LDAP & Kerberos interaction & SIGPIPE

[Peter Shoults]

I am sending this out again as the customer has tested a fix I provided
and determined that it does resolve their issue.  I would like to move
forward with a fix, and the one I am using now is the one mentioned
below with the addition of signal().  If I can get some comments, I
would appreciate it.  Otherwise - I guess I will just proceed to put
this into Solaris code.

Pete

On 09/16/09 10:48, Peter Shoults wrote:
> Hi,
>
> Customer has brought forward an issue they were having with Kerberos and
> LDAP, where LDAP is being used to store the database information for
> Kerberos.  The issue is that if the LDAP server is restarted for any
> reason, then Kerberos does not automatically resync back with the LDAP
> server when the LDAP server is back up and running.  Specifically, one
> can run and login into kadmin, but any commands that are run will fail
> with the error:
>
> "Communication failure with server while retrieving list."
>
> It turns out if the user exits from kadmin and logs back in a second
> time, then the command do work fine.
>
> I have determined that the cause of this problem is that when the LDAP
> server is restarted, all the connections we have on port 636 to the LDAP
> server go into a CLOSE_WAIT/FIN_WAIT_2 state.  When we log into kadmin,
> we attempt to contact the LDAP server on these connections, and we
> received SIGPIPE in response to our writes.  Here is a snippet from truss:
>
> 3200/1:         57.2401 write(14, 0x0010B810, 23)                      
> Err#32 EPIPE
> 3200/1:                             150301\012941A 60F Y P87A7BE9318B6
> c8C |0F   v
> 3200/1:         57.2404     Received signal #13, SIGPIPE [caught]
>
> This is fine - the sig_pipe handler is invoked and we do print out the
> syslog message.  However, we never reset the signal disposition for
> SIGPIPE.    kadmind process immediately proceeds to try the next
> connection to the LDAP server, and again gets SIGPIPE.  This time
> though, the default handler is invoked, which terminates kadmind.  At
> this point, SMF realizes kadmind has died and restarts it, which
> re-establishes all our connections to the LDAP server and that explains
> why a subsequent login to kadmin will work.
>
> I have two questions about this.  The first why do we have a handler for
> SIGPIPE in the kadmin code, unlike the krb5kdc code, which sets SIGPIPE
> disposition to SIG_IGNORE.  This handler in the kadmin code has not
> changed in a long long time.  I tested setting SIGPIPE to SIG_IGN and
> this does allow a user to enter commands into kadmin after LDAP server
> restarts and run commands without issue.
>
> Assuming we have the SIGPIPE handler specifically to output the syslog
> message, then I propose that we have in the handler a resetting of the
> signal disposition to sig_pipe.  I have also tested this fix and
> verified that this also resolves the problem and allows the user to
> enter kadmin commands after LDAP server restarts.  Here is my change:
>
> file modified is ovsec_kadmd.c
>
> void
> sig_pipe(int unused)
> {
> +        signal(SIGPIPE, sig_pipe);
>         krb5_klog_syslog(LOG_NOTICE, gettext("Warning: Received a SIGPIPE; "
>                 "probably a client aborted.  Continuing."));
> }
>
>
> Pete
>
>