DNS Scavenging Revealed

I thought I'd write up how scavenging works in DNS in conjunction with DHCP. There are places that describe the process well, but not everywhere, and nowhere that I've seen has it quite right in terms of timings.

DNS has 2 intervals relating to aging: the no refresh interval and the refresh interval:

No refresh interval - during this time, no refreshes may be made to the timestamp of a DNS record (more about refreshes in a moment), but updates are ok. This is done to prevent unnecessary frequent replication of DNS record changes.

Refresh interval - during this time, the timestamp of the record may be refreshed, at which point the no refresh interval kicks in again.

After the refresh interval has passed, if there has been no refresh (or update) of that record then it is available to be scavenged.

How to set up scavenging:
On the properties of the DNS zone, click on "Aging". Set whatever no refresh and refresh intervals you want. Ensure that you also tick the box to "scavenge stale resource records".
You then also need to tell the server to actually scavenge periodically. I am in favour of a single DC doing the scavenging.
In the properties of the DNS server, on the Advanced tab, check the box to "enable automatic scavenging of stale records", and set the scavenging period. Don't make this too frequent as scavenging hits the server.
Note that if you right click on the DNS server, there is a "set aging/scavenging for all zones..." option. This will, if you wish, allow you to set consistent scavenging options across all hosted zones. I don't use it personally.

That's the easy bit and what you can find on any website.

DNS record Refresh vs Update
A refresh is defined as when the record tries to update its timestamp but its IP address hasn't changed.
An update is when the IP address changes. This kind of operation can always occur as the no refresh interval does not impact on it.

So, let's see the processes in action:
Firstly, DNS refreshes being prevented by the no refresh interval:
The scenario here is that we have a DHCP server, with DHCP lease times set to 10 mins.
This means that DHCP half-life (T1) will occur every 5 mins.
Scavenging is enabled and scheduled every hour, with no refresh interval set to 1 hour, and refresh interval set to 2 hours:

1. Scavenging is occurring at 25 past each hour.
2. Client powers on at 09:44.
3. The timestamp for the record is set to 09:00:00 because all timestamps are rounded down to the nearest hour.
4. At 09:49, it attempts to refresh its timestamp in DNS. It is not allowed to do so due to it being in the no refresh interval.
5. This goes on every 5 minutes. At 10:04, it tries again. Now, logic dictates that it should be able to refresh its timestamp this time because the no refresh interval is only 1 hour. However (and this is what I've never seen written anywhere else), it can't. This is because, all DNS registrations that occur from 09:00 to 09:59 will have a timestamp of 09:00:00. Therefore, counting only commences at 10:00, so the no refresh interval is officially from 10:00 - 11:00 (and any time before 10:00). Therefore the record timestamp will be refreshed on the first attempt after 11:00 and will be set to 11:00:00.

Now, let's see scavenging in action in the same scenario:

1. Scavenging still occurs at 25 past the hour.
2. Client is powered on at 13:38. DNS timestamp is set to 13:00:00.
3. It tries to refresh its timestamp every 5 mins but is unable to.
4. At 14:16 it is powered down.
5. At 15:00, its DNS record enters the refresh interval. However, the client is powered down so does not refresh it.
6. At 17:00 (2 hour refresh interval), the refresh interval expires and the record becomes eligible for scavenging.
7. At 17:25, scavenging runs and the record is deleted.

Matching DHCP and DNS scavenging periods correctly is vital.

Here's an example where they aren't matched. DCs re-register their records every 24 hours by default, or whenever Netlogon is restarted. All clients automatically attempt to refresh their DNS records every 24 hours too.

In this scenario, we have the same scavenging settings as above, but this time, DHCP lease times are set to 4 days.

1. Scavenging occurs at 25 past the hour again.
2. Client powers on at 08:12 and receives an IP address from DHCP. DNS timestamp is set to 08:00:00. Since the lease time is 4 days, the client will not attempt to refresh its record for 2 days if left up and running.
3. Therefore, by the end of the refresh interval at 12:00, the client has not refreshed its record and consequently at 12:25 the record is scavenged, despite the client being up and running and in use all that time.

Therefore, take care with over-zealous scavenging. Know your DHCP lease times!

What are the best settings?

I've heard talk that with a default DHCP lease time of 7 days, no refresh of 3 days and 4 days refresh intervals are good (the idea being that T1 is 3.5 days in, when the refresh interval is in place - again be careful its not the other way round with a no refresh of 4 days as T1 will always fail, though T2 at 87.5% of the way through the lease should succeed).

However, I'm not convinced. If the person goes on holiday for a week, the record will be scavenged - is that good? It might be depending on your point of view. However, by leaving it there it's not necessarily a bad thing. If the client is turned off for a few days and the lease expires, then as soon as it comes back on line and picks up a new IP address, the record is immediately updated anyway so it will always show the latest IP address.

Using DNSCMD to see record timestamps:

I don't like the DNS console much. It's slow to update so I can never trust the information it is giving me is up to date. DNSCMD often will show me that a timestamp is updated well before the console ever does.

If you want to see the timestamps for records in the console, you need to turn on advanced features. If you want to see them using DNSCMD, it's:

dnscmd servername /EnumRecords DNSZoneName @ /Type A

This will show you all A RRs, with timestamps.

DNS Zone info in AD

Server 2003 introduced DomainDnsZones and ForestDnsZones application partitions. These moved DNS data out of the domain partition where it was replicated to the GC to its own partition.

In ADSIEdit you can see where each of the partitions stores DNS zone information (2 domain forest example):
Domain partition - cn=MicrosoftDNS,cn=System,dc=childdomain,dc=example,dc=com
DomainDnsZones partition (child domain) - cn=MicrosoftDNS,dc=DomainDnsZones,dc=childdomain,dc=example,dc=com
DomainDnsZones partition (root domain) - cn=MicrosoftDNS,dc=DomainDnsZones,dc=example,dc=com
ForestDnsZones partition - cn=MicrosoftDNS,dc=ForestDnsZones,dc=example,dc=com

If you try to change the replication scope from one to the other, and receive a message like:
"the name limit for the local computer network adapter card was exceeded"
then this likely to be because the zone already exists in the other location.

e.g. you try and move the replication scope of a DNS zone to DomainDnsZones and receive the message above.
Open up adsiedit and navigate to the relevant location in DomainDnsZones. You will likely already see a copy of the partition you are trying to move there.
Delete it, along with any ...in progress... zones you see there as these are the failed attempts to create the zone.
You will immediately be able to change the replication scope.

How did this get created?
Well, in my case it was creating a Server 2003 forest. This automatically installs DNS to the new partitions. However, I wanted to move them back, as well as move _msdcs.forestname to a sub domain of the forest root to represent how a Windows 2000 forest would look. Therefore, when I came to "upgrade" it to a Server 2003 forest-style DNS configuration again, the partition had already existed from when it was first installed.

_msdcs.forestname
Under Server 2003, this is in its own zone, replicating forest wide.
In Windows 2000, it's a sub-domain of the forest root domain. Which was always a pain as you had to ensure the child DCs also had a copy of the forest root DNS data, often using a secondary zone.
It can be upgraded using the instructions in kb817470.
To downgrade it, as I did here, I had to dfo the following:
The newly created _msdcs zone was deleted followed by the deletion of the automatically created delegation (under the forest name DNS zone), then create a new _msdcs subdomain under the forest name DNS zone and restart the netlogon service on every Domain Controller in the forest in order to force related DC RR registration.
Finally, you may speed the ForestDNSzone NC replication by running repadmin /syncall or /replicate /force commands.

Legacy group membership

Once you update the forest to 2003FFL, linked-value replication (LVR) comes into play.
However, this doesn't affect existing group membership entries.

This can be seen through repadmin:
repadmin /showobjmeta DC_Name Dn_of_group

Distinguished Name =============================
ABSENT member 2010-02-19 18:02:53 US\Server1
205479704 205479704 2
CN=user1,OU=Accounts,DC=test,DC=com
PRESENT member 2010-02-22 10:21:05 UK\Server2
208940639 143787850 1
CN=user2,OU=Accounts,DC=test,DC=com

Key
ABSENT - LVR enabled member. This is similar to a tombstoned object - it records a membership that has been removed. This will be garbage collected after the tombstone lifetime has expired.
PRESENT - LVR enabled member.
LEGACY - legacy membership. LVR will not take affect for this member.

These can apply to members and also to "managed by" properties.

LVR enabled basically means that there is additional metadata associated with the membership, stored in msDS-ReplValueMetaData attribute for the group.
If you compare an entry for this attribute before and after a membership is LVR-enabled, you see something like this:

LEGACY member:


member
CN=User1,OU=Accounts,DC=test,DC=com
0

1601-01-01T00:00:00Z
2003-04-11T20:55:09Z
0
1601-01-01T00:00:00Z
00000000-0000-0000-0000-000000000000
0
0

"

Updated PRESENT member:

member
CN=User1,OU=Accounts,DC=test,DC=com
0

1601-01-01T00:00:00Z
2003-04-11T20:55:09Z
2
2010-02-17T18:56:26Z
ebc213fa-1e61-423d-92ca-d938777a2348
7694183
203002792
CN=NTDS Settings,CN=Server1,CN=Servers,CN=UK,CN=Sites,CN=Configuration,DC=test,DC=com

Updating Legacy Memberships
In order to update all memberships and managed-by's to be LVR-enabled, you need to remove that membership and re-add it. This updates the metadata.

I'll try and paste some Powershell to do this later on as a response to this post.

Old password still valid for an hour

Not sure why I haven't blogged this one before.

With Windows Server 2003 Sp1, MS decided to modify NTLM authentication behaviour (this also affects LDAP authentication) so that a user can use their old password for an hour to access the network after it has changed.
This is to allow things like service accounts to still be able to login while the new password propagates.
However, I've never been convinced that the case is there (why does kerberos not need it?).

To disable this feature, you need to modify HKLM\System\CurrentControlSet\Control\LSA\OldPasswordAllowedPeriod. This is a DWORD value.
Setting it to 0 means that you disable the use of old passwords.
If this value is not in the registry the default of 60 minutes is used.

See kb906305 for further details.

Using Winbind in Linux to authenticate to AD

Linux has Pluggable Authentication Modules (PAM) to provide a common set of authentication of APIs.
Also has Name Server Switch (NSS) APIs for looking up user information.
This resides on each linux machine.

AD authentication method 1
Have PAM and NSS communicate directly with AD using LDAP. However, LDAP authentication passes in clear text (so would need to use LDAPS). Also need to specify DC.

AD authentication method 2
Get PAM to use Kerberos and NSS to do group membership lookup via LDAP. More secure but still has problem where have to specify DC. Also doesn't manage passwords well.

AD authentication method 3
Use Samba.
Samba has a Winbind daemon that interprets PAM and NSS calls and interprets them into AD calls, using either Kerberos, LDAP or RPC, depending upon which is appropriate.
Winbind can also replicate the DC locator mechanism to find the best DC using SRV records.

Winbind parameters
Winbind domain: Netbios name of AD domain.
Security model: ADS.
Winbind ADS realm: DNS name of domain.
Winbind domain controllers: name of DC or * to get Winbind to use its DC locator mechanism.

ID mapping
NIS is used in Linux often to ensure that a Linux UID is unique on machines.

Can therefore have a mechanism to do a 1:1 mapping of AD users and groups to Linux UIDs. This is a big admin overhead.

Easier method is to get Winbind to extract the RID from the SID when the user logs in and use the RID as the UID.
This has zero admin overhead but won't work in a multi domain environment in case the RID is repeated.

Implementing mapping
To implement mapping, edit etc/samba/smb.conf
AD mapping: add "idmap backend = ad"
RID mapping: add "idmap backend = rid"

Synchronisation of the Directory

There are 2 primary methods for synchronising with the AD.

The first is DirSync. The usual way of doing this is to have it operate in "per partition" mode.
This requires that the permission Replicating Directory Changes is granted to the user account under which the DirSync control is running.
This permission has to be granted on the top level of the partition.
It then uses a cookie to first sync with the partition and then to subsequently poll the directory, after which the cookie will contain only the directory changes.
Since this permission is at the top level it will contain all changes in the directory and an ldap query will then filter it for just the objects and attributes required.

The Replicating Directory Changes permission basically appears to grant the right to read and copy the data from AD.

This does give some concern over how confidential data is read (using the confidential bit for example). The data will still be detected as changed. However I'm unsure if the data can then be read by the program, so you may end up with a change recorded but no record of what the change actually was. This also really applies to all objects - does the DirSync account also need read permissions on any attributes that it wants to sync the changes for?
By default it will also record deletions as tombstoned objects, but again, may have to grant permissions on the Deleted Objects container.

There is also a "per object" mode which allows you to not record all changes and does not require the Replicating Directory Changes permission. However, details here are sketchy.


The second method is to monitor the uSNChanged attribute.
Since every time a change is made on a DC, it updates its highestCommittedUSN (check this attribute out on RootDSE).The uSNChanged attribute of the object that was changed is then set to the same as the highestCommittedUSN value.
uSNChanged is not replicated.

Therefore, by doing a complete AD sync and then for subsequent reads only reading the uSNChanged attribute where it is greater than the previous high, you get a list of objects that have changed.
Deletions are captured either by reading the Deleted Objects container or by periodically doing a full AD sync.

A single DC will need to be read otherwise the uSNChanged attribute becomes meaningless. If the DC does change a full sync will be required.

Objects should be tracked using the objectGUID since this is the only attribute guaranteed not to change.

Some useful links:
http://msdn.microsoft.com/en-us/library/aa772153(VS.85).aspx
http://support.microsoft.com/kb/891995
http://msdn.microsoft.com/en-us/library/ms677626(VS.85).aspx
http://www.experts-exchange.com/Software/Server_Software/File_Servers/Active_Directory/Q_24206134.html

Change Notification For Intersite Replication

You can override the default replication interval for intersite links.
To do it, go into adsiedit with an Enterprise Admin account.
Navigate to the siteLink object you want to set change notification for, and edit the options attribute.
The first bit needs to be 1. Generally, the attribute will not be set so just set it to 1. Otherwise you'll need to do a bitwise OR calculation on it to add the 1.
Once this is done, allow the configuration partition change to replicate to all affected DCs in those sites. After it has done, all further changes made to the AD on DCs in those sites will replicate between the 2 sites faster using change notification.