[Dovecot] v2.1.10: Director director_servers order issue

[Bertrand Jacquin]

Hi,

I'm setting up an IMAP/POP cluster using dovecot director for some NFS
reasons and I'm getting some troubles with director_servers directive.

Configuration is shared on both nodes :

director_servers      = 10.0.50.50    10.0.50.51
director_mail_servers = 192.168.0.150 192.168.0.151
director_user_expire = 15 min

service director {
  unix_listener login/director {
    mode = 0666
  }
  fifo_listener login/proxy-notify {
    mode = 0600
    user = $default_login_user
  }
  unix_listener director-userdb {
    mode = 0600
  }
  inet_listener {
    port = 9090
  }
}

service ipc {
  unix_listener ipc {
    user = $default_login_user
  }
}

Here 10.0.50.50 is node1, 10.0.50.51 is node2

# ring status on node1
 director ip port type last failed
 10.0.50.50  9090 self never
 10.0.50.51  9090      never

# ring status on node2
 director ip port type last failed
 10.0.50.50  9090 self never
 10.0.50.51  9090      never

self is the same on both nodes, and that generate some cases that can
be observed on logs :

# logs on node1
 dovecot: director: Error: Remote director thinks it's someone else (connected to 10.0.50.51:9090, remote says it's 10.0.50.50:9090)

And this is true, TCP packet contain bad DATA :

 # tcpdump -i eth0 -nn src host 10.0.50.51 and port 9090 -s 0 -w - -l | strings | egrep '^ME'
 ME	10.0.50.50	9090

# logs on node2
 dovecot: director: Error: connect(10.0.50.51:9090) failed: Invalid argument

Invalid argument is present as bind() is not done with the appropriate
IP :

 # strace -p 6063 -fF -s 1024 -e trace=bind,connect
 bind(28, {sa_family=AF_INET, sin_port=htons(0), sin_addr=inet_addr("10.0.50.50")}, 16) = 0
 connect(28, {sa_family=AF_INET, sin_port=htons(9090), sin_addr=inet_addr("10.0.50.51")}, 16) = -1 EINVAL (Invalid argument)

Also, during this time, CPU is ~100% used.

bind() is done with node1 IP address.

I also tested using a DNS :

 director_servers = director-all

director-all is two A entries, first containing 10.0.50.50, second
10.0.50.51 and the result here is exactly the same.

Using two configurations :

# director_servers on node1
 director_servers = 10.0.50.50 10.0.50.51

# director_servers on node2
 director_servers = 10.0.50.51 10.0.50.50

Give the same result as showed before :

# ring status on node1
 director ip port type last failed
 10.0.50.50  9090 self never
 10.0.50.51  9090      never

# ring status on node2
 director ip port type last failed
 10.0.50.50  9090 self never
 10.0.50.51  9090      never

It seems that first sorted entry in director_servers is considered as
"self".

I'm using dovecot 2.1.10-0~auto+55 debian package from rename-it
repository.

Thanks for your help.

-- 
Beber
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: not available
URL: <http://dovecot.org/pipermail/dovecot/attachments/20121120/3cad8bb1/attachment-0002.bin>