The change in question was intended to prevent the need to reload in
order to update qualifies on contacts when an AOR changes. However, this
ended up causing a deadlock instead.
Joshua Colp [Tue, 7 Feb 2017 18:01:03 +0000 (18:01 +0000)]
srv: Fix crash when ast_srv_lookup is used and 0 records are returned.
When performing an SRV lookup using the ast_srv_lookup function it
did not properly handle the situation where 0 records are returned.
If this happened it would wrongly assume that at least one record
was present.
This change fixes the code so it will exit early if an error occurs
or if 0 records are returned.
ASTERISK-26772
patches:
srv_lookup.patch submitted by nappsoft (license 6822)
Joshua Colp [Mon, 6 Feb 2017 17:40:45 +0000 (17:40 +0000)]
res_stasis_device_state: Protect the adding/removing of subscriptions.
The adding and removing of device state subscriptions did not protect
fully against simultaneous manipulation. In particular the subscribe
case allowed a small window where two subscriptions could be added for
the same device state instead of just one.
This change makes the code hold the subscriptions lock for the entirety
of each operation to ensure that two are not occurring at the same time.
Mark Michelson [Wed, 1 Feb 2017 19:54:50 +0000 (13:54 -0600)]
Update qualifies when AOR configuration changes.
Prior to this change, qualifies would only update in the following
cases:
* A reload of res_pjsip.so was issued.
* A dynamic contact was re-registered after its AOR's qualify_frequency
had been changed
This does not work well if you are using realtime for your AORs. You can
update your database to have a new qualify_frequency, but the permanent
contacts on that AOR will not have their qualifies updated. And the
dynamic contacts on that AOR will not have their qualifies updated until
the next registration, which could be a long time.
This change seeks to fix this problem by making it so that whenever AOR
configuration is applied, the contacts pertaining to that AOR have their
qualifies updated.
Additions from this patch:
* AOR sorcery objects now have an apply handler that calls into a newly
added function in the OPTIONS code. This causes all contacts
associated with that AOR to re-schedule qualifies.
* When it is time to qualify a contact, the OPTIONS code checks to see
if the AOR can still be retrieved. If not, then qualification is
canceled on the contact.
Alterations from this patch:
* The registrar code no longer updates contact's qualify_frequence and
qualify_timeout. There is no point to this since those values already
get updated when the AOR changes.
* Reloading res_pjsip.so no longer calls the OPTIONS initialization
function. Reloading res_pjsip.so results in re-loading AORs, which
results in re-scheduling qualifies.
Richard Mudgett [Wed, 1 Feb 2017 00:28:15 +0000 (18:28 -0600)]
channel.c: Fix unbalanced read queue deadlocking local channels.
Using the timerfd timing module can cause channel freezing, lingering, or
deadlock issues. The problem is because this is the only timing module
that uses an associated alert-pipe. When the alert-pipe becomes
unbalanced with respect to the number of frames in the read queue bad
things can happen. If the alert-pipe has fewer alerts queued than the
read queue then nothing might wake up the thread to handle received frames
from the channel driver. For local channels this is the only way to wake
up the thread to handle received frames. Being unbalanced in the other
direction is less of an issue as it will cause unnecessary reads into the
channel driver.
ASTERISK-26716 is an example of this deadlock which was indirectly fixed
by the change that found the need for this patch.
* In channel.c:__ast_queue_frame(): Adding frame lists to the read queue
did not add the same number of alerts to the alert-pipe. Correspondingly,
when there is an exceptionally long queue event, any removed frames did
not also remove the corresponding number of alerts from the alert-pipe.
Richard Mudgett [Tue, 31 Jan 2017 22:38:49 +0000 (16:38 -0600)]
res_agi: Prevent an AGI from eating frames it should not. (Re-do)
A dialplan intercept routine is equivalent to an interrupt routine. As
such, the routine must be done quickly and you do not have access to the
media stream. These restrictions are necessary because the media stream
is the responsibility of some other code and interfering with or delaying
that processing is bad. A possible future dialplan processing
architecture change may allow the interception routine to run in a
different thread from the main thread handling the media and remove the
execution time restriction.
* Made res_agi.c:run_agi() running an AGI in an interception routine run
in DeadAGI mode. No touchy channel frames.
Richard Mudgett [Tue, 31 Jan 2017 22:32:18 +0000 (16:32 -0600)]
Frame deferral: Revert API refactoring.
There are several issues with deferring frames that are caused by the
refactoring.
1) The code deferring frames mishandles adding a deferred frame to the
deferred queue. As a result the deferred queue can only be one frame
long.
2) Deferrable frames can come directly from the channel driver as well as
the read queue. These frames need to be added to the deferred queue.
3) Whoever is deferring frames is really only doing the __ast_read() to
collect deferred frames and doesn't care about the returned frames except
to detect a hangup event. When frame deferral is completed we must make
the normal frame processing see the hangup as a frame anyway. As such,
there is no need to have varying hangup frame deferral methods. We also
need to be aware of the AST_SOFTHANGUP_ASYNCGOTO hangup that isn't real.
That fake hangup is to cause the PBX thread to break out of loops to go
execute a new dialplan location.
4) To properly deal with deferrable frames from the channel driver as
pointed out by (2) above, means that it is possible to process a dialplan
interception routine while frames are deferred because of the
AST_CONTROL_READ_ACTION control frame. Deferring frames is not
implemented as a re-entrant operation so you could have the unsupported
case of two sections of code thinking they have control of the media
stream.
A worse problem is because of the bad implementation of the AMI PlayDTMF
action. It can cause two threads to be deferring frames on the same
channel at the same time. (ASTERISK_25940)
* Rather than fix all these problems simply revert the API refactoring as
there is going to be only autoservice and safe_sleep deferring frames
anyway.
Joshua Colp [Tue, 31 Jan 2017 17:17:50 +0000 (17:17 +0000)]
res_pjsip: Handle invocation of callback on outgoing request when error occurs.
There are some error cases in PJSIP when sending a request that will
result in the callback for the request being invoked. The code did not
handle this case and assumed on every error case that the callback was not
invoked.
The code has been changed to check whether the callback has been invoked
and if so to absorb the error and treat it as a success.
Sean Bright [Mon, 30 Jan 2017 15:02:14 +0000 (10:02 -0500)]
res_rtp_asterisk: Swap byte-order when sending signed linear
Before Asterisk 13, signed linear was converted into network byte order by a
smoother before being sent over the network. We restore this behavior by
forcing the creation of a smoother when slinear is in use and setting the
appropriate flags so that the byte order conversion is always done.
George Joseph [Wed, 25 Jan 2017 12:50:43 +0000 (05:50 -0700)]
debug_utilities: Add ast_logescalator
The escalator works by creating a set of startup commands in cli.conf
that set up logger channels and issue the debug commands for the
subsystems specified. If asterisk is running when it is executed,
the same commands will be issued to the running instance. The original
cli.conf is saved before any changes are made and can be restored by
executing '$prog --reset'.
The log output will be stored in...
$astlogdir/message.$uniqueid
$astlogdir/debug.$uniqueid
$astlogdir/dtmf.$uniqueid
$astlogdir/fax.$uniqueid
$astlogdir/security.$uniqueid
$astlogdir/pjsip_history.$uniqueid
$astlogdir/sip_history.$uniqueid
Some minor tweaks were made to chan_sip, and res_pjsip_history
so their history output could be send to a log channel as packets
are captured.
A minor tweak was also made to manager so events are output to verbose
when "manager set debug on" is issued.
kkm [Wed, 25 Jan 2017 01:51:07 +0000 (17:51 -0800)]
make_build_h: handle backslashes in external strings
LikewiseOpen creates user names with a backslash in them. A gentle
massage with sed(1) allows such strings to be inserted into build.h
properly quoted. I am also adding the same for host name and other
strings used in the script that are more or less user-controlled.
kkm [Wed, 18 Jan 2017 02:46:43 +0000 (18:46 -0800)]
ast_careful_fwrite to support EPIPE gracefully
When a reading end of the network socket is closed by an AMI manager,
the EPIPE is signaled when writing to our end, resulting in the
spurious log error message
ast_careful_fwrite: fwrite() returned error: Broken pipe
Previously EPIPE was handled in ast_carefulwrite() a few lines above,
but not in this function.
kkm [Wed, 25 Jan 2017 04:31:38 +0000 (20:31 -0800)]
app_queue: Fix queues randomly disappearing on reload
With 500+ queues and a reload every minute, a random queue disappears
upon reload. The cause is mususe of the 'dead' flag. Namely, all queues
were marked dead up front, and then "resurrected" by dropping this flag
for those found in the configuration. But a queue marked dead can be
removed also when control leaves the app entry point on a PBX thread.
With this change, the queue is marked only not found, and at the end of
reload only the queues that are still not found are actually marked as
dead, so the dead flag is never reset, and set only on positively dead
queues.
Mark Michelson [Wed, 25 Jan 2017 21:26:53 +0000 (15:26 -0600)]
Add reload options to CLI/AMI stale object commands.
Marking an object as stale in a memory cache is supposed to prime the
cache so that the next time the item is retrieved, the stale item is
deleted from the cache and a background task is run to re-populate the
cache with a fresh version of the object.
The problem is, there are some object types out there for which there is
no natural reason that they would be retrieved from the backend with any
regularity. Outbound PJSIP registrations are a good example of this. At
startup, they are read, and an object-specific state is created that
refers to the initially-retrieved object for all time.
Adding the "reload" option to the CLI/AMI commands gives the cache the
opportunity to manually re-retrieve the object from the backend, both
storing the new object in the cache and applying the new object's
configuration to the module that uses that object.
Joshua Colp [Tue, 24 Jan 2017 21:39:39 +0000 (21:39 +0000)]
res_pjsip_endpoint_identifier_ip: Ensure error defaults to 0.
When configuring a match using a netmask the error variable was
not defaulting to 0. For some people this would cause the code
to think an error occurred when adding the match when in reality
it added perfectly fine.
Richard Mudgett [Tue, 10 Jan 2017 19:11:20 +0000 (13:11 -0600)]
main/app.c: Memory corruption from early format destruction.
* make_silence() created a malloced silence slin frame without adding a
slin format ref. When the frame is destroyed it will unref the slin
format that never had a ref added. Memory corruption is expected to
follow.
* Simplified and fixed counting the number of samples in a frame list for
make_silence().
* Eliminated an unnecessary RAII_VAR associated with the make_silence()
frame.
George Joseph [Thu, 19 Jan 2017 15:05:36 +0000 (08:05 -0700)]
ari: Implement 'debug all' and request/response logging
The 'ari set debug' command has been enhanced to accept 'all' as an
application name. This allows dumping of all apps even if an app
hasn't registered yet. To accomplish this, a new global_debug global
variable was added to res/stasis/app.c and new APIs were added to
set and query the value.
'ari set debug' now displays requests and responses as well as events.
This required refactoring the existing debug code.
* The implementation for 'ari set debug' was moved from stasis/cli.{c,h}
to ari/cli.{c,h}, and stasis/cli.{c,h} were deleted.
* In order to print the body of incoming requests even if a request
failed, the consumption of the body was moved from the ari stubs
to ast_ari_callback in res_ari.c and the moustache templates were
then regenerated. The body is now passed to ast_ari_invoke and then
on to the handlers. This results in code savings since that template
was inserted multiple times into all the stubs.
An additional change was made to the ao2_str_container implementation
to add partial key searching and a sort function. The existing cli
code assumed it was already there when it wasn't so the tab completion
was never working.
Richard Mudgett [Sat, 21 Jan 2017 03:13:34 +0000 (21:13 -0600)]
PJPROJECT logging: Fix detection of max supported log level.
The mechanism used for detecting the maximum log level compiled into the
linked pjproject did not work. The API call simply stores the requested
level into an integer and does no range checking. Asterisk was assuming
that there was range checking and limited the new value to the allowable
range. To get the actual maximum log level compiled into the linked
pjproject we need to get and save off the initial set log level from
pjproject. This is the maximum log level supported.
* Get and save off the initial log level setting before altering it to the
desired level on startup. This has to be done by a macro rather than
calling a core function to avoid incorrectly linking pjproject.
* Split the initial log level warning messages to warn if the linked
pjproject cannot support the requested startup level and if it is too low
to get the pjproject buildopts for "pjproject show buildopts".
* Adjust the CLI "pjproject set log level" to check the saved max log
level and to generate normal output messages instead of a warning message.
Mark Michelson [Mon, 23 Jan 2017 22:18:18 +0000 (16:18 -0600)]
Free endpoint ACLs when destroying PJSIP endpoints.
If endpoint ACLs were specified, they were not being freed
when endpoints were destroyed. On systems with realtime endpoints, this
could add up quickly since each DB lookup would allocate the ACL without
freeing it.
George Joseph [Mon, 23 Jan 2017 15:10:50 +0000 (08:10 -0700)]
pjproject_bundled: Fix setting max log level
An earlier attempt to prevent pjsua from spitting out an extra 6795
lines of debug output every time the testsuite called it was also
turning off the ability for asterisk to output debug info when it
needed to. This patch reverts the earlier fix and instead adds
a pjproject patch that sets the startup log level to 1 for pjsua
pjsystest and the pjsua python binding. This is an asterisk-only
patch that does not affect pjproject functionality and will not be
submitted upstream.
Joshua Colp [Mon, 23 Jan 2017 16:08:37 +0000 (16:08 +0000)]
res_pjsip_endpoint_identifier_ip: Read settings before resolving.
An option has been added, srv_lookups, which controls whether
SRV lookups are performed on the provided match hosts or not.
It was possible for this option to be applied after resolution
had already happened.
This change makes it so hosts are stored away, settings are read
and applied, and then resolution is done. This ensures that no
matter the ordering the srv_lookups option is in effect.
Richard Mudgett [Sun, 22 Jan 2017 23:25:57 +0000 (17:25 -0600)]
LISTFILTER: Remove outdated ERROR message.
Feeding LISTFILTER an empty variable results in an invalid ERROR message.
Earlier changes made the message useless because we can no longer tell if
the variable is empty or does not exist. It is valid to try to remove a
value from an empty list just as it is valid to try to remove a value that
is not in a non-empty list.
* Removed the outdated ERROR message.
* Added more test cases to the LISTFILTER unit test.
Richard Mudgett [Thu, 5 Jan 2017 21:11:12 +0000 (15:11 -0600)]
res_pjsip_pubsub.c: Fix AMI event list counts.
Fix the AMI PJSIPShowSubscriptionsInbound, PJSIPShowSubscriptionsOutbound,
and PJSIPShowResourceLists actions event counts. The reported counts may
not necessarily be accurate depending on what happens.
The subscriptions count would be wrong if Asterisk ever has outbound
subscriptions.
The resource list count could be wrong if a list were added or removed
during the AMI action being processed.
George Joseph [Fri, 13 Jan 2017 17:03:15 +0000 (10:03 -0700)]
debug_utilities: Create ast_loggrabber
ast_loggrabber gathers log files from customizable search patterns,
optionally converts POSIX timestamps to a readable format and
tarballs the results.
It was possible for a frame to be re-inserted into a jitter buffer after it
had been removed from it. A case when this happened was if a frame was read
out of the jitterbuffer, passed to the translation core, and then multiple
frames were returned from said translation core. Upon multiple frames being
returned the first is passed on, but sebsequently "chained" frames are put
back into the read queue. Thus it was possible for a frame to go back into
the jitter buffer where this would cause problems.
This patch adds a flag to frames that are inserted into the channel's read
queue after translation. The abstract jitter buffer code then checks for this
flag and ignores any frames marked as such.
Richard Mudgett [Sat, 14 Jan 2017 03:23:49 +0000 (21:23 -0600)]
taskprocessor.c: Change when high water warning logged.
The task processor queue reached X scheduled tasks message was originally
intended to get logged only once per task processor to prevent spamming
the log. This is no longer necessary since high and low water thresholds
can better control when the message is logged.
It is beneficial to generate the warning each time a task processor
reaches the high water level because PJSIP stops processing new requests
while any high water alert is active. Without this change you would have
to enable at least debug level 3 logging to know about a repeated alert
trigger.
* Made generate the warning message whenever a task is pushed into the
task processor that triggers the high water alert.
* Appended 'again' to the warning for a repeated high water alert trigger.
Aaron An [Tue, 10 Jan 2017 11:54:16 +0000 (19:54 +0800)]
res_rtp_asterisk: Fix bug in function CHANNEL(rtcp, all_rtt)
Function CHANNEL(rtcp,all_rtt) CHANNEL(rtcp,all_loss) CHANNEL(rtcp,all_jitter)
always return 0.0 due to wrong define of macro "AST_RTP_SATA_SET" and
"AST_RTP_STAT_STRCPY".
It should compare "combined" with "stat" not "current_stat".
ASTERISK-26710 #close Reported-by: Aaron An Tested-by: AaronAn
Change-Id: Id4140fafbf92e2db689dac5b17d9caa009028a15
George Joseph [Wed, 11 Jan 2017 00:10:39 +0000 (17:10 -0700)]
debug_utilities: Create the ast_coredumper utility
This utility allows easy manipulation of asterisk coredumps.
* Configurable search paths and patterns for existing coredumps
* Can generate a consistent coredump from the running instance
* Can dump the lock_infos table from a coredump
* Dumps backtraces to separate files...
- thread apply 1 bt full -> <coredump>.thread1.txt
- thread apply all bt -> <coredump>.brief.txt
- thread apply all bt full -> <coredump>.full.txt
- lock_infos table -> <coredump>.locks.txt
* Can tarball corefiles and optionally delete them after processing
* Can tarball results files and optionally delete them after processing
* Converts ':' in coredump and results file names '-' to facilitate
uploading. Jira for instance, won't accept file names with colons
in them.
Tested on Fedora24+, Ubuntu14+, Debian6+, CentOS6+ and FreeBSD9+[1].
[1] For *BSDs, the "devel/gdb" package might have to be installed to
get a recent gdb. The utility will check all instances of gdb
it finds in $PATH and if one isn't found that can run python, it
prints a friendly error.
Joshua Colp [Thu, 5 Jan 2017 12:11:43 +0000 (12:11 +0000)]
res_pjsip_endpoint_identifier_ip: Add support for SRV lookups.
This change implements SRV support for the IP based endpoint
identifier module. All possible addresses through SRV are looked
up and added as matches. If no SRV records are available a
fallback to normal host resolution is done. If an IP address
is provided then no SRV lookup occurs.
This is configured using the "srv_lookups" option on the
identify section and defaults to "yes".
projects / thirdparty / asterisk.git / log
