Re: (ITS#7723) slapd crashes on multi core machines if a search request is *immediately* followed by an unbind

[jsynacek@redhat.com]

On 10/15/2013 01:10 PM, michael.vishchers@7p-group.com wrote:
> It is not the client loop that is multithreading but the ldap server.
> 
> And it is not a misuse of the API but a problem that may be raised by day t=
> o day network problems.
> 
> I've boiled down the problem to a few simple configurations that work (or b=
> etter, fail ;-) with both 2.4.23 and 2.4.36. A tgz file containing a setup =
> with start script and testclient is attached. It should be sufficient to re=
> produce the fault.
> 
> The problem occurs only if we use session variable substitution in the rwm =
> overlay, and only if a search is *immediately* (e.g. caused by network loss=
>  and client timeout) followed by an unbind.
> 

I modified the reproducer a bit (the start script) and find out a few things.
You can find the reproducer I'm using at [1].

Valgrind's helgrind shows some lock problems in the rwm overlay and also in
back-ldap and connection.c. After correcting those the issue seems to be gone.

You can find helgrind logs at [2] (before the fix) and [3] (after).

Also, ElectricFence reveals some problems [4], which I didn't fix yet.

A fix attempt can be found at [5]. I'm not sure if that is a correct fix, or it
just masked the real issue. But I didn't to manage to reproduce the problem
after applying it.

[1] http://jsynacek.fedorapeople.org/openldap/its7723/reproducer/
[2]
http://jsynacek.fedorapeople.org/openldap/its7723/results/slapd1-helgrind-broken.log
[3]
http://jsynacek.fedorapeople.org/openldap/its7723/results/slapd1-helgrind-fixed.log
[4]
http://jsynacek.fedorapeople.org/openldap/its7723/results/slapd1-broken-efence-gdb.txt
[5]
http://jsynacek.fedorapeople.org/openldap/its7723/0001-fix-possible-race-conditions.patch

Cheers,
-- 
Jan Synacek
Software Engineer, Red Hat