[asterisk-dev] [Code Review] 2961: Add channel locking for calls that result in channel snapshot creation

rmudgett reviewboard at asterisk.org
Tue Oct 29 16:57:40 CDT 2013 

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviewboard.asterisk.org/r/2961/#review10030
-----------------------------------------------------------


I think you may be creating just as many bugs as fixing with this complicated locking patch.  The bridge snapshot locking analysis appears to be incomplete.


/branches/12/channels/sig_pri.c
<https://reviewboard.asterisk.org/r/2961/#comment19254>

    Revert this change to send_mcid().  The chan is either already locked or NULL.



/branches/12/include/asterisk/stasis_bridges.h
<https://reviewboard.asterisk.org/r/2961/#comment19255>

    This cannot be honored.  The bridge is either already locked or not.
    
    In fact, this function must lock the bridge itself before calling ast_bridge_snapshot_create() because there exists a caller (publish_blind_transfer()) that does not have the bridge locked.  The ast_bridge_snapshot_create() walks the bridge channels list.



/branches/12/include/asterisk/stasis_bridges.h
<https://reviewboard.asterisk.org/r/2961/#comment19256>

    The only caller of this function must always have the bridge locked.



/branches/12/include/asterisk/stasis_bridges.h
<https://reviewboard.asterisk.org/r/2961/#comment19257>

    The only caller of this function must always have the bridge locked.


- rmudgett


On Oct. 25, 2013, 5:07 p.m., Mark Michelson wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviewboard.asterisk.org/r/2961/
> -----------------------------------------------------------
> 
> (Updated Oct. 25, 2013, 5:07 p.m.)
> 
> 
> Review request for Asterisk Developers.
> 
> 
> Bugs: ASTERISK-22709
>     https://issues.asterisk.org/jira/browse/ASTERISK-22709
> 
> 
> Repository: Asterisk
> 
> 
> Description
> -------
> 
> John Bigelow discovered that if the newly-created atxfer_threeway_nominal test is run in a loop, eventually a test run will cause the UUT instance of Asterisk to crash. Investigation of the crash showed that the crash was being caused while trying to get the string representation of a channel's translation paths while creating a channel snapshot. The attended transfer manager thread was creating the channel snapshot as part of the attended transfer stasis message while in another thread (in this case, the bridge channel thread) the translation paths were being altered to make the channel compatible with others in the bridge.
> 
> It's clear that the issue is that the channel's contents are not lock protected while creating a channel snapshot. Luckily, the channel's lock is held already when updating translation paths. By adding locking, I was able to run the test over 450 times straight with no negative consequences.
> 
> When adding locking, there were two strategies to follow:
> 
> 1) Take advantage of the fact that channel locks are recursive, and just add a single SCOPED_CHANNELLOCK() call to ast_channel_snapshot_create() and call it a day.
> 2) Make an audit of channel-related calls that end up creating a channel snapshot and require locks be held prior to making those calls.
> 
> I like approach 2 more, even though it's harder to pull off. In investigating what paths result in channel snapshots being created, I found that channel locks should be required before calling the following functions:
> 
> * ast_channel_snapshot_create()
> * ast_channel_blob_create()
> * ast_channel_publish_snapshot()
> * ast_publish_channel_state()
> * ast_channel_publish_blob()
> * ast_channel_publish_varset()
> * ast_channel_amaflags_set()
> * ast_channel_callid_set()
> * ast_channel_whentohangup_set()
> * ast_channel_callgroup_set()
> * ast_channel_pickupgroup_set()
> * ast_channel_internal_bridge_set()
> * ast_channel_language_set()
> * ast_channel_accountcode_set()
> * ast_channel_peeraccount_set()
> * ast_channel_linkedid_set()
> * ast_channel_stage_snapshot()
> * ast_setstate()
> * ast_aoc_manager_event()
> * ast_channel_setwhentohangup_tv()
> * ast_channel_setwhentohangup()
> * ast_set_variables()
> 
> All of these functions' documentation have been updated to indicate a precondition of having the channel locked.
> 
> In addition, there are some more complex functions that create channel snapshots but could not easily be made to have the precondition of having the channel locked. These are:
> 
> * ast_bridge_blob_create()
> * ast_publish_attended_transfer_fail()
> * ast_publish_attended_transfer_bridge_merge()
> * ast_publish_attended_transfer_threeway()
> * ast_publish_attended_transfer_app()
> * ast_publish_attended_transfer_link()
> * ast_bridge_publish_enter()
> * ast_bridge_publish_leave()
> * ast_bridge_publish_blind_transfer()
> 
> These functions' documentation now have a precondition of having no channels or bridges locked.
> 
> This review seeks only to make sure that channels are properly locked prior to creating channel snapshots. A similar effort is likely needed to ensure that bridges are locked during bridge snapshot creation.
> 
> 
> Diffs
> -----
> 
>   /branches/12/tests/test_stasis_channels.c 401703 
>   /branches/12/tests/test_cel.c 401703 
>   /branches/12/tests/test_cdr.c 401703 
>   /branches/12/res/res_stasis.c 401703 
>   /branches/12/res/res_pjsip_refer.c 401703 
>   /branches/12/res/res_agi.c 401703 
>   /branches/12/res/parking/parking_manager.c 401703 
>   /branches/12/res/parking/parking_bridge_features.c 401703 
>   /branches/12/res/ari/resource_bridges.c 401703 
>   /branches/12/pbx/pbx_realtime.c 401703 
>   /branches/12/main/stasis_channels.c 401703 
>   /branches/12/main/stasis_bridges.c 401703 
>   /branches/12/main/pickup.c 401703 
>   /branches/12/main/pbx.c 401703 
>   /branches/12/main/endpoints.c 401703 
>   /branches/12/main/dial.c 401703 
>   /branches/12/main/core_unreal.c 401703 
>   /branches/12/main/core_local.c 401703 
>   /branches/12/main/channel.c 401703 
>   /branches/12/main/cel.c 401703 
>   /branches/12/main/bridge_channel.c 401703 
>   /branches/12/main/bridge.c 401703 
>   /branches/12/include/asterisk/stasis_channels.h 401703 
>   /branches/12/include/asterisk/stasis_bridges.h 401703 
>   /branches/12/include/asterisk/channelstate.h 401703 
>   /branches/12/include/asterisk/channel.h 401703 
>   /branches/12/include/asterisk/aoc.h 401703 
>   /branches/12/funcs/func_timeout.c 401703 
>   /branches/12/channels/sig_pri.c 401703 
>   /branches/12/channels/sig_analog.c 401703 
>   /branches/12/channels/chan_vpb.cc 401703 
>   /branches/12/channels/chan_unistim.c 401703 
>   /branches/12/channels/chan_skinny.c 401703 
>   /branches/12/channels/chan_sip.c 401703 
>   /branches/12/channels/chan_pjsip.c 401703 
>   /branches/12/channels/chan_phone.c 401703 
>   /branches/12/channels/chan_oss.c 401703 
>   /branches/12/channels/chan_nbs.c 401703 
>   /branches/12/channels/chan_motif.c 401703 
>   /branches/12/channels/chan_misdn.c 401703 
>   /branches/12/channels/chan_mgcp.c 401703 
>   /branches/12/channels/chan_jingle.c 401703 
>   /branches/12/channels/chan_iax2.c 401703 
>   /branches/12/channels/chan_h323.c 401703 
>   /branches/12/channels/chan_gtalk.c 401703 
>   /branches/12/channels/chan_dahdi.c 401703 
>   /branches/12/channels/chan_console.c 401703 
>   /branches/12/channels/chan_alsa.c 401703 
>   /branches/12/apps/app_voicemail.c 401703 
>   /branches/12/apps/app_userevent.c 401703 
>   /branches/12/apps/app_queue.c 401703 
>   /branches/12/apps/app_meetme.c 401703 
>   /branches/12/apps/app_disa.c 401703 
>   /branches/12/apps/app_dial.c 401703 
>   /branches/12/apps/app_confbridge.c 401703 
>   /branches/12/apps/app_agent_pool.c 401703 
>   /branches/12/addons/chan_ooh323.c 401703 
>   /branches/12/addons/chan_mobile.c 401703 
> 
> Diff: https://reviewboard.asterisk.org/r/2961/diff/
> 
> 
> Testing
> -------
> 
> As stated in the description, I ran the atxfer_threeway_nominal test in a loop. After approximately 450 test runs, there were no negative consequences. Prior to this changeset, running the test 20-50 times would result in a crash.
> 
> 
> Thanks,
> 
> Mark Michelson
> 
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.digium.com/pipermail/asterisk-dev/attachments/20131029/1b2038be/attachment-0001.html>

More information about the asterisk-dev mailing list