Why a logical IP is marked as DEPRECATED?

Reported Firewall Problems

We've had many questions, comments and complaints about IP address "problems" when using highly available services in a Sun Cluster environment. We found out that most, if not all of these were related to configurations where firewalls were configured between the service running on the cluster, and the clients connecting to the cluster.

So, what is the problem? The firewall administrators often make the assumption that a packet sent from a client to the logical IP address of an HA service will generate a response IP packet with exactly the same logical IP address as the source address. So, they configure an appropriate firewall rule and wonder why this rule does not work, i.e., instead there were IP packets coming back from an HA service that did not match this rule.

Then they start researching the network configuration on the cluster node that hosts the HA service and find out that the logical IP address used by that service was set to a state called "DEPRECATED". And they think this is the root cause of their problem - which (we think) is not the case.

How does Address Selection really work?

As address selection can become very complicated in complex network setups, the following will be true for the typical simple network setup found at most installations.

Let's look at the address selection for an outgoing packet a bit more closely. First we must make a distinction between TCP (RFC 793) and UDP (RFC 768). TCP is a connection-oriented protocol, i.e. a connection is established between a client and a service. Using this connection, source and target addresses are always used appropriately; in a Sun Cluster environment the source address of a packet sent by the service to a client will usually be the logical IP address of that HA service - but only if the client used the logical service address to send its request to the service.

So, this will not cause any problems with firewalls, because you know exactly which IP addresses will be used as source addresses for outgoing IP addresses.

Let's look into UDP now. UDP is a connectionless protocol, i.e., there is no established connection between a client and a server (service). A UDP-based service can choose its source address for outgoing packets by binding itself to a fixed address, but most services don't do this. Instead, they accept incoming packets from all network addresses configured. For those readers who are familiar with network programming, the typical code segment has the following lines in it:


struct sockaddr_in address;
...
address.sin_addr.s_addr = INADDR_ANY;
...
if (bind (..., (struct sockaddr \*) &address, ...) == 0)

Using this typical piece of code, the UDP service listens on all configured IP addresses, and the outbound source address is set by the IP layer and the choosing algorithm is complex and cannot be influenced. Details can be read in Infodoc 204569 (access on SunSolve for SPECTRUM contract holders only); but we think they are not that relevant here, except for this quote: "IP addresses associated with interfaces marked as DEPRECATED will not normally be used as source addresses by IP unless deprecated interfaces are all that is available, in which case they will be used."

The DEPRECATED flag

So, now DEPRECATED comes into play. A DEPRECATED address will - normally - not be used as a source address! First, why does Sun Cluster set HA IP addresses, i.e. logical or floating addresses into state DEPRECATED? Because they are floating addresses - there is no guarantee that they will stay on one node. In failure situations an HA IP address will float to another node together with its service. Or if the administrator decides to migrate a service; or when the service is stopped, the logical IP address might disappear on one node.

Let's have a look at services where IP communication is initiated from a cluster node. E.g. a cluster node might try to mount an external NFS share on this node temporarily. Whether this is UDP or TCP based NFS would not matter in this case! The IP layer would choose a source address; it could be the logical IP address of an HA service that happens to run on the same system - if it were not DEPRECATED. Now, imagine the NFS mount is successful, is using the logical IP address and NFS transfers work fine. Now, the HA service that owns the HA IP address is switched to another node in the cluster; its IP address would also switch. What would happen to the NFS traffic between this node and the external NFS server? It would fail. Packets coming from the NFS server would reach a different node now; namely that of the HA service that switched, taking its IP address with it. (And the NFS client on the cluster node would fail as well.....)

So, that is the reason for setting the DEPRECATED flag on HA IP addresses; remember the quote above: "...marked as DEPRECATED will not normally be used...". Although not setting the DEPRECATED flag would improve the probability that the address potentially be used by the IP layer as a source address, there is no guarantee and at the end, this would not help. But the DEPRECATED flag helps to prevent major problems on cluster nodes.

The Solution

Back to the original question: how can I make my firewall rules work? There are 4 possibilities - in prioritized order, best practice first:

  1. change your firewall rules to accept all possible addresses from the nodes where packets could be originating from;
  2. change your service, by binding only to the HA service IP address - which is only possible if its configuration lets you do this or if you have access to the source code;
  3. move your HA service into a Solaris 10 container, that only uses the logical IP address; in this configuration the logical IP address will always be used as source address, even though it is in state DEPRECATED;
  4. try to manipulate the decision process of the IP layer - which is a very bad idea.

To summarize

Sun Cluster sets the DEPRECATED flag on HA service IP addresses by design and it is a good thing, as it prevents strange problems with IP based clients on cluster nodes to happen. Not setting it, would not solve the problems reported.

Hartmut Streppel
Principal Field Technologist
Systems Practice
Comments:

Hartmut,

In many years of clarifying these points in discussions of SCX.x clustering and delivering training on the same, one of the best write-ups of the current state of the IPMP implementation of Solaris I've come across is the high-level design doc for Clearview (http://www.opensolaris.org/os/project/clearview/ipmp-highlevel-design.pdf).

In highlighting the inadequacies of the current implementation and identifying where the new implementation should (will) improve it, they've beautifully shone the light upon a number of existing pathologies. Fantastic reading for better understanding of the current, yet also the new architecture to come ...

Roll on b107's integration of Clearview ... ;-)

HTH

Craig

Posted by Craig Morgan on March 02, 2009 at 05:58 AM PST #

Can it possible that we can prevent this issue without configuring the SO based IPMP ?

Any Suggestion ?

Thanks
M|Tanvir

Posted by Mohammed Tanvir Alam on March 29, 2009 at 08:37 AM PDT #

I am not sure I understand your question, especially "SO based IPMP". It is my understanding that only points a,b and c in the "Solution" paragraph can solve this issue.
Hartmut

Posted by Hartmut Streppel on March 29, 2009 at 06:55 PM PDT #

an libc based wrapper or kernel net namespaces (not sure if this linux feature exists on Solaris outside of zones) could make the default ip selection configurable for a process.

And BTW: outgoing TCP connections also have this problem. Especially nasty in RMI scenarios.

Posted by Bernd Eckenfels on May 05, 2009 at 03:24 PM PDT #

There is fourth way to force service to choose logical IP address. It is enough to close service in zone, and configure zones as cluster nodes. Do not assign any IP address to the zone. So zone will get one address (logical one) after switching our group to it.
Disadvantage is that when zone does not have any running group we can only access it from zlogin on physical host.

Posted by guest on July 26, 2009 at 04:55 AM PDT #

Putting the service into a container is a very good idea. If that container only has the logical IP address visible, this will be taken as source address; even if it is marked DEPRECATED. I'll add that to my article.

Posted by Hartmut Streppel on July 28, 2009 at 11:54 PM PDT #

Post a Comment:
  • HTML Syntax: NOT allowed
About

mkb

Search

Archives
« April 2014
SunMonTueWedThuFriSat
  
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
   
       
Today