George Joseph [Tue, 11 Jul 2017 12:26:27 +0000 (06:26 -0600)]
res_musiconhold: Add kill_escalation_delay, kill_method to class
By default, when res_musiconhold reloads or unloads, it sends a HUP
signal to custom applications (and all descendants), waits 100ms,
then sends a TERM signal, waits 100ms, then finally sends a KILL
signal. An application which is interacting with an external
device and/or spawns children of its own may not be able to exit
cleanly in the default times, expecially if sent a KILL signal, or
if it's children are getting signals directly from
res_musiconhoild.
* To allow extra time, the 'kill_escalation_delay'
class option can be used to set the number of milliseconds
res_musiconhold waits before escalating kill signals, with the
default being the current 100ms.
* To control to whom the signals are sent, the "kill_method" class
option can be set to "process_group" (the default, existing
behavior), which sends signals to the application and its
descendants directly, or "process" which sends signals only to the
application itself.
George Joseph [Wed, 5 Jul 2017 20:31:43 +0000 (14:31 -0600)]
http.c: Reduce log spam
Messages like "fwrite() failed: Connection reset by peer" are no
help whatsoever, especially since they can be caused simply by a
client disconnecting.
* Make those WARNINGs DEBUGs.
* Check the return of the headers fprintf.
res_pjsip: Fix crash with from_user containing invalid characters.
If the from_user field contains certain characters (like @, {, ^, etc.),
PJSIP will return a null value for the URI when attempting to parse it.
This causes a crash when trying to dial out through a trunk that contains
these invalid characters in its from_user field.
This change checks the configuration and ensures that an endpoint will
not be created if the from_user contains an invalid character. It also
adds a null check to the PJSIP URI parsing as a backup.
Richard Mudgett [Wed, 5 Jul 2017 18:39:45 +0000 (13:39 -0500)]
res_rtp_asterisk.c: Fix TURN deadlock by using ICE session group lock.
When a message is received on the TURN socket, the code processing the
message needs to call into the ICE/STUN session for further processing.
This code path locks the TURN group lock then the ICE/STUN group lock. In
another thread an ICE/STUN timer can fire off to send a keep alive message
over the TURN socket. In this code path, the ICE/STUN group lock is
obtained then the TURN group lock is obtained to send the packet. A
classic deadlock case if the group locks are not the same.
* Made TURN get created using the ICE/STUN session's group lock.
NOTE: I was originally concerned that the ICE/STUN session can get
recreated by ice_reset_session() for an event like RTCP multiplexing
causing a change during SDP negotiation. In this case the TURN group lock
would become different. However, TURN is also recreated as part of the
ICE/STUN recreation in ice_create() when all known ICE candidates are
added to the new ICE session. While the ICE/STUN and TURN sessions are
being recreated there is a period where the group locks could be
different.
ASTERISK-27023 #close
Patches:
res_rtp_asterisk-turn-deadlock-fix.patch (license #6502)
patch uploaded by Michael Walton (modified)
Richard Mudgett [Fri, 23 Jun 2017 16:17:51 +0000 (11:17 -0500)]
bridge_native_rtp.c: Fix direct media video RTP instance ACL check.
The video stream was using the audio stream RTP instance addresses to
check if the video RTP gets directed to an allowed direct media Access
Control List (ACL) address. There is no guarantee that the video RTP
instance uses the same addresses as the audio RTP instance.
This looks like it has been a bug since v11 when direct media ACL was
first added to chan_sip and then faithfully reproduced through a couple
code refactorings into the new bridging architecture.
Sean Bright [Wed, 5 Jul 2017 12:42:07 +0000 (08:42 -0400)]
core: Fix segfault when invoking 'data get' CLI command
Invoking 'data get /asterisk/core/channeltypes' caused a crash because
of an assumption of a tech's capabilities to be non-NULL. The
'Surrogate' tech, however, does have a NULL capabilities member,
resulting in a crash.
Alexander Traud [Mon, 3 Jul 2017 15:59:43 +0000 (17:59 +0200)]
chan_sip: Only when different, add TCP|TLS in autodomain (SIP Domain Support).
When sip.conf contained tcpenable=yes and autodomain=yes, the TCP domain was
added in any case, because of a local Boolean-negation error of the return value
of ast_sockaddr_cmp. After fixing this error for TCP and TLS, the TLS domain was
still always added with tlsenable=yes, because the domains were not compared
just on the address but also on the port – and TLS is always on a different port
than UDP/TCP.
Alexander Traud [Mon, 3 Jul 2017 15:38:32 +0000 (17:38 +0200)]
chan_sip: Fix a typo for tlsbindaddr in autodomain (SIP Domain Support).
Because of a copy-and-paste error when the struct ast_sockaddr changed,
tlsbindaddr was not added, when sip.conf contained autodomain=yes; see
"show sip domains" on the command-line interface (CLI) of Asterisk.
Sean Bright [Thu, 29 Jun 2017 18:58:35 +0000 (14:58 -0400)]
app_voicemail: Cleanup ODBC connection handling
The primary focus of this patch is adding a missing call to
ast_odbc_release_obj(), but is also a general cleanup of the ODBC
related code in app_voicemail.
Richard Mudgett [Thu, 29 Jun 2017 23:27:20 +0000 (18:27 -0500)]
pjsip_distributor.c: Fix deadlock with TCP type transports.
When a SIP message comes in on a transport, pjproject obtains the lock on
the transport and pulls the data out of the socket. Unlike UDP, the TCP
transport does not allow concurrent access. Without concurrency the
transport lock is not released when the transport's message complete
callback is called. The processing continues and eventually Asterisk
starts processing the SIP message. The first thing Asterisk tries to do
is determine the associated dialog of the message to determine the
associated serializer. To get the associated serializer safely requires
us to get the dialog lock.
To send a request or response message for a dialog, pjproject obtains the
dialog lock and then obtains the transport lock. Deadlock can result
because of the opposite order the locks are obtained.
* Fix the deadlock by obtaining the serializer associated with the dialog
another way that doesn't involve obtaining the dialog lock. In this case,
we use an ao2 container to hold the associated endpoint and serializer.
The new locks are held a brief time and won't overlap other existing lock
times.
George Joseph [Fri, 30 Jun 2017 13:31:52 +0000 (07:31 -0600)]
pjproject_bundled: Allow passing configure options to bundled
There wasn't any good way to pass options like --host or --build
down to the pjproject configure which makes cross-compiling difficult.
* Added a new PJPROJECT_CONFIGURE_OPTS environment variable which
can be used to pass arbitrary options to pjproject configure.
* Automatically set the pjproject configure --host and --build
options to match those supplied for the asterisk configure.
ASTERISK-27097 #close Reported-by: Kinsey Moore
Change-Id: I5fa776e110262851173002a26ffe1172e4c35b2e
George Joseph [Thu, 29 Jun 2017 19:50:14 +0000 (13:50 -0600)]
chan_pjsip: Fix ability to send UPDATE on COLP
When connected_line_method is "invite", we're supposed to determine
if the client can support UPDATE and if it can, send UPDATE instead
of INVITE to avoid the SDP renegotiation. Not only was pjproject
not setting the PJSIP_INV_SUPPORT_UPDATE flag, we were testing
that invite_tsx wasn't NULL which isn't always the case.
* Updated chan_pjsip/update_connected_line_information to drop the
requirement that invite_tsx isn't NULL.
* Submitted patch to pjproject sip_inv.c that sets the
PJSIP_INV_SUPPORT_UPDATE flag correctly.
* Updated pjsip.conf.sample to clarify what happens when "invite"
is specified.
Ivan Poddubny [Tue, 27 Jun 2017 09:37:11 +0000 (11:37 +0200)]
app_queue: Fix returning to dialplan when a queue is empty
The fix for ASTERISK-25665 introduced a regression.
The return value of queue_exec used to be 0 in case of leavewhenempty
but it was changed to -1 (returned from wait_our_turn and passed
transparently by queue_exec), thus leading to hangup instead of returning
back to dialplan.
This commit resets the value back to 0 in this case, restoring
original behavior.
Joshua Colp [Wed, 28 Jun 2017 14:03:31 +0000 (14:03 +0000)]
res_rtp_asterisk: Fix issues with ICE renegotiation.
When re-inviting to add more streams it is possible for
the role of existing ICE sessions to be changed to the
incorrect value. This results in subsequent refreshes
within the sessions getting a role conflict and the ICE
session breaking down. This change only sets the role to
be the new value if an ICE renegotiation is actually
going to happen, otherwise the existing role is preserved.
As well if we encounter a situation where a unidirectional
ICE negotiation happens and the other side does not send us
candidates we will not store any information for sending
traffic, even though we know where they are reachable. This
change fixes this by using the source of the ICE traffic
itself as the target if no candidates are known and we
receive some ICE traffic.
George Joseph [Fri, 9 Jun 2017 03:50:43 +0000 (21:50 -0600)]
bridge_native_rtp: Keep rtp instance refs on bridge_channel
There have been reports of deadlocks caused by an attempt to send a frame
to a channel's rtp instance after the channel has left the native bridge
and been destroyed. This patch effectively causes the bridge channel to
keep a reference to the glue and both the audio and video rtp instances
so what gets started will get stopped.
ASTERISK-26978 #close Reported-by: Ross Beer
Change-Id: I9e1ac49fa4af68d64826ccccd152593cf8cdb21a
Torrey Searle [Thu, 15 Jun 2017 08:12:41 +0000 (10:12 +0200)]
res_pjsip: Add DTMF INFO Failback mode
The existing auto dtmf mode reverts to inband if 4733 fails to be
negotiated. This patch adds a new mode auto_info which will
switch to INFO instead of inband if 4733 is not available.
Alexei Gradinari [Fri, 16 Jun 2017 23:08:30 +0000 (19:08 -0400)]
res_pjsip_mwi: update unsolicited MWI subscriptions on updating contact
Do not need to unsubscribe/subscribe on creating the ednpoint's contact.
The modified function create_mwi_subscriptions_for_endpoint adds
the subscription only if it does not exist.
The subscriptions aren't added for active contacts
which are retrieved on startup from realtime
if mwi_disable_initial_unsolicited=yes.
Because the mwi_contact_added is not called.
So the subscriptions also should be created on updating contact.
Kevin Harwell [Tue, 20 Jun 2017 20:41:14 +0000 (15:41 -0500)]
core_local: local channel data not being properly unref'ed and unlocked
In an earlier version of Asterisk a local channel [un]lock all functions were
added in order to keep a crash from occurring when a channel hung up too early
during an attended transfer. Unfortunately, when a transfer failure occurs and
depending on the timing, the local channels sometime do not get properly
unlocked and deref'ed after being locked and ref'ed. This happens because the
underlying local channel structure gets NULLed out before unlocking.
This patch reworks those [un]lock functions and makes sure the values that get
locked and ref'ed later get unlocked and deref'ed.
Kevin Harwell [Tue, 20 Jun 2017 21:01:48 +0000 (16:01 -0500)]
bridge: stuck channel(s) after failed attended transfer
If an attended transfer failed it was possible for some of the channels
involved to get "stuck" because Asterisk was not hanging up the transfer target.
This patch ensures Asterisk hangs up the transfer target when an attended
transfer failure occurs.
George Joseph [Fri, 16 Jun 2017 14:31:04 +0000 (08:31 -0600)]
res_stasis: Plug reference leak on stolen channels
When a stasis channel is stolen by another app, the control
structure is unreffed but never unlinked from the app_controls
container. This causes the channel reference to leak.
Added OBJ_UNLINK to the callback in channel_stolen_cb.
Also added some additional channel lifecycle debug messages to
channel.c.
ASTERISK-27059 #close Repoorted-by: George Joseph
Change-Id: Ib820936cd49453f20156971785e7f4f182c56e14
Core/PBX: Deadlock between dialplan execution and application unregistration.
Not easy to reproduce, but we have noticed deadlocks when unloading a module
while dialplan is handling a request.
The deadlock is between :
1) Dialplan execution: pbx_extension_helper() first taking conlock,
then pbx_findapp() [when called] asking for lock on apps list.
2) Application unregistration: ast_unregister_application() first taking lock
on apps list, then unreference_cached_app() [when called] asking for conlock.
As a protection, I suggest to modify ast_unregister_application(), so that it
anticipates the need of conlock, before taking the lock on apps list.
The side effect is a longer unavailability of conlock when unregistering an
application.
Alexei Gradinari [Mon, 12 Jun 2017 14:23:56 +0000 (10:23 -0400)]
res_pjsip: New endpoint option "notify_early_inuse_ringing"
This option was added to control whether to notify dialog-info state
'early' or 'confirmed' on Ringing when already INUSE.
The value "yes" is useful for some SIP phones (Cisco SPA)
to be able to indicate and pick up ringing devices.
Jan Friesse [Thu, 30 Mar 2017 14:33:51 +0000 (16:33 +0200)]
res_corosync: Change thread stack size
In Corosync 2.x libraries were changed to use LibQB IPC.
Sadly LibQB IPC doesn't support copy-free access to received buffer, so
Corosync libraries were rewritten to use stack as buffer. Mostly the
needed stack size is quite small, but for all *_dispatch functions, 1MiB
is needed.
Asterisk function ast_pthread_create_background set stack size for new
thread to much smaller AST_BACKGROUND_STACKSIZE (~500KiB).
This results in Asterisk crash when running with Corosync 2.x.
Patch solves this issue by creating it's own version of
ast_pthread_create_background which sets stack size to much higher value
(actually it's AST_BACKGROUND_STACKSIZE + 3MiB).
Another problem may appear when "corosync show members" netconsole
command is executed. It is also executed in thread and also has only
500KiB stack size. Sadly it calls corosync_cfg_get_node_addrs which
again needs at least 1MiB stack.
Solution is to use HAVE_COROSYNC_CFG_STATE_TRACK as a discriminator
between Corosync 1.x and 2.x. If 1.x is found, nothing changes. If 2.x
is found, NodeID is displayed instead of IP address.
George Joseph [Tue, 13 Jun 2017 16:33:34 +0000 (10:33 -0600)]
res_ari: Add "module loaded" check to ari stubs
The recent change to make the use of LOAD_DECLINE more consistent
caused res_ari to unload itself before declining if the ari.conf
file wasn't found. The ari stubs though still tried to use the
configuration resulting in segfaults.
This patch creates a new CHECK_ARI_MODULE_LOADED macro which tests
to see if res_ari is actually loaded and causes the stubs to also
decline if it isn't. The macro was then added to the mustache
template's "load_module" function.
ASTERISK-27026 #close Reported-by: Ronald Raikes
Change-Id: I263d56efa628ee3c411bdcd16d49af6260c6c91d
Richard Mudgett [Thu, 15 Jun 2017 17:33:22 +0000 (12:33 -0500)]
chan_pjsip: Fix PJSIP_MEDIA_OFFER dialplan function read.
The construction of the returned string assumed incorrectly that the
supplied buffer would always be initialized as an empty string. If it is
not an empty string we could overrun the supplied buffer by the length of
the non-empty buffer string plus one. It is also theoreticaly possible
for the supplied buffer to be overrun by a string terminator during a read
operation even if the supplied buffer is an empty string.
* Fix the assumption that the supplied buffer would already be an empty
string. The buffer is not guaranteed to contain an empty string by all
possible callers.
Joshua Colp [Thu, 15 Jun 2017 12:32:32 +0000 (12:32 +0000)]
channel: Fix reference counting in ast_channel_suppress.
The ast_channel_suppress function wrongly decremented the
reference count of the underlying structure used to keep
track of what should be suppressed on a channel if the
function was called multiple times on the same channel.
This change cleans up the reference counting a bit so
this no longer occurs.
George Joseph [Wed, 14 Jun 2017 13:54:45 +0000 (07:54 -0600)]
res_rtp_asterisk: Fix ssrc change for rtcp srtp
It looks like there was a copy/paste error in ast_rtp_change_source
where if there was a rtcp srtp instance, instead of updating its
ssrc we were updating the srtp instance ssrc twice.
ASTERISK-27022 #close Reported-by: Michael Walton
Change-Id: Ic88f3aee7227b401c58745ac265ff92c19620095
Joshua Colp [Thu, 8 Jun 2017 19:38:51 +0000 (19:38 +0000)]
bridge: Add a deferred queue.
This change adds a deferred queue to bridging. If a bridge
technology determines that a frame can not be written and
should be deferred it can indicate back to bridging to do so.
Bridging will then requeue any deferred frames upon a new
channel joining the bridge.
This change has been leveraged for T.38 request negotiate
control frames. Without the deferred queue there is a race
condition between the bridge receiving the T.38 request
negotiate and the second channel joining and being in the
bridge. If the channel is not yet in the bridge then the T.38
negotiation fails.
A unit test has also been added that confirms that a T.38
request negotiate control frame is deferred when no other
channel is in the bridge and that it is requeued when a new
channel joins the bridge.
Kevin Harwell [Tue, 13 Jun 2017 19:17:29 +0000 (14:17 -0500)]
res_pjsip_refer/session: Calls dropped during transfer
When doing an attended transfer it's possible for the transferer, after
receiving an accepted response from Asterisk, to send a BYE to Asterisk,
which can then be processed before Asterisk has time to start and/or
complete the transfer process. This of course causes the transfer to not
complete successfully, thus dropping the call.
This patch makes it so any BYEs received from the transferer, after the REFER,
that initiate a session end are deferred until the transfer is complete. This
allows the channel that would have otherwise been hung up by Asterisk to
remain available throughout the transfer process.
Alexei Gradinari [Mon, 12 Jun 2017 14:57:24 +0000 (10:57 -0400)]
res_pjsip_mwi: don't create mwi subscriptions if initial unsolicited disabled
If sending unsolicited mwi to all endpoints on startup is disabled
(mwi_disable_initial_unsolicited=yes) do not need to create subscriptions.
If there are many (thousands) realtime endpoints configured with unsolicited mwi
and Vociemail Storage configured as ODBC or IMAP there will be huge number of
DB/IMAP requests on startup.
Guido Falsi [Thu, 8 Jun 2017 15:54:46 +0000 (17:54 +0200)]
BuildSystem: Add patches to allow building with recent LibreSSL
Add some #if defined checks which allow building against LibreSSL.
These patchess come from OpenBSD ports:
https://cvsweb.openbsd.org/cgi-bin/cvsweb/ports/telephony/asterisk/patches/
projects / thirdparty / asterisk.git / log
