Re: [naviserver-devel] "pthread_join failed in Ns_ThreadJoin: Invalid argument"

[Gustaf N. <ne...@wu...>]

Am 26.03.15 um 13:11 schrieb David Osborne:
> Hi Gustaf,
>
> The crashes we are seeing are not during shutdown, but during normal 
> server operation which is why we are concerned about them. The system 
> recovers quickly due to daemontools, but the scheduled tasks running 
> at the time of the crash are aborted.
can you check, what thread was shutting down (in your backtrace: 
0x7f585181dda0)
what value for connsperthread are you using?
>
> There is nothing in the syslog to indicate a memory problem or oom 
> killer being invoked.
>
> As for a Tcl bug, we might potentially be able to upgrade the version 
> of Tcl use to 8.5.17 from the Debian testing distribution (currently 
> we use 8.5.11). But I wouldn't like to do that on the production 
> system unless there was very good reason.
there is no change regarding to this problem in the Tcl 8.5.* family
>
> Do you know more about the Tcl bug that caused your last crash?
I have to dig this out. The problem was with the memory management of 
tcl-objs,
where appending to a tcl-obj could lead to overshoot the capabilities of 
the length field.
This was fixed in tcl 8.6 and tcl 8.5 (at least 8.5.15). If you were hit 
by this
bug, the backtrace would look very differently.

-g
>
>
>
> On 26 March 2015 at 11:18, Gustaf Neumann <ne...@wu... 
> <mailto:ne...@wu...>> wrote:
>
>     Hi David,
>
>     We have this issue of shutdown in test-cases and during server
>     shutdown since
>     several years. It is annoying, but mostly harmless. The last time
>     i looked into
>     the problem, i got the impression that it depends on a not fully
>     predictable
>     shutdown order, which is not completely in our hands due to tcl
>     interactions.
>
>     i would be alert, if you see this problem in other situations. For
>     our production
>     system, i saw the last crash months ago (happend due to a Tcl-bug
>     with
>     doubling the size of a dstring over 2gb). Another "crash" on a
>     different system
>     turned out to be due to linux's oom killer.
>
>     -g
>


-- 
Univ.Prof. Dr. Gustaf Neumann
WU Vienna
Institute of Information Systems and New Media
Welthandelsplatz 1, A-1020 Vienna, Austria