Install at Boot on Citrix XenApp PVS Servers

The Problem

Whilst designing a XenApp 7.x infrastructure for a customer, a requirement surfaced that they wanted to be able to update their in-house developed applications, of which there were three separate ones, written by three separate teams, on a weekly basis. Given that they did not have App-V packaging skills, so weekly repackaging would have been cost prohibitive, we looked at using Citrix AppDisks so that we could hand Server 2012R2 virtual machines over to the customer with AppDisks in write mode, let them install and test their apps and then hand them back to us for sealing and promotion to UAT and thence Production. Unfortunately, we hit a number of issues with the AppDisk technology which meant that we had to seek an alternative way of delivering these applications.

The Solution

Using a computer startup PowerShell script assigned via Group Policy to the OU where the servers earmarked for the customer’s applications resided, I came up with a robust mechanism which would install the applications during boot. The Server 2012R2 VMs with the Citrix VDA installed, which were providing published applications and desktops to the end users, were booting from Provisioning Services with Cache in device RAM and overflow to local disk so we therefore needed to ensure that the installation of the software didn’t use the RAM cache as this could have had a negative effect on performance. It was deemed too high risk to have the software explicitly installed to a drive other than the system drive, namely C:, even if the installers allowed it, particularly as some of the required prerequisite packages were quite old and out of support. We therefore added another virtual hard disk to the VMs which the installation script would find, format and then create symbolic links to folders on the additional drive from the installation folders on the C: drive so that their software could be installed without using C: drive space.

As the customer’s previous XenApp 6.x implementation had sprawled to over twenty different PVS vDisks, including the ones the developers used which had been continuously in private, writeable, mode for many years so were very different from the rigidly change controlled production images, the solution needed to keep the number of vDisks to an absolute minimum so the install on boot apps would install on the single base image that was also used for COTS (Commercial Off The Shelf) apps.

Solution Detail

Support for Multiple Applications

Since there were three distinct applications which required different installers to be run, with possibly more applications to come, I tried to keep the complexity low, particularly in terms of scripts since trying to maintain multiple scripts was something that could prove problematic for BAU operations. I therefore used AD groups to designate which XenApp servers got which applications where the application name was part of the group name so it was simple, via good old regular expressions, to figure out which application a server should be getting by enumerating its AD group memberships via ADSI (so that the Active Directory PowerShell module isn’t required).

$filter = "(&(objectCategory=computer)(objectClass=computer)(cn=$env:COMPUTERNAME))"
$properties = ([adsisearcher]$filter).FindOne().Properties
## We will get computer group memberships to decide what to install if( $properties )
{
[string[]]$groups = @( $properties.memberof )
if( ! $groups -or ! $groups.Count )
{
## Already determined via computer name that we are a custom app server so having no groups means we do not know what to install
$fatalError = "Failed to get any AD groups via $filter therefore unable to install any custom applications"
throw $fatalError
}
else
{
Write-Verbose "Found $($groups.Count) groups via $filter"
ForEach( $group in $groups )

This was preferred to using different OUs for the different apps so then only one copy of the startup script was required and the chances of a server being in the wrong OU were reduced, especially in Production where a single OU was used for servers providing these custom apps and those not providing them. The script differentiated between the two types of server via NetBIOS name as a different naming convention was used for those providing custom apps so the startup script could exit almost immediately if the NetBIOS name was not that of a custom apps server and thus not delay the boot noticeably.

In order to keep it simple from a Citrix Studio perspective, we used Citrix tags to designate each application so that we could use a single Delivery Group for all of the custom applications and use tag restrictions on Application Groups to direct them to the correct XenApp server. This was in preference to using multiple Delivery Groups. There did not unfortunately appear to be an easy and reliable method of getting tag information on the XenApp server during boot otherwise we would have considered just using tags as opposed to tags and AD groups.

Disk Preparation

The startup script enumerates local disks by way of the Get-Disk cmdlet, filtering on a friendly name of “VMware*”, since that is the hypervisor in use, and where the disk size is 4GB or less so that it does not try to use the non-persistent system drive or the persistent drive used for the PVS overflow disk. These filters are passed to the script as parameters with defaults to make the script more easily repurposed.

$disks = Get-Disk -FriendlyName $diskDriveName | Where-Object { $_.Size -le $agileDriveSize }
if( ! $disks -or $disks.GetType().IsArray )
{
$fatalError = "Unable to find single drive of type $diskDriveName of size $($agileDriveSize / 1GB)GB or less"
throw $fatalError
}

If the disk’s partition style is “raw” then it is initialised otherwise all of its partitions are deleted in order to remove all vestiges of any previous application installation.

if( $disks.PartitionStyle -eq 'raw' )
{
$disks = $disks | Initialize-Disk -PartitionStyle MBR -PassThru
}
else ## partitions exist so delete them
{
$disks|Get-Partition|Remove-Partition -Confirm:$false
}

The drive is then formatted and a drive letter assigned although the drive letter is not hard-coded which removes the chances of clashing with others disks and optical drives.

$formatted = $disks | New-Partition -AssignDriveLetter -UseMaximumSize | Format-Volume -FileSystem NTFS -NewFileSystemLabel $label -Confirm:$false

if( ! $? -or ! $formatted -or [string]::IsNullOrEmpty( $formatted.DriveLetter ) )
{
$fatalError = "Failed to format drive with label "$label""
throw $fatalError
}

Finally, permissions are changed on the root of the new drive since non-admin users get write access by default which is not desired (or secure!).

As the underlying storage tier had built-in deduplication, we hoped/believed that the overhead of say ten instances of this additional disk for one application was nowhere near ten times the size of the data written to one disk.

If using the XenDesktop Setup Wizard within the PVS console to create one or more VMs with additional disks, ensure that the “UseTemplateCache” registry value is set to 0 where the PVS MMC snap-in is run from and that it is set before the mmc process is started otherwise additional disks in your template VM will not appear in the newly created VMs. See this aticle for more information.

usetemplatecache

Symbolic Links

Before the installation script for the custom applications can be run, symbolic links have to be created from the system (C:) drive to the newly created drive in order to preserve the PVS RAM cache. This was achieved by having a simple text file for each application which had one folder per line where the folder was the name of a location on C: where that app would install software components. For instance, a line in the text file might be “\ProgramData\Chocolatey” which would result in the following PowerShell code being executed where $dest =
“D:\ProgramData\Chocolatey” and $source =
“C:\ProgramData\Chocolatey” :

if( ! ( Test-Path -Path $dest -PathType Container -EA SilentlyContinue  ) )
{
$newDir = New-Item -Path $dest -ItemType Container -Force
if( ! $newDir )
{
$fatalError = "Failed to create folder $dest`n$_"  
throw $fatalError
  }
}
## This requires PowerShell 5.0+
$link = New-Item -ItemType SymbolicLink -Path $source -Value $dest
if( ! $link )
{
  $fatalError = "Failed to create link $source to $dest`n$_"
throw $fatalError
}

Where D: is the drive letter that was assigned to the 4GB additional drive.

One slight hiccup we encountered was that although the startup script was running with administrative privilege, it still didn’t have sufficient privilege to create the symbolic link, which also requires PowerShell version 5.x. I therefore added functionality to the script to take a local copy of the running script and elevate it to run under the system account using the SysInternals psexec utility.

Local Administrator Access

In order for this solution to not be onerous on BAU support and potentially introduce delays for the customer’s developers, specific AD groups were added to the local administrators group during boot if the server was a development one, as opposed to UAT or Production. The fact that a server was a development one, which the customer had full control over in terms of rebooting, since it was booting from a read-only PVS vDisk, was simply gleaned from the fact that the string “_DEV_” appeared at an expected position in the AD group name that was also used to determine which application was to be installed on it. On UAT and Production servers, changes to the local administrators group membership were not done.

There is a script here which uses the same code to change local group membership.

Application Installation

From our perspective, the installation of the customer’s applications was straightforward as they were responsible for the application installation scripts so the startup script simply had to call the customer’s script, checking for any exceptions so these could be percolated up into our error handling mechanism, once the additional drive was ready. Having determined which of the three applications were to be installed on a particular XenApp server by way of its AD group membership as previously described, the startup script would simply invoke the “XenApp.Boot.Time.Installer.ps1” script in the folder for that application or throw a fatal error if the script did not exist, could not be accessed or raised an exception itself.

In the development environment, where the customer had full control over the share containing their install script and Chocolatey packages and also had local admin access to the XenApp servers hosting their applications, the customer could test their own script whenever they wanted by simply rebooting their development servers.

In the UAT and Production environments, the customer had no local administrator access to the servers or write access to the share containing their install script and packages – a change request was required in order for us to copy, albeit via another PowerShell script with simple prompts to reduce the risk of BAU operations making incorrect changes, the files from the development shares into the production ones. The Delivery Groups containing these servers had a weekly reboot set to commence at 0200 on a Sunday.

One of the few limitations of this approach is that any software being installed that requires a server reboot cannot be supported since the reboot would cause the changes to be lost, because they are PVS booted in shared mode, and a cycle of infinite reboots would ensue. However, the nature of the applications, particularly as they would not be in use during installation as the server would be starting up, meant this was unlikely to happen. It would be more likely that a change to a required component, such as the .NET Framework, would require a reboot which is covered later.

Multiple Datacentre Considerations

The whole solution was provided out of two datacentres for resiliency so the solution for these bespoke apps also had to fit within this. To that end, two shares were provided for the UAT and Production environments such that the startup script would examine the IPv4 address of the booting XenApp server and select the installation share local to that datacentre although with a mechanism to fallback to the share in the other datacentre if the local one was not available. Data was replicated between the two shares by way of the script that BAU operations would use when promoting the customers scripts and packages from Development after a change request had been approved. This script had its own sanity checking in such as ensuring that there were actually changes, additions or removals to the files that were being promoted, as it kept a copy of the previous installation files, this being achieved via checksumming files rather than just relying on comparison of file sizes and datestamps.

At the Citrix level, there were two machine catalogues containing the bespoke app servers – one for each datacentre although this did not affect the startup script at all other than as mentioned in the previous paragraph.

Testing PVS vDisk Changes

Since the bespoke app servers are booting off the same vDisk as normal XenApp servers, in order to reduce complexity, regular updating and patching of the base disk is occurring which needs to also be incorporated into the bespoke apps to ensure they continue functioning as expected. In the development environment, the bespoke app servers are not subject to an automated reboot schedule, unlike the normal servers, since the developers are in control of their servers. It was agreed with the customer that there would be a ten working day grace period from when a new vDisk version was put into production before the development bespoke app servers would be forcibly rebooted by BAU operations if any had not been rebooted already. In order to automate this, a “message of the day” (MOTD) mechanism was implemented which launches notepad with a text file when an administrator logs on where the text file contains the date by which the reboot needs to be performed and the reasons why. The creation of this startup item and the contents of the text file are part of the script used to promote vDisks so is all automated and when a server is rebooted the startup item is removed via the startup script so that erroneous reboot notifications are not shown at admin logon. A scheduled task could have been created to perform the forced reboot after ten days, and removed at boot by the startup script, but the customer did not want this functionality automating.

As mentioned earlier, there are also potentially customer instigated PVS base disk changes where the developers need to test a change to one of their components which requires a reboot, such as the .NET Framework or database drivers. This is catered for in development by making the development servers boot off a copy of the main vDisk which most of the time is identical but can have a new version created if required. Once the new version with the potential changes has been created and promoted to a production version, just for the development servers, the customer can then test their applications and subsequently either raise a change request for the change to be incorporated into the main vDisk, which will eventually be promoted to UAT and Production, or will ask for the changes to be reverted, as in they don’t need them, so the development servers are then rebooted off the previous PVS production version and the newer disk version discarded.

Error Handling and SCOM Integration

Everything is error/double checked with exception handlers writing to a log file which is the first point of call for troubleshooting. If a fatal error, as in exception, is detected, the Citrix Broker Agent service is stopped and disabled by the script so that the affected server could not be accessed by end users, since it cannot register with a Delivery Controller, in case the application wasn’t available at all or in an inconsistent state.

A pair of events was implemented such that a successful boot and installation would log the one, information event, and a fatal error would give rise to an error event, with descriptive error test, with a different event id, being raised locally on the server. As SCOM was being used, a custom monitor and thence alert was implemented so that the BAU support team had pro-active notifications if any installation failures occurred as these would need investigating and rectifying (which would generally mean another reboot was required once the issue had been identified and resolved).

Author: guyrleech

I wrote my first program, in BASIC, in 1980, was a Unix developer after graduation from Manchester University (Computer Science) and then became a consultant, initially with Citrix WinFrame, in 1995 and later into Terminal Server/Services and thence EUC. I currently hold the Citrix CTP, Microsoft MVP, VMware vExpert and Parallels VIPP awards. I invented and wrote the first few versions of the security product which is now Ivanti Application Control (formerly AppSense Application Manager). I now work as an freelance consultant-cum-developer, live in West Yorkshire, England; have a wife, three children, one grandchild and two dogs and was a keen competitive runner until health problems put an end to that fun.

Leave a Reply

Please log in using one of these methods to post your comment:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: