| Summary: | ASTERISK-20682: deadlock when servicing MWI event | ||
| Reporter: | Andrew Parisio (parisioa) | Labels: | |
| Date Opened: | 2012-11-14 12:55:34.000-0600 | Date Closed: | 2012-12-31 11:07:08.000-0600 |
| Priority: | Critical | Regression? | |
| Status: | Closed/Complete | Components: | Channels/chan_sip/Subscriptions |
| Versions: | SVN | Frequency of Occurrence | Occasional |
| Related Issues: | |||
| Environment: | 1.8 svn | Attachments: | ( 0) locks.txt ( 1) threads.txt |
| Description: | This is a 2 or 3 week old copy of 1.8/svn. We've had 3 or so lock ups in the last 2 weeks, have to kill -9 it. Will gdb bt full next time it crashes (didn't know you could do it on a live instance).
Thanks! | ||
| Comments: | By: Mark Michelson (mmichelson) 2012-11-14 13:50:05.542-0600 I've attached ASTERISK-20682.patch for you to try using. It moves event subscription callbacks to a point where there is no lock being held. In addition, I've made event subscriptions ref-counted in the slim chance that the event subscription gets destroyed in between us finding the event and calling the callback. By: Mark Michelson (mmichelson) 2012-11-14 14:45:30.446-0600 Richard Mudgett has pointed out a possibility for this patch to be dangerous since it can result in referencing a freed structure. In other words, this could turn your deadlock into a crash given the right circumstances. I'm going to remove the patch as a result. By: Rusty Newton (rnewton) 2012-11-15 18:39:47.683-0600 Backtrace may help further, https://wiki.asterisk.org/wiki/display/AST/Getting+a+Backtrace#GettingaBacktrace-GettingInformationForADeadlock Thanks! By: Andrew Parisio (parisioa) 2012-12-19 18:07:30.349-0600 No crashes on an SVN copy about 3 weeks old. I think this is fixed. Thanks! | ||