier call to
create_addr. Doing this got the translation paths set up as expected when using
transcode_via_sln.

The changes presented in channel.c fixed the second issue for me. First and
foremost, when Asterisk is started, we'll read codecs.conf to see the value of
the genericplc option. If this option is set, and ast_write is called for a
frame with no data, then we will attempt to fill in the missing samples for
the frame. The implementation uses a channel datastore for maintaining the
PLC state and for creating a buffer to store PLC samples in. Even when we
receive a frame with data, we'll call plc_rx so that the PLC state will have
knowledge of the previous voice frame, which it can use as a basis for when
it comes time to actually do a PLC fill-in.

So, reviewers, now I ask for your help. First off, there's the one line change
in chan_sip that I have put in. Is it right? By my logic it seems correct, but
I'm sure someone can tell me why it is not going to work. This is probably the
change I'm least concerned about, though. What concerns me much more is the
set of changes in channel.c. First off, am I even doing it right? When I run
tests, I can clearly see that when PLC is activated, I see a significant increase
in RTP traffic where I would expect it to be. However, in my humble opinion, the
audio sounds kind of crappy whenever the PLC fill-in is done. It sounds worse to
me than when no PLC is used at all. I need someone to review the logic I have used
to be sure that I'm not misusing anything. As far as I can see my pointer arithmetic
is correct, and my use of AST_FRIENDLY_OFFSET should be correct as well, but I'm
sure someone can point out somewhere where I've done something incorrectly.

As I was writing this review request up, I decided to give the code a test run under
valgrind, and I find that for some reason, calls to plc_rx are causing some invalid
reads. Apparently I'm reading past the end of a buffer somehow. I'll have to dig around
a bit to see why that is the case. If it's obvious to someone reviewing, speak up!

Finally, I have one other proposal that is not reflected in my code review. Since
without transcode_via_sln set, one cannot predict or control where a translation
path will be up, it seems to me that the current practice of using PLC only when
transcoding to SLINEAR is not useful. I recommend that once it has been determined
that the method used in this code review is correct and works as expected, then
the code in translate.c that invokes PLC should be removed.

Diffs (updated)
-----

  /trunk/channels/chan_sip.c 262149 
  /trunk/main/channel.c 262149 

Diff: https://reviewboard.asterisk.org/r/622/diff

Testing
-------

I ran a test where I would place a call from my Polycom phone through one Asterisk box
to a second one. The second box would play back the demo-congrats file. I modified the
RTP code in the second box to drop approximately 5 percent of the outgoing packets. The
audio between the two Asterisk boxes was GSM. The audio between the Polycom phone and
Asterisk was ulaw.

I used wireshark on the first Asterisk box to analyze the RTP streams. I noticed an increase
from an average of 2650 packets on the path from the first Asterisk box to the Polycom to
an average of 2786 packets when PLC was enabled.

Thanks,

Mark