Corey Farrell [Tue, 21 Nov 2017 05:10:58 +0000 (00:10 -0500)]
loader: Rework of load_dynamic_module.
* Split off load_dlopen to perform actual dlopen, check results and log
warnings when needed.
* Use flags which minimize number of calls to dlopen required. First
attempt always uses RTLD_GLOBAL when global_symbols_only is enabled,
RTLD_LOCAL when it is not.
This patch significantly reduces the number of dlopen's performed. With
299 modules my system ran dlopen 857 times before this patch, 655 times
after this patch.
Corey Farrell [Wed, 22 Nov 2017 02:34:56 +0000 (21:34 -0500)]
loader: Minor fix to module registration.
This protects the module loader itself against crashing if dlopen is
called on a module from outside loader.c.
* Expand scope of lock inside ast_module_register to include reading of
resource_being_loaded.
* NULL check resource_being_loaded.
* Set resource_being_loaded NULL as soon as dlopen returns. This fixes
some error paths where it was not NULL'ed.
* Create module_destroy function to deduplicate code from
ast_module_unregister and modules_shutdown.
* Resolve leak that occured if a module did not successfully register.
* Simplify checking for successful registration.
Richard Mudgett [Tue, 12 Dec 2017 00:20:06 +0000 (18:20 -0600)]
res_rtp_asterisk.c: Disable packet flood detection for video streams.
We should not do flood detection on video RTP streams. Video RTP streams
are very bursty by nature. They send out a burst of packets to update the
video frame then wait for the next video frame update. Really only audio
streams can be checked for flooding. The others are either bursty or
don't have a set rate.
* Added code to selectively disable packet flood detection for video RTP
streams.
Joshua Colp [Thu, 30 Nov 2017 16:12:55 +0000 (16:12 +0000)]
AST-2017-012: Place single RTCP report block at beginning of report.
When the RTCP code was transitioned over to Stasis a code change
was made to keep track of how many reports are present. This count
controlled where report blocks were placed in the RTCP report.
If a compound RTCP packet was received this logic would incorrectly
place a report block in the wrong location resulting in a write
to an invalid location.
This change removes this counting logic and always places the report
block at the first position. If in the future multiple reports are
supported the logic can be extended but for now keeping a count
serves no purpose.
Joshua Colp [Wed, 6 Dec 2017 14:24:03 +0000 (14:24 +0000)]
pjsip: Ignore state changes from old transactions.
When we fail over to a new target we create a new transaction
and it becomes the current INVITE transaction. This does not
prevent the previous transaction from raising state changes
and causing the session to be prematurely disconnected if a
transport error occurs immediately.
This change backports a fix from PJSIP that eliminates the
incorrect state change and reduces when they would be raised
in the first place.
Ivan Poddubny [Tue, 12 Dec 2017 21:38:01 +0000 (22:38 +0100)]
app_queue: Fix extension state subscriptions removed on dialplan reload
The approach with having a single global subscription to all extension
state changes has one issue: dynamically created hints don't have any
watchers and are therefore garbage collected on the first dialplan
reload.
This change creates a state subscription for every queue member with a
hint as state_interface, thus increasing the count of watches for
hints, so they are not destroyed prematurely anymore.
There are 2 side effects:
1. The state change callback in app_queue is not executed when
there are no members referring to the extension.
2. The callback is called multiple times for the same hint if it's
associated with more than one queue member.
Corey Farrell [Tue, 12 Dec 2017 21:16:38 +0000 (16:16 -0500)]
menuselect: Tweak check for recently run configure.
Recently menuselect has randomly produced an error stating that
configure was just run and make had to be restarted. I believe this is
due to an incorrect menuselect/Makefile rule. The original rule
produced an error if makeopts or autoconfig.h were older than
makeopts.in or autoconfig.h.in. I believe this can create an issue if
makeopts is older than autoconfig.h.in or if autoconfig.h is older than
makeopts.in. The new rules compare files independently.
This patch does three things associated with the initial incoming INVITE
request URI.
1) Add access to the full initial incoming INVITE request URI.
2) We were not setting DNID on incoming PJSIP channels. The DNID is the
user portion of the initial incoming INVITE Request-URI. The value is
accessed by reading CALLERID(dnid).
3) Fix CHANNEL(pjsip,target_uri) documentation.
* The initial incoming INVITE request URI is now available using
CHANNEL(pjsip,request_uri).
* Set the DNID on PJSIP channel creation so CALLERID(dnid) can return the
initial incoming INVITE request URI user portion.
* CHANNEL(pjsip,target_uri) now correctly documents that the target URI is
the contact URI.
* Refactored print_escaped_uri() out of channel_read_pjsip() to handle
pjsip_uri_print() error condition when the buffer is too small.
Sean Bright [Thu, 7 Dec 2017 20:19:40 +0000 (15:19 -0500)]
astdb: Improve prefix searches in astdb
Using the LIKE operator requires a full table scan of 'astdb', whereas a
comparison operation is able to use the primary key index.
This patch adds a new function to the AstDB API for quick prefix matches
and updates res_sorcery_astdb to utilize it. This showed substantial
performance improvement in my test environment.
Related to ASTERISK~26806, but does not completely resolve it.
Corey Farrell [Sat, 9 Dec 2017 00:19:34 +0000 (19:19 -0500)]
loader: Refactor resource_name_match.
Optimize resource_name_match. This change eliminates use of
ast_strdupa, instead verifying that both basename's are the same length,
then using strncasecmp.
Sean Bright [Thu, 7 Dec 2017 15:52:39 +0000 (10:52 -0500)]
utils: Add convenience function for setting fd flags
There are many places in the code base where we ignore the return value
of fcntl() when getting/setting file descriptior flags. This patch
introduces a convenience function that allows setting or clearing file
descriptor flags and will also log an error on failure for later
analysis.
Corey Farrell [Fri, 8 Dec 2017 01:41:14 +0000 (20:41 -0500)]
res_stasis and res_speech: Fix load order.
res_stasis was missing AST_MODFLAG_LOAD_ORDER. Set res_stasis and
res_speech to start at (AST_MODPRI_APP_DEPEND - 1) so they are ready for
dependent modules.
Kevin Harwell [Fri, 8 Dec 2017 00:22:34 +0000 (18:22 -0600)]
pjsip_options: contacts sometimes not being updated on reload
For both dynamic and static contacts it was possible that potential AOR
changes were not being applied to all contacts. This was because the qualify
and schedule code was only retrieving AOR's, and contacts with frequencies
greater than zero.
For instance the following could happen: and AOR/contact has a frequency of 5,
it then gets set to 0, and then a reload occurs. All scheduled OPTIONS are
stopped, a list of AOR's is retrieved with frequency > 0, but none are
selected since in this scenario all are 0. The contact for the one previously
set to 5 though does not get updated, so it's status remains "AVAILABLE".
This patch makes it so all contacts (static and dynamic) are selected, and
appropriately updated if need be.
Kevin Harwell [Fri, 8 Dec 2017 00:18:00 +0000 (18:18 -0600)]
pjsip_options: dynamic contact's fields not updated on reload
Dynamic contacts were not being properly updated on reload. As a matter of
fact any changes to the AOR that a dynamic contact was associated with were
not being applied.
On reload, this patch makes it so for each dynamic contact, the associated
AOR is now retrieved and the AOR's fields are applied to the contact.
Corey Farrell [Thu, 7 Dec 2017 16:35:39 +0000 (11:35 -0500)]
CLI: Fix remote console completion.
Duplicate checking was done incorrectly when parsing completion options
from a remote console causing all options to be ignored as duplicates.
Once fixed I had to separate processing of the best match to ensure it
was not identified as a duplicate when it is the only match.
Corey Farrell [Wed, 6 Dec 2017 20:49:32 +0000 (15:49 -0500)]
sounds_index: Avoid repeatedly reindexing.
The sounds index is rebuilt each time a format is registered or
unregistered. This causes the index to be repeatedly rebuilt during
startup and shutdown.
This patch significantly reduces the work done by delaying sound index
initialization until after modules are loaded. This way a reindex only
occurs if a format module is loaded after startup. We also skip
reindexing when format modules are unloaded during shutdown.
Corey Farrell [Wed, 6 Dec 2017 18:42:55 +0000 (13:42 -0500)]
media_index: Improve startup.
This eliminates some wasteful operations in media_index startup.
* Replace statically set string-fields with char[0].
* Eliminate pointless RAII_VAR's.
* alloc_variant: Avoid pointless ao2_find on new info->variant.
* Stop trying find_variant before alloc_variant.
* process_media_file: replace ast_str with ast_asprintf. This avoids
reallocation of file_id_str.
Overall sounds_index.c is about 27% of Asterisk startup time when using
sample configs. This patch reduces it to 20%. This is a half-fix. The
real problem is that the media_index is regenerated repeatedly - 68
times in my test.
Richard Mudgett [Wed, 6 Dec 2017 00:04:47 +0000 (18:04 -0600)]
CDR: Fix deadlock setting some CDR values.
Setting channel variables with the AMI Originate action caused a deadlock
when you set CDR(amaflags) or CDR(accountcode). This path has the channel
locked when the CDR function is called. The CDR function then
synchronously passes the job to a stasis thread. The stasis handling
function then attempts to lock the channel. Deadlock results.
* Avoid deadlock by making the CDR function handle setting amaflags and
accountcode directly on the channel rather than passing it off to the CDR
processing code under a stasis thread to do it.
* Made the CHANNEL function and the CDR function process amaflags the same
way.
* Fixed referencing the wrong message type in cdr_prop_write().
Richard Mudgett [Wed, 6 Dec 2017 13:36:02 +0000 (07:36 -0600)]
bridge_basic.c: Update transfer diagnostic messages addendum.
* Added start DTMF transfer verbose messages.
* Made associated transfer messages use a similar message format.
* Adjusted message verbose level as requested by initial reporter.
The SuccessfulAuth using_password field was declared as a pointer to a
uint32_t when the field was later read as a uint32_t value. This resulted
in unnecessary casts and a non-portable field value reinterpret in
main/security_events.c:add_json_object(). i.e., It would work on a 32 bit
architecture but not on a 64 bit big endian architecture.
Sungtae Kim [Mon, 4 Dec 2017 09:40:18 +0000 (10:40 +0100)]
Add new object for VoicemailUserEntry
Currently, when the app_voicemail sending VoicemailUserEntry AMI event, there's
no OldMessageCount info for default.
To check the OldMessageCount info, it required IMAP_STORAGE define, but this is
not correct.
Added OldMessageCount item as a default.
More complicated direct media reinvite negotiations can result in longer
delays before direct media flows. The strictrtp learning timeout time
was too short. One log showed that the first RTP packet came in just
after three seconds.
* Increase the strictrtp learning timeout time from 1.5 to 5 seconds.
Alexander Traud [Mon, 4 Dec 2017 14:33:16 +0000 (15:33 +0100)]
res_rtp_asterisk: Correct default in sample configuration file.
With Asterisk 12 (commit 866d968), the default of "icesupport" changed to
- "yes" in the module "res_rtp_asterisk" and
- "no" in the module "chan_sip".
The latter was reflected in the sample configuration file for "sip.conf". The
former did not make it into "rtp.conf.sample".
Joshua Colp [Mon, 4 Dec 2017 00:49:14 +0000 (00:49 +0000)]
pjproject: Clean up disabling of WebRTC support.
The definition in config_site.h and the argument to the
configure script are not necessary to disable WebRTC
support. The correct argument, --disable-libwebrtc, is
already passed.
Corey Farrell [Sat, 2 Dec 2017 21:55:31 +0000 (16:55 -0500)]
autoconf: Remove use of m4_ifblank.
The m4_ifblank macro is not available on CentOS 6, reverse conditionals
to allow use of m4_ifval instead. ./bootstrap.sh was run but this patch
does not result in any difference to the generated configure script.
George Joseph [Thu, 30 Nov 2017 20:38:50 +0000 (13:38 -0700)]
AST-2017-013: chan_skinny: Call pthread_detach when sess threads end
chan_skinny creates a new thread for each new session. In trying
to be a good cleanup citizen, the threads are joinable and the
unload_module function does a pthread_cancel() and a pthread_join()
on any sessions that are active at that time. This has an
unintended side effect though. Since you can call pthread_join on a
thread that's already terminated, pthreads keeps the thread's
storage around until you explicitly call pthread_join (or
pthread_detach()). Since only the module_unload function was
calling pthread_join, and even then only on the ones active at the
tme, the storage for every thread/session ever created sticks
around until asterisk exits.
* A thread can detach itself so the session_destroy() function
now calls pthread_detach() just before it frees the session
memory allocation. The module_unload function still takes care
of the ones that are still active should the module be unloaded.
Sean Bright [Fri, 1 Dec 2017 16:01:01 +0000 (11:01 -0500)]
config: Speed up config template lookup
ast_category_get() has an (undocumented) implementation detail where it
tries to match the category name first by an explicit pointer comparison
and if that fails falls back to a normal match.
When initially building an ast_config during ast_config_load, this
pointer comparison can never succeed, but we will end up iterating all
categories twice. As the number of categories using a template
increases, this dual looping becomes quite expensive. So we pass a flag
to category_get_sep() indicating if a pointer match is even possible
before trying to do so, saving us a full pass over the list of current
categories.
In my tests, loading a file with 3 template categories and 12000
additional categories that use those 3 templates (this file configures
4000 PJSIP endpoints with AOR & Auth) takes 1.2 seconds. After this
change, that drops to 22ms.
Sean Bright [Fri, 1 Dec 2017 14:29:43 +0000 (09:29 -0500)]
config: Speed up ACO & sorcery initialization
When starting Asterisk in the foreground, there is a perceptible delay
when loading modules that use the ACO and sorcery config frameworks.
For example, a lightly configured res_pjsip took 853ms to load on my
VM.
I tracked down the slowness to the XPath queries used to associate the
relevant documentation with the config options. One improvement was
adding a call to xmlXPathOrderDocElems after loading an XML document.
From the libxml2 docs:
Call this routine to speed up XPath computation on static documents.
The second change was to remove recursive descent and wildcard
operators from the XPath queries. After these changes, res_pjsip takes
85ms to load on my VM and there is no longer a perceptible delay when
starting Asterisk in the foreground.
projects / thirdparty / asterisk.git / log
