Bug 798627

Description of problem:

Since upgrading to glibc-2.5-65.el5_7.3 we noticed that gethostbyname() can now return duplicate addresses - if the name appears in /etc/hosts multiple times.

Machines running with glibc-2.5-58.el5_6.4 do not show this behaviour.

I've not tested glibc-2.5-81 yet but the announcement https://rhn.redhat.com/errata/RHBA-2012-0260.html does not mention anything obviously related.

Note in the test setup we have nsswitch set to look in /etc/hosts first and host.conf has multi set to on, but so do the systems with the previous glibc which do no show the new behaviour.

Running with/without nscd doesn't affect this other than that it caches the results.

Version-Release number of selected component (if applicable):

glibc-2.5-65.el5_7.3

How reproducible:

100%

Steps to Reproduce:

Here is my test procedure.

I wrote a trivial test script ~/scripts/gethostbyname.pl :

#! /usr/bin/perl
#
use Socket;
foreach $i (@ARGV) {
    ($name,$aliases,$addrtype,$length,@addrs)=gethostbyname($i);
    printf "%-25s %s\n", $name, $i;
    foreach $ad (@addrs) {
        print " ".inet_ntoa($ad)."\n";
    }
}

On a machine running the newer glibc:

# grep multi /etc/host.conf
multi on
# grep '^host' /etc/nsswitch.conf
hosts:      files dns
# service nscd stop
Stopping nscd:                                  [  OK  ]
# grep beast /etc/hosts
# ~jp107/scripts/gethostbyname.pl beast
                          beast
# echo "10.6.6.6 beast" >> /etc/hosts
# grep beast /etc/hosts| wc -l
1
# ~jp107/scripts/gethostbyname.pl beast
beast                     beast
 10.6.6.6
# echo "10.6.6.6 beast" >> /etc/hosts
# grep beast /etc/hosts| wc -l
2
# ~jp107/scripts/gethostbyname.pl beast
beast                     beast
 10.6.6.6
 10.6.6.6
# echo "10.6.6.6 beast" >> /etc/hosts
# grep beast /etc/hosts| wc -l
3
# ~jp107/scripts/gethostbyname.pl beast
beast                     beast
 10.6.6.6
 10.6.6.6
 10.6.6.6
#
...

Actual results:

duplicate (idential) addresses returned

Expected results:

each unique address returned only once.

Additional info:

The 'old' glibc behaviour is what we also get on Solaris.  It is possible that gethostbyname() was always intended to return the address multiple times, and the previous behaviour was a (very longstanding) bug.

We originally noticed this because a script we use to generate /etc/host sticks some names in twice.  Apache got confused if a "listen name:port" where gethostbyname(name) returns the same address multiple times - it tries to bind to the same address/port multiple times which fails (and the apache error message doesn't make it clear what is going wrong, but that isn't glibc's fault)...

We have corrected our script to avoid those duplicates, but the change of behaviour in glibc might affect others.