First rtp activity check was performed after 500ms regardless of the rtp_timeout setting. Having a call in ringing state for more than rtp_timeout and the first rtp package is received more than 500ms after sdp negotiation and before the rtp_timeout, erronously caused the call to be hungup. Changed to perform the first rtp inactivity check after the timeout setting preventing calls to be disconnected before the rtp_timeout has elapsed since sdp negotiation.
Fixes#710
* Fixed possible memory leak in tn_config:tn_get_etn() where we
weren't releasing etn if tn or eprofile were null.
* We now canonicalize TNs before using them for lookups or adding
them to Identity headers.
* Fixed a typo in stir_shaken.conf.sample.
Resolves: #716
Add a new identify_by option to res_pjsip_endpoint_identifier_ip
called 'transport' this matches endpoints based on the bound
ip address (local) instead of the 'ip' option, which matches on
the source ip address (remote).
UserNote: set identify_by=transport for the pjsip endpoint. Then
use the existing 'match' option and the new 'transport' option of
the identify.
Fixes: #672
* OpenSSL 1.0.2 doesn't support X509_get0_pubkey so we now use
X509_get_pubkey. The difference is that X509_get_pubkey requires
the caller to free the EVP_PKEY themselves so we now let
RAII_VAR do that.
* OpenSSL 1.0.2 doesn't support upreffing an X509_STORE so we now
wrap it in an ao2 object.
* OpenSSL 1.0.2 doesn't support X509_STORE_get0_objects to get all
the certs from an X509_STORE and there's no easy way to polyfill
it so the CLI commands that list profiles will show a "not
supported" message instead of listing the certs in a store.
Resolves: #676
There were a few references in the embedded documentation XML
where the case didn't match or where the referenced app or function
simply didn't exist any more. These were causing 404 responses
in docs.asterisk.org.
Add ability to match against PJSIP request URI.
UserNote: this new feature let users match endpoints based on the
indound SIP requests' URI. To do so, add 'request_uri' to the
endpoint's 'identify_by' option. The 'match_request_uri' option of
the identify can be an exact match for the entire request uri, or a
regular expression (between slashes). It's quite similar to the
header identifer.
Fixes: #599
This commit introduces configurable TCP keepalive settings for both TCP and TLS transports. The changes allow for finer control over TCP connection keepalives, enhancing stability and reliability in environments prone to connection timeouts or where intermediate devices may prematurely close idle connections. This has proven necessary and has already been tested in production in several specialized environments where access to the underlying transport is unreliable in ways invisible to the operating system directly, so these keepalive and timeout mechanisms are necessary.
Fixes#657
There was functionality in chan_sip to get REFER headers, with GET_TRANSFERRER_DATA variable. This commit implements the same functionality in pjsip, to ease transfer from chan_sip to pjsip.
Fixes: #579
UserNote: the GET_TRANSFERRER_DATA dialplan variable can now be used also in pjsip.
The prometheus exposition format requires each line to be unique[1].
This is handled by struct prometheus_metric having a list of children
that is managed when registering a metric. In case the scrape callback
is used, it is the responsibility of the implementation to handle this
correctly.
Originally the bridge callback didn't handle NULL snapshots, the crash
fix lead to NULL metrics, and fixing that lead to duplicates.
The original code assumed that snapshots are not NULL and then relied on
"if (i > 0)" to establish the parent/children relationship between
metrics of the same class. This is not workerable as the first bridge
might be invisible/lacks a snapshot.
Fix this by keeping a separate array of the first metric by class.
Instead of relying on the index of the bridge, check whether the array
has an entry. Use that array for the output.
Add a test case that verifies that the help text is not duplicated.
Resolves: #642
[1] https://prometheus.io/docs/instrumenting/exposition_formats/#grouping-and-sorting
Currently, if a parking lot is full, bridge setup returns -1,
causing dialplan execution to terminate without TryExec.
However, such failures should be handled more gracefully,
the same way they are on other paths, as indicated by the
module's author, here:
http://lists.digium.com/pipermail/asterisk-dev/2018-December/077144.html
Now, callers will hear the parking failure announcement, and dialplan
will continue, which is consistent with existing failure modes.
Resolves: #624
In handle_negotiated_sdp the pending_media_state->read_callbacks must be
reset before they are added in the SDP handlers in
handle_negotiated_sdp_session_media. Otherwise, old callbacks for
removed streams and file descriptors could be added to the channel and
Asterisk would poll on non-existing file descriptors.
Resolves: #611
* Added checks for missing session, session->channel and rdata
in stir_shaken_incoming_request.
* Added checks for missing session, session->channel and tdata
in stir_shaken_outgoing_request.
Resolves: #645
In as_check_common_config, we were calling ast_std_free on
raw_key but raw_key was allocated with ast_malloc so it
should be freed with ast_free.
Resolves: #636
Why do we need a refactor?
The original stir/shaken implementation was started over 3 years ago
when little was understood about practical implementation. The
result was an implementation that wouldn't actually interoperate
with any other stir-shaken implementations.
There were also a number of stir-shaken features and RFC
requirements that were never implemented such as TNAuthList
certificate validation, sending Reason headers in SIP responses
when verification failed but we wished to continue the call, and
the ability to send Media Key(mky) grants in the Identity header
when the call involved DTLS.
Finally, there were some performance concerns around outgoing
calls and selection of the correct certificate and private key.
The configuration was keyed by an arbitrary name which meant that
for every outgoing call, we had to scan the entire list of
configured TNs to find the correct cert to use. With only a few
TNs configured, this wasn't an issue but if you have a thousand,
it could be.
What's changed?
* Configuration objects have been refactored to be clearer about
their uses and to fix issues.
* The "general" object was renamed to "verification" since it
contains parameters specific to the incoming verification
process. It also never handled ca_path and crl_path
correctly.
* A new "attestation" object was added that controls the
outgoing attestation process. It sets default certificates,
keys, etc.
* The "certificate" object was renamed to "tn" and had it's key
change to telephone number since outgoing call attestation
needs to look up certificates by telephone number.
* The "profile" object had more parameters added to it that can
override default parameters specified in the "attestation"
and "verification" objects.
* The "store" object was removed altogther as it was never
implemented.
* We now use libjwt to create outgoing Identity headers and to
parse and validate signatures on incoming Identiy headers. Our
previous custom implementation was much of the source of the
interoperability issues.
* General code cleanup and refactor.
* Moved things to better places.
* Separated some of the complex functions to smaller ones.
* Using context objects rather than passing tons of parameters
in function calls.
* Removed some complexity and unneeded encapsuation from the
config objects.
Resolves: #351Resolves: #46
UserNote: Asterisk's stir-shaken feature has been refactored to
correct interoperability, RFC compliance, and performance issues.
See https://docs.asterisk.org/Deployment/STIR-SHAKEN for more
information.
UpgradeNote: The stir-shaken refactor is a breaking change but since
it's not working now we don't think it matters. The
stir_shaken.conf file has changed significantly which means that
existing ones WILL need to be changed. The stir_shaken.conf.sample
file in configs/samples/ has quite a bit more information. This is
also an ABI breaking change since some of the existing objects
needed to be changed or removed, and new ones added. Additionally,
if res_stir_shaken is enabled in menuselect, you'll need to either
have the development package for libjwt v1.15.3 installed or use
the --with-libjwt-bundled option with ./configure.
Media Experience Score relies on incorrect pseudo_mos variable
calculation. According to forming an opinion section of the
documentation, calculation relies on ITU-T G.107 standard:
https://docs.asterisk.org/Deployment/Media-Experience-Score/#forming-an-opinion
ITU-T G.107 Annex B suggests to calculate MOS with a coefficient
"seven times ten to the power of negative six", 7 * 10^(-6). which
would mean 6 digits after the decimal point. Current implementation
has 7 digits after the decimal point, which downrates the calls.
Fixes: #597
When started with a verbose level of 3, asterisk can emit over 1500
verbose message that serve no real purpose other than to fill up
logs. When asterisk shuts down, it emits another 1100 that are of
even less use. Since the testsuite runs asterisk with a verbose
level of 3, and asterisk starts and stops for every one of the 700+
tests, the number of log messages is staggering. Besides taking up
resources, it also makes it hard to debug failing tests.
This commit changes the log level for those verbose messages to 5
instead of 3 which reduces the number of log messages to only a
handful. Of course, NOTICE, WARNING and ERROR message are
unaffected.
There's also one other minor change...
ast_context_remove_extension_callerid2() logs a DEBUG message
instead of an ERROR if the extension you're deleting doesn't exist.
The pjsip_config_wizard calls that function to clean up the config
and has been triggering that annoying error message for years.
Resolves: #582
This introduces a setting for outbound registrations to override the
global User-Agent header setting.
Resolves: #515
UserNote: PJSIP outbound registrations now support a per-registration
User-Agent header
Resolves a regression identified by @justinludwig involving the
rendering of IPv6 addresses in outgoing SDP.
Also updates `media_address` on PJSIP endpoints so that if we are able
to parse the configured value as an IP we store it in a format that we
can directly use later. Based on my reading of the code it appeared
that one could configure `media_address` as:
```
[foo]
type = endpoint
...
media_address = [2001:db8::]
```
And that value would be blindly copied into the outgoing SDP without
regard to its format.
Fixes#541
This reverts commit 315eb551db.
Over the past year, we've had several reports of "topology storms"
occurring where 2 external facing channels connected by one or more
local channels and bridges will get themselves in a state where
they continually send each other topology change requests. This
usually manifests itself in no-audio calls and a flood of
"Exceptionally long queue length" messages. It appears that this
commit is the cause so we're reverting it for now until we can
determine a more appropriate solution.
Resolves: #530
* Since ICE candidates are used for the check and pjproject is
required to use ICE, res_rtp_asterisk was failing to compile
when pjproject wasn't available. The check is now wrapped
with an #ifdef HAVE_PJPROJECT.
* The rtp->ice_active_remote_candidates container was being
used to check the address on incoming packets but that
container doesn't contain peer reflexive candidates discovered
during negotiation. This was causing the check to fail
where it shouldn't. We now check against pjproject's
real_ice->rcand array which will contain those candidates.
* Also fixed a bug in ast_sockaddr_from_pj_sockaddr() where
we weren't zeroing out sin->sin_zero before returning. This
was causing ast_sockaddr_cmp() to always return false when
one of the inputs was converted from a pj_sockaddr, even
if both inputs had the same address and port.
Resolves: #500Resolves: #503Resolves: #505
When updating an existing header the 'update' code incorrectly
just copied the new value into the existing buffer. If the
new value exceeded the available buffer size memory outside
of the buffer would be written into, potentially causing
a crash.
This change makes it so that the 'update' now duplicates
the new header value instead of copying it into the existing
buffer.
Add patch to split the log level for invalid packets received on the
signaling port. The warning regarding the packet will move to level 2
so that it can still be displayed, while the raw packet will be at level
4.
When ICE is in use, we can prevent a possible DOS attack by allowing
DTLS protocol messages (client hello, etc) only from sources that
are in the active remote candidates list.
Resolves: GHSA-hxj9-xwr8-w8pq
This fixes a number of broken links throughout the
tree, mostly caused by wiki.asterisk.org being replaced
with docs.asterisk.org, which should eliminate the
need for sporadic fixes as in f28047db36.
Resolves: #430
There are valid scenarios where res_odbc's connection pool might have some dead
or stuck connections while others are healthy (imagine network
elements/firewalls/routers silently timing out connections to a single DB and a
single IP address, or a heterogeneous connection pool connected to potentially
multiple IPs/instances of a replicated DB using a DNS front end for load
balancing and one replica fails).
In order to time out those unhealthy connections without blocking access to
other parts of Asterisk that may attempt access to the connection pool, it would
be beneficial to not lock/block access around the entire pool in
_ast_odbc_request_obj2 while doing potentially blocking operations on connection
pool objects such as the connection_dead() test, odbc_obj_connect(), or by
dereferencing a struct odbc_obj for the last time and triggering a
odbc_obj_disconnect().
This would facilitate much quicker and concurrent timeout of dead connections
via the connection_dead() test, which could block potentially for a long period
of time depending on odbc.ini or other odbc connector specific timeout settings.
This also would make rapid failover (in the clustered DB scenario) much quicker.
This patch changes the locking in _ast_odbc_request_obj2() to not lock around
odbc_obj_connect(), _disconnect(), and connection_dead(), while continuing to
lock around truly shared, non-immutable state like the connection_cnt member and
the connections list on struct odbc_class.
Fixes: #465
See UserNote below.
Exposed the existing Hangup AMI action in manager.c so we can use
all of it's channel search and AMI protocol handling without
duplicating that code in dialplan_functions.c.
Added a lookup function to res_pjsip.c that takes in the
string represenation of the pjsip_status_code enum and returns
the actual status code. I.E. ast_sip_str2rc("DECLINE") returns
603. This allows the caller to specify PJSIPHangup(decline) in
the dialplan, just like Hangup(call_rejected).
Also extracted the XML documentation to its own file since it was
almost as large as the code itself.
UserNote: A new dialplan app PJSIPHangup and AMI action allows you
to hang up an unanswered incoming PJSIP call with a specific SIP
response code in the 400 -> 699 range.
If too many ciphers are specified in the PJSIP config,
include the maximum number of ciphers that may be
specified in the user-facing error message.
Resolves: #396
* Allow res_speech to translate the input channel if the
format is translatable to a format suppored by the
speech provider.
Resolves: #129
UserNote: res_speech now supports translation of an input channel
to a format supported by the speech provider, provided a translation
path is available between the source format and provider capabilites.
Per RFC8827:
Implementations MUST NOT implement DTLS renegotiation and MUST
reject it with a "no_renegotiation" alert if offered.
So we disable it when webrtc=yes is set.
Fixes#378
UpgradeNote: The dtls_rekey will be disabled if webrtc support is
requested on an endpoint. A warning will also be emitted.
Commit f66f77f last year prevents the res_pjsip_exten_state and
res_pjsip_mwi modules from unloading due to possible pjproject
asserts if the modules are reloaded. A side effect of the
implementation is that the taskprocessors these modules use aren't
being released. When asterisk is doing a graceful shutdown, it
waits AST_TASKPROCESSOR_SHUTDOWN_MAX_WAIT seconds for all
taskprocessors to stop but since those 2 modules don't release
theirs, the shutdown hangs for that amount of time.
This change allows the modules to be unloaded and their resources to
be released when ast_shutdown_final is true.
Resolves: #379
This commit introduces an extension to the endpoint and relevant
resource sizes for PJSIP, transitioning from its current 40-character
constraint to a more versatile 255-character capacity. This enhancement
significantly overcomes limitations related to domain qualification and
practical usage, ultimately delivering improved functionality. In
addition, it includes adjustments to accommodate the expanded realm size
within the ARI, specifically enhancing the maximum realm length.
Resolves: #345
UserNote: With this update, the PJSIP realm lengths have been extended
to support up to 255 characters.
UpgradeNote: As part of this update, the maximum allowable length
for PJSIP endpoints and relevant resources has been increased from
40 to 255 characters. To take advantage of this enhancement, it is
recommended to run the necessary procedures (e.g., Alembic) to
update your schemas.
res_statsis's app loop sleeps for up to .2s waiting on input
to a channel before re-checking the command queue. This can
cause delays between channel setup and bridge.
This change is to send a SIGURG on the sleeping thread when
a new command is enqueued. This exits the sleeping thread out
of the ast_waitfor() call triggering the new command being
processed on the channel immediately.
Resolves: #362
UserNote: Call setup times should be significantly improved
when using ARI.
Make it possible to start a playback and the calling party
to receive audio on a bridge before the call is connected.
Model the implementation after play_on_channel and deliver a
AST_CONTROL_PROGRESS before starting the playback.
For a PJSIP channel this will result in sending a SIP 183
Session Progress.
The documentation on qualify_timeout does not explicitly state that the timeout
includes any time required to perform any needed DNS queries on the endpoint.
If the OPTIONS response is delayed due to the DNS query, it can still render an
endpoint as Unreachable if the net time is enough for qualify_timeout to expire.
Resolves: #352
res_speech_aeap previously did not register an error handler
with aeap, so it was not notified of a disconnect. This resulted
in SpeechBackground never exiting upon a websocket disconnect.
Resolves: #303
From the gdb information, ast_websocket_read reads a message successfully,
then transport_read is called in the serializer. During execution of pjsip_transport_down,
ws_session->stream->fd is closed; ast_websocket_read encounters an error and exits the while loop.
After executing transport_shutdown, the transport's reference count becomes 0, causing a crash when sending SIP messages.
This was due to pjsip_transport_dec_ref executing earlier than pjsip_rx_data_clone, leading to this issue.
In websocket_cb executeing pjsip_transport_add_ref, this we now ensure the transport is not destroyed while in the loop.
Resolves: asterisk#299
In function rtp_ioqueue_thread_remove counter in ioqueue object is not decreased
which prevents unused ICE TURN threads from being removed.
Resolves: #301
