Russell Bryant [Wed, 2 Jun 2010 22:46:58 +0000 (22:46 +0000)]
Merged revisions 267352 via svnmerge from
https://origsvn.digium.com/svn/asterisk/trunk
........
r267352 | russell | 2010-06-02 17:46:37 -0500 (Wed, 02 Jun 2010) | 7 lines
try to fix some random chan_h323 compilation failures
After some debugging, the random chan_h323 build failures appear to be due
to complications introduced by some chan_h323 specific build stuff getting
triggered during a clean. Simplify this by moving the h323 clean commands
down into channels/makefile.
........
Russell Bryant [Wed, 2 Jun 2010 21:42:41 +0000 (21:42 +0000)]
Merged revisions 267303 via svnmerge from
https://origsvn.digium.com/svn/asterisk/trunk
........
r267303 | russell | 2010-06-02 16:41:54 -0500 (Wed, 02 Jun 2010) | 6 lines
Ensure the -Wno-strict-aliasing flag makes it, even if ASTCFLAGS has been specified.
When ASTCFLAGS was specified with the make command, Makefile.rules was using
the specified value from the command line and not the one here, making it so this
flag would go missing.
........
This changes the sample slinear frame data to contain non-zero data so that
translation calculations for speex works when preprocessing and VAD is turned
on. The encoder expects samples to be returned, but when attempted with the
mentioned two options and silent sample frames everything was discarded.
Eliminate stale manager events after a set interval, even if AMI clients don't query for them.
Actions (or failures to act) by external clients should not cause memory leaks
in Asterisk, especially when those continued leaks could cause Asterisk to
misbehave later.
Use sigaction for signals which should persist past the initial trigger, not signal.
If you call signal() in a Solaris signal handler, instead of just resetting
the signal handler, it causes the signal to refire, because the signal is not
marked as handled prior to the signal handler being called. This effectively
causes Solaris to immediately exceed the threadstack in recursive signal
handlers and crash.
Properly use peer's outboundproxy for outbound REGISTERs.
The logic used in transmit_register to get the outboundproxy for a peer
was flawed since this value would be overridden shortly afterwards when
create_addr was called.
In addition, this also fixes some logic used when parsing users.conf so
that the peer name is placed in the internally-generated register string
so that an outboundproxy set in the Asterisk GUI will be used for outbound
REGISTERs.
........
It is possible to connect to the manager interface before all Asterisk modules
are loaded. To ensure that an application does not send AMI actions that might
require a module that has not yet loaded, the application can listen for the
FullyBooted manager event. It will be sent upon connection if all modules have
been loaded, or as soon as loading is complete. The event:
Calendaring support for Exchange Server 2007+ via EWS
This commit adds support for calendaring with Exchange Server 2007+ via
Exchange Web Services. Full write support and for querying attendees. Many
thanks to Jan Kaláb for the feature.
Mark Michelson [Fri, 21 May 2010 21:57:24 +0000 (21:57 +0000)]
Fix memory hogging behavior of app_queue.
From reviewboard:
This review request is for the patch on issue 17081.
A user reported that he saw increasing numbers of allocations stemming
from app_queue.c when he would run the "queue show" CLI command. The
user reported that he was using approximately 40 realtime queues and
as he ran the CLI command more and more, the memory usage would shoot up.
As it turns out, there was a memory leak and a separate usage of memory
that, while not really a leak, was very irresponsible.
Both memory problems can be attributed to the function init_queue(). When
the "queue show" command is run, all realtime queues have the init_queue()
function called on the in-memory queue. The idea is to place the queue in
its default state and then overwrite options specified in the realtime backend
as we read them.
The first problem, the memory leak, had to do with the fact that the string
field for the name of the first periodic announcement file was being re-created
every time init_queue was called. This patch corrects the behavior by only
calling ast_str_create if the memory has not already been allocated.
The other problem is a bit more complicated. The majority of the strings
in the call_queue structure were changed to use the ast_string_fields API
for 1.6.0 and beyond. init_queue resets all string fields on the queue to
their default values. Then, later in the realtime queue loading process,
these string fields are set to their configured values.
For those unfamiliar with string fields, frequent resizing of a string like
this is not what the string fields API is designed for. The result of this
constant resizing is that as the queue gets loaded, eventually space for
the string runs out and so a new memory pool, at twice the size of the
previously allocated one, is created for the string fields. The reporter
of issue 17081 wrote a script that ran the "queue show" CLI command 2100
times. By the end, each of his 40 queues was taking about a megabyte of
memory apiece just for their string fields.
My fix for this problem is to revert the call_queue structure from using
string fields. In my patch here, I have moved the queue back to using
fixed-sized buffers. I ran the script provided by the reporter of 17081
and determined that I no longer saw the steadily-increasing memory usage
that I had seen before applying the patch.
Allow ast_safe_sleep to defer specific frames until after the sleep has concluded.
From reviewboard
Background:
A Digium customer discovered a somewhat odd bug. The setup is that parties A
and B are bridged, and party A places party B on hold. While party B is
listening to hold music, he mashes a bunch of DTMF. Party A takes party
B off hold while this is happening, but party B continues to hear hold
music. I could reproduce this about 1 in 5 times.
The issue:
When DTMF features are enabled and a user presses keys, the channel that
the DTMF is streamed to is placed in an ast_safe_sleep for 100 ms, the
duration of the emulated tone. If an AST_CONTROL_UNHOLD frame is read
from the channel during the sleep, the frame is dropped. Thus the
unhold indication is never made to the channel that was originally placed
on hold.
The fix:
Originally, I discussed with Kevin possible ways of fixing the specific
problem reported. However, we determined that the same type of problem
could happen in other situations where ast_safe_sleep() is used. Using
autoservice as a model, I modified ast_safe_sleep_conditional() to
defer specific frame types so they can be re-queued once the sleep has
finished. I made a common function for determining if a frame should
be deferred so that there are not two identical switch blocks to
maintain.
ast_callerid_parse() had a path that left name uninitialized.
Several callers of ast_callerid_parse() do not initialize the name
parameter before calling thus there is the potential to use an
uninitialized pointer.
........
................
Fix transcode_via_sln option with SIP calls and improve PLC usage.
From reviewboard:
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.
fixes infinite loop during udptl.c's decode_open_type
When decode_length returns the length there is a check to see if that
length is negative, if so the decode loop breaks as this means the
limit has been reached. The problem here is that length is an
unsigned int, so length can never be negative. This resulted in
an infinite loop.
During deadlock avoidance the sip dialog pvt is locked and
unlocked. When this occurs we have no guarantee the pvt's owner
is still valid. We were trying to access the pvt's owner after
this without checking to see if it still existed first.
(closes issue #17271)
Reported by: under
Patches:
check_rtp_timeout.diff uploaded by under (license 914)
Tested by: dvossel
........
Internal timing is now on by default, if you're using DAHDI 2.3 or above.
The reason for ensuring DAHDI 2.3 or above is that this version ensures that
a timer is always available, whereas in previous versions, it was possible
for DAHDI to be loaded, but have no drivers to actually generate timing. If
internal_timing was turned on in this circumstance, a complete lack of audio
would result. This is the reason why internal_timing was not on by default.
However, now that DAHDI ensures the availability of a timer, there is no
reason for this setting to be off (and in fact, it solves a great many initial
user problems).
Keep track of digit duration, when we're decoding inband to pass DTMF frames.
(closes issue #17235)
Reported by: frawd
Patches:
new_dtmf_dsp_len.patch uploaded by frawd (license 610)
20100518__issue17235.diff.txt uploaded by tilghman (license 14)
Tested by: frawd
........
During the processing of Cisco dtmf the dtmf samples were
not being calculated correctly. In an attempt to determine
what sample rate was being used, a NULL frame was processed
which caused a crash. This patch resolves this.
Modify directory name reading to be interrupted with operator or pound escape.
In the case of accidentally entering the wrong first three letters for the
reading, users could be very frustrated if the name listing is very long. This
allows interrupting the reading by pressing 0 or #. 0 will attempt to execute
a configured operator (o) extension and # will exit and proceed in the
dialplan.
Fix logic error when checking for a devstate provider.
When using strsep, if one of the list of specified separators is not found,
it is the first parameter to strsep which is now NULL, not the pointer returned
by strsep.
This issue isn't especially severe in that the worst it is likely to do is waste
some cycles when a device with no '/' and no ':' is passed to ast_device_state.
........
................
The Version field in the cookies we're setting contain quotes around the version
number which is not compatible with RFC2109 and breaks some implementations.
(closes issue #17231)
Reported by: ecarruda
Patches:
manager_rfc2109-trunk-v1.patch uploaded by ecarruda (license 559)
manager_rfc2109-1.6.2-v1.patch uploaded by ecarruda (license 559)
Tested by: ecarruda, russell
........
................
The latest version of the core sounds files 1.4.19 now includes the missing
queue-minute sound file which is called by app_queue but which has been
missing.
Marquis42 suggested a better method of doing what I wanted because I ended up
removing the WARNING message for all instances when really I just wanted to
remove it for the 'return' keyword, not everything.
We attempted to detect silence after translating a frame
from signed linear. This caused a flooding of errors. To
resolve this the code to detect silence was moved before the
translation.
As usual, the way they do things is different, so we need to account for that.
automake is versioned ala BSD/Linux, but autoconf is not. We don't actually
need to specify a version there, since AC_PREREQ will cover it for us. Things
will fail pretty loudly if AC_PREREQ isn't met.
There is a race condition between console_hangup()
and start_stream(). It is possible for console_hangup()
to be called and then the stream thread to begin after the hangup.
To avoid this a check in start_stream() to make sure the pvt-owner
still exists while the pvt lock is held is made. If the owner
is gone that means the channel hung up and start_stream should
be aborted.
........
Use the detected pthread building flags in every place, instead of hardcoding -lpthread.
We nicely detect the right flags on each system for building Asterisk with
pthreads, then ignore it for every other build option that requires us to
build with pthreads. This caused some items to return a false negative.
Also cleanup some minor naming issues that caused "library library" redundancy
in the output.
Russell Bryant [Thu, 6 May 2010 14:04:36 +0000 (14:04 +0000)]
Merged revisions 261496 via svnmerge from
https://origsvn.digium.com/svn/asterisk/trunk
........
r261496 | russell | 2010-05-06 08:58:07 -0500 (Thu, 06 May 2010) | 40 lines
Fix handling of removing nodes from the middle of a heap.
This bug surfaced in 1.6.2 and does not affect code in any other released
version of Asterisk. It manifested itself as SIP qualify not happening when
it should, causing peers to go unreachable. This was debugged down to scheduler
entries sometimes not getting executed when they were supposed to, which was in
turn caused by an error in the heap code.
The problem only sometimes occurs, and it is due to the logic for removing an entry
in the heap from an arbitrary location (not just popping off the top). The scheduler
performs this operation frequently when entries are removed before they run (when
ast_sched_del() is used).
In a normal pop off of the top of the heap, a node is taken off the bottom,
placed at the top, and then bubbled down until the max heap property is restored
(see max_heapify()). This same logic was used for removing an arbitrary node
from the middle of the heap. Unfortunately, that logic is full of fail. This
patch fixes that by fully restoring the max heap property when a node is thrown
into the middle of the heap. Instead of just pushing it down as appropriate, it
first pushes it up as high as it will go, and _then_ pushes it down.
Lastly, fix a minor problem in ast_heap_verify(), which is only used for
debugging. If a parent and child node have the same value, that is not an
error. The only error is if a parent's value is less than its children.
A huge thanks goes out to cappucinoking for debugging this down to the scheduler,
and then producing an ast_heap test case that demonstrated the breakage. That
made it very easy for me to focus on the heap logic and produce a fix. Open source
projects are awesome.
Voicemail transfer to operator should occur immediately, not after main menu.
There were two scenarios in the advanced options that while using the
operator=yes and review=yes options, the transfer occurred only after exiting
the main menu (after sending a reply or leaving a message for an extension).
Now after the audio is processed for the reply or message the transfer occurs
immediately as expected.
Improve handling of T.38 re-INVITEs that arrive before a T.38-capable
application is executing on a channel.
This patch addresses an issue found during working with end-users
using res_fax. If an incoming call is answered in the dialplan, or
jumps to the 'fax' extension due to reception of a CNG tone (with
faxdetect enabled), and then the remote endpoint sends a T.38
re-INVITE, it is possible for the channel's T.38 state to be
'T38_STATE_NEGOTIATING' when the application starts up. Unfortunately,
even if the application wants to use T.38, it can't respond to the
peer's negotiation request, because the AST_CONTROL_T38_PARAMETERS
control frame that chan_sip sent originally has been lost, and the
application needs the content of that frame to be able to formulate a
reply.
This patch adds a new 'request' type to AST_CONTROL_T38_PARAMETERS,
AST_T38_REQUEST_PARMS. If the application sends this request, chan_sip
will re-send the original control frame (with
AST_T38_REQUEST_NEGOTIATE as the request type), and the application
can respond as normal. If this occurs within the five second timeout
in chan_sip, the automatic cancellation of the peer reinvite will be
stopped, and the application will 'own' the negotiation process from
that point onwards.
This also improves the code path in chan_sip to allow sip_indicate(),
when called for AST_CONTROL_T38_PARAMETERS, to be able to return a
non-zero response, which should have been in place before since the
control frame *can* fail to be processed properly. It also modifies
ast_indicate() to return whatever result the channel driver returned
for this control frame, rather than converting all non-zero results
into '-1'. Finally, the new request type intentionally returns a
positive value, so that an application that sends
AST_T38_REQUEST_PARMS can know for certain whether the channel driver
accepted it and will be replying with a control frame of its own, or
whether it was ignored (if the sip_indicate()/ast_indicate() path had
properly supported failure responses before, this would not be
necessary).
This patch also modifies res_fax to take advantage of the new request.
In addition, this patch makes sip_t38_abort() actually lock the
private structure before doing its work... bad programmer, no donut.
This patch also enhances chan_sip's 'faxdetect' support to allow
triggering on T.38 re-INVITEs received as well as CNG tone detection.
Ensure channel state is not incorrectly set in the case of a very early answer.
The needringing bit was being read in dahdi_read after answering thereby
setting the state to ringing from up. This clears needringing upon answering
so that is no longer possible.
Fix potential crash from race condition due to accessing channel data without the channel locked.
In res_musiconhold.c, there are several places where a channel's
stream's existence is checked prior to calling ast_closestream on it. The issue
here is that in several cases, the channel was not locked while checking the
stream. The result was that if two threads checked the state of the channel's
stream at approximately the same time, then there could be a situation where
both threads attempt to call ast_closestream on the channel's stream. The result
here is that the refcount for the stream would go below 0, resulting in a crash.
I have added proper channel locking to res_musiconhold.c to ensure that
we do not try to check chan->stream without the channel locked. A Digium customer
has been using this patch for several weeks and has not had any crashes since
applying the patch.
projects / thirdparty / asterisk.git / log
