Math Faculty Computing Facility
University of Waterloo
N2L 3G1 Canada
July 10, 1991
Xhier is a system for software maintenance and distribution, currently in use at the University of Waterloo. It allows easy, automatic software installation on a variety of UNIX systems. This paper describes some of the design goals of xhier, its structure and operation, as well as some of the problems we have encountered, and some future goals.
The Math Faculty Computing Facility (MFCF) is a major UNIX support organization at the University of Waterloo, and, in addition to other responsibilities specific to the Math Faculty, provides software support for UNIX machines all over campus using the xhier software maintenance and distribution system.
MFCF started doing UNIX support before networks and workstations were so widespread. It was not too many years ago when the MFCF UNIX environment was a single DEC VAX 11/780 running 4.2BSD UNIX. In that environment, adding and updating software was relatively easy, since operating system releases were infrequent and the amount of third party software was small. It was easy to just add some source to /usr/src and install some commands in /usr/bin. If there were bugs in the OS that needed fixing, or enhancements that were desired, it was a (relatively) simple matter to modify the BSD source, and install the change. A few more BSD VAXes were added without too much pain, and careful use of tools like rdist(1) and rcp(1) made it possible to keep things up to date.
Once UNIX machines started being more common on campus, and different models and brands of machines started appearing, it quickly become obvious that a different approach was required. With operating systems being updated as often as every few months, different operating systems using different file system organizations, and with useful third party software proliferating, a more structured and automated approach was needed.
Development of the xhier system was started in early 1989 and has been ongoing ever since, with many enhancements and fixes still planned. MFCF currently provides software support for 11 different machine ``architectures'', using the automated tools provided by xhier. Xhier is currently used by over 250 machines, with over 1200 commands in more than 200 software ``packages'' available, with over 14,000 package installations on the various machines.
MFCF was faced with the task of providing software support for multiple architectures, serving multiple machine administrations, and different user groups and purposes. A research machine in the Computer Science department might have different needs than a student workstation cluster in Electrical Engineering, though both would want some form of software support, and some selection of locally installed commands. Additionally, most users want a consistent environment, regardless of the architecture and operating system of the machine they currently happen to be using.
Accordingly, it was necessary to find some way to distribute software to, and make it available on, large numbers of different machines. This had to be done with a large amount of configuration flexibility (so it could be tailored to the needs of different machines and administrations), and it had to be as automatic as possible, in order to make the most efficient use of the software support staff.
Faced with this environment, MFCF's work on xhier has proceeded with these primary goals:
Xhier takes a different and more automated approach than systems like the depot. Appendix B discusses some of the differences between xhier and the depot.
In order to minimize the changes to stock file organizations, it was decided to locate xhier'd software in a separate (hopefully) unique location - under /software.
Under /software are the package directories. In xhier, software is organized into ``packages'' of related software. For example, there is a tex package, containing TEX related software, and a gnu package containing GNU software. Separating software into separate packages has two main advantages: it allows a system administrator to pick and choose what software to make available on a machine (which may depend on licensing or size considerations), and it allows the grouping of related software for maintenance and similar purposes. Additionally, it allows multiple versions of the same software to be installed and available at the same time.
The other goals are approached using the details of the organization of software packages, and by the use of support programs.
Xhier recoginizes that there are three main steps in making software available to users:
Each software package contains all the files associated with that software, along with other files required for the distribution and installation of the package. A package typically consists of a number of sub-directories, with the common ones being:
% lc /software/x11 Directories: .admin bin data doc export include lib logs man
Two sub-directories are used by xhier for distribution and installation.
% lc /software/x11/export Files: boottime crontab inetd.conf services
% lc /software/x11/.admin Files: Dependencies Install Maintainer Options Targets file-types
The package structure makes each package a separate entity, containing all the knowledge required to install and use the software it contains. It also provides a consistent structure so that packages tend to look very similar, which makes maintenance easier.
See appendix A for an example of how packages are created and installed.
Now for the complication. Recent versions of UNIX have reorganized parts of the file system to make file sharing easier. For example, SunOS has /usr/share, which contains those files that can be shared among all machines, regardless of CPU type. Xhier takes this idea a step (or several steps) further.
Separating files based on their type has two main advantages, both of which help to satisfy the goals of xhier. It makes it easier to share software via remote file systems (e.g. NFS), because you can share as much of a package as possible, and minimize duplication across machines of the same or different hardware architectures. And it makes software distribution easier because it makes it easier to use a tool like rdist(1) to distribute the appropriate file types to the appropriate machines e.g. shared files like man pages can be distributed everywhere, while executable binaries can be distributed to just those machines with the same architecture.
Xhier currently recognizes six file types:
This is accomplished through the use of the /.software directory. /.software contains one sub-directory (or, more typically, a symbolic link to, or an NFS mount of, an appropriately located directory) for each of the six basic file types, with each sub-directory containing package directories. For example, the man pages for the tex package are in /.software/share/tex/man, while the command binaries are in /.software/arch/tex/bin. Note that /software and /.software/share are the same place.
Every package has a shared component; in particular, the .admin directory must exist, and is always a shared directory. And since /software and /.software/share are the same place, this ensures that, for a package called pkg, /software/pkg will always exist.
It is desirable to hide the file type structure of a package from the outside world, so that a user or another program does not need to be concerned with the internal structure of a package. So, even though every file in a package can be referenced through /.software type directories, references from outside the package itself are always made through /software/pkg. (A package is allowed to know its own structure, and refer through /.software as appropriate.) This is accomplished by the (liberal) use of symbolic links in the /software/package directory. For example, /software/tex/bin is actually a symbolic link to /.software/arch/tex/bin. It is also possible to have a shared directory contain real files along with symbolic links to files in directories of other types. For example, a package might have commands that are primarily shareable shell scripts, with the few compiled commands being referenced through symbolic links in the shared bin directory. Note that these links, while pointing to different absolute locations on different machines, are shareable, since they link through /.software.
The result is that the file type directories under /.software provide separate hierarchies for different types, which makes file sharing and distribution much easier. For example, diskless workstations will usually share and mount all file types from their server, with the exception of the local type, which each client will have its own version of. The /software link provides a naming convention for access to the components of a package. Note that a given machine will have the correct type hierarchies for itself only, e.g. /.software/arch on a VAX contains only VAX binaries, and on a Sun will contain only Sun binaries.
Xhier is built on the idea that automation is good. Accordingly, there is an xhier program to do just about anything to do with software installation and distribution.
There are currently about 80 commands in the xhier package, most of them shell scripts. Commands are used for such things as package creation, maintenance and installation, and distribution, along with some utilities that are useful for day-to-day operation (log file rollers, etc.). The following is a summary of some of the more prominent commands.
The xh-mkpkg command creates a package skeleton in the
/software directory, with command and support directories, and
some template files, that can be filled in by the package creator. This
makes the initial creation of a package simpler, because a
package creator doesn't have to remember all the details.
xh-make is used as a cover for make(1) and
imake(1) and uses a standard imake template to make
it easy to compile and install a program
#include "../PackageName" Program(callpat,c,user)
Once a package has been compiled and installed into the package directory, the xh-install command makes it available for use. It does this by running the package's Install script, to do any package-specific tasks (such as package configuration, or replacement of any stock files), and then, if that is successful, invoking other programs to patch system files (like /etc/services) from the fragments in the package's export directory.
Once a package has been installed, and is working on one machine, it can be distributed to, and built on other machines. This is a two-step process. The xh-distribute command is used to distribute the package contents and structure to other machines. It creates a Distfile for use with rdist that sends the appropriate files to the appropriate machines lower on the tree. For example, it will send the shared files everywhere, but architecture specific files will only be sent to machines of the same architecture (though it will complain if files on a machine of a different architecture are out of date, or missing). xh-distribute then invokes rdist and summarizes the output for easier human consumption.
The second step is xh-sdist, which distributes the source for the package to machines of other hardware/software architectures (the ``architecture master'' machines), and runs xh-make on those machines in the background, collecting the output and mailing it to the invoker. xh-sdist has a number of options and configuration files to control exactly what gets distributed to the remote machines (e.g. .o files are not distributed). xh-sdist has turned out to be an incredibly useful tool - it makes updating software and installing fixes almost trivial.
To help ensure that software is kept up to date, working, and consistent on all machines, xh-maintenance is run weekly from the cron (on every xhier'd machine). xh-maintenance does some housekeeping and consistency inspection, and runs xh-install and xh-distribute on all packages, arranging to send the appropriate output to the appropriate package and system maintainers.
Since xh-maintenance is run weekly, starting at the same time on every machine, it can take several weeks for package updates to reach the bottom of the distribution tree. And so the other important distribution-related tool is xh-dist-recurse, which attempts to distribute a package recursively down the distribution tree. This can be very useful in distributing fixes quickly, in case you've inadvertently broken something and don't want to wait for the usual propagation delays to distribute the fix.
In a conventional software installation, software is installed in one location only (commonly /usr/local/bin), or in other OS-specific directories (such as Sun's /usr/openwin/bin), which the user is expected to add to PATH. Expecting a user to know where software is installed is not always reasonable. For many years, MFCF has had a command, showpath, to help users set their PATH without having to know the specifics of a particular machine.
With xhier, once the packages are installed, there are many different package directories, each containing command directories (e.g. bin), man directories, include directories, and/or lib directories. The showpath command could simply produce a list of all the bin directories in all the installed packages. This, however, quickly becomes too cumbersome to handle - a 3,000 character PATH containing 200 directories is a little large. It also becomes inefficient since some shells hash the PATH contents, but often only the first few elements (csh apparently hashes the first 8 elements of PATH).
Xhier approaches this problem by creating a directory that contains a link to each of the installed, packaged commands, using the xh-make-links command. These links have to be symbolic links, since the commands are not always going to be on the same filesystem (or even the same machine). There are actually three of these directories, one for user commands, one for maintenance commands, and one for commands used by other programs (servers). These directories are system-wide and system-specific in that they contain links to the commands for only those packages that the system administrator wishes to be ``default'' packages. A user's PATH consists of the directory of links, the appropriate stock directories (for that architecture), and any other directories the user wishes to include. These directories of symbolic links might be expected to cause a performance penalty on command invocation, but, in practice, it seems not to be a problem.
This ``searchrules'' approach to commands works well, but it is harder to apply it to the other parts of a package. Many man(1) commands understand a MANPATH variable, but most compilers and loaders don't have an easy way (i.e. an environment variable) to extend the searchrules for include files and libraries. Include files and libraries are linked (again using xh-make-links) into the standard locations honoured by the OS, usually /usr/include and /usr/local/lib.
One more minor complication. MFCF had been using the rman remote man command to provide man pages to smaller machines, where there is often not enough disk space to keep man pages locally. Local modifications to rman allow the remote man page server to deal with packaged man pages by special interpretation of the MANPATH variable.
setenv PATH `showpath usedby=user $HOME/bin standard` setenv MANPATH `showpath class=man standard` setenv EDITOR `showpath findfirst=vi` % echo $PATH /u/jmsellens/bin:/.software/local/.admin/bins/bin:/usr/ucb:/bin:/usr/bin % echo $MANPATH /software/.admin/man/:/usr/man/ % echo $EDITOR /.software/local/.admin/bins/bin/vi
Xhier tends to point out some of the more obscure points of software installation. One prime lesson is that doing software distribution can turn out to be a complicated process, with a certain amount of overhead.
Automation is a necessity, once more than a few machines are involved, but it turns out that some things can be hard to automate. For example, it's not too hard to add an entry to the /etc/services file, but it's a little harder to notice conflicts or replacements, and some packages might want a particular userid to exist (e.g. ``games''), but it's hard to add an appropriate account to a passwd file automatically. The current xhier utilities handle inetd, cron, and /etc/rc entries fairly well (by automatically editing the appropriate system files, and disabling conflicting stock entries), but passwd and group file entries still have to be applied by hand (local administrators are prompted to do it when a package is installed). And while automation allows easy installation and distribution, it also makes it easy to distribute broken software and automatically destroy dozens of machines.
Some software comes with particular pathnames built-in, often hard-wired pathnames are sprinkled throughout the code (especially files destined for /etc), and it's often hard to change these pathnames to refer to locations under /software instead. In other cases, they are easy to change, but the same symbolic name is used for different kinds of files. For example, the make variable ``USRLIB'' might be set to /usr/lib, and be the location for both object libraries and data files, or ``ETCDIR'' might be used to contain local configuration files and system maintenance commands. Files are often named for their location, rather than for their function. Xhier-ing a package forces an installer to interpret the given pathnames, and choose appropriate xhier pathnames instead.
In addition, certain pathnames, like /usr/lib/sendmail and /usr/ucb/lpr, are known to many programs, and so any replacement versions of these programs must also replace the well-known stock pathnames (with links to xhier file locations).
It's hard to organize NFS (or, presumably, other remote file system) mounts to make software sharing foolproof. This problem is especially evident with symbolic links. A symlink to an absolute pathname will start looking for the file at / on the local machine, even though the link is on the remote machine. And conversely, a relative symlink will tend to stay on the remote machine, unless it passes up through a mount point. Some uses seem to call for absolute symlinks, and some seem to call for relative symlinks, but so far we have been unable to come up with a rule (or even a guideline) that works in all cases.
Choosing what goes in what package is hard. It's not always easy to group software into appropriate packages, and different administrators might want different parts of a package, instead of the whole thing. Large packages are easier to deal with, but small packages give greater flexibility.
It's easy to generate lots of junk mail, so efficient ways of summarizing and distributing it are required.
The current distribution mechanism uses a tree structure, with each member in the tree ``pushing'' packages to lower nodes in the tree. This is less than optimal, but was necessary since the only distribution tool available at the time of implementation was rdist. Some of the problems with this approach are
This, however, requires a program like track(1) that can do most of the same things that (the locally enhanced) rdist(1) can do. This work is currently underway. It also requires fairly major changes to the mechanisms that determine which parts of which packages get sent to which machines. These changes are almost in place, to be used with the current distribution mechanism until something ``track-like'' is ready.
Most machines don't have enough local disk space to contain all packages, or even just the ones they want to use. Current space usage on the primary maintenance machine is
% du -s /.software/* 105 /.software/admin 272164 /.software/arch 11823 /.software/local 727 /.software/regional 231340 /.software/shareor a total of over 500 megabytes, in addition to the space required by the OS. As a result, remote file system mounts (currently only via NFS) are required to make the necessary software available on many machines. This is usually done in bulk, where a machine might mount all of /.software/share and /.software/arch from a willing server. Things get complicated when you wish to have some packages locally and some remote mounted, since you either have to mount each package individually, or arrange that /.software/*/package are either real directories or symbolic links through the appropriate mount points.
At present, all remote mounting is done manually, which is getting less and less practical. Some method needs to be devised to automatically make the appropriate mounts and links, which will allow a local administrator to easily add a software package to a machine. It's not at all clear what the best approach to this problem is.
Xhiering a software package forces a software maintainer to look beyond the local machine, and consider how a package would be installed and used on different machines, with different administrators, perhaps wishing different defaults or environments. However, the current outlook tends to stop at the edges of the campus. A ``shared'' file tends to mean one that applies to the campus as a whole, rather than to the ``world''. In fact, there is currently no way in xhier to indicate a file with type ``campus'' i.e. there is no /.software/campus. It would be nice if it was possible to distribute xhier and xhier'd packages to other organizations, but that brings other problems to light, such as the necessity to trust a remote software provider.
With Xhier, it is now almost trivial to make a new software package available to hundreds of machines. It's easy to make and distribute fixes or enhancements.
Xhier allows MFCF to provide software support, and a consistent user environment, on many more machines than would be otherwise possible.
Xhier does have a number of faults
First, the source is unpacked into /usr/source/foo, and the Makefiles are modified as appropriate, to set the right options, and installation locations. Many times this is done using an xh-imakefile (see figure 4 for an example) that is processed with xh-imake to allow easier per-architecture customizing.
% xh-mkpkg foois used to create a skeleton structure in /software/foo. This structure is then adjusted by hand as required. Usually this just means removing the parts of the skeleton not required by this package.
Then the files in the .admin directory are edited as appropriate. The appropriate options are set from the defaults in the Options file, and the Install script is modified to do the appropriate things on installation and uninstallation. Often this means just ending up with the script
#!/bin/csh -f exit 0The foo package name is then ``registered'' by adding it to the software(i) man page.
The software is installed via
% xh-make installin the source directory. Once that completes, the commands
% xh-install foo % xh-distribute foo % xh-sdist -m installinstall the package locally, and then distribute the package structure and source to the other architecture master machines, and compile and install the package there.
Once this has been done, the foo package is available to any xhier'd machine on campus, with very little effort.
Both systems separate software into packages, and define a structure for the individual packages. The depot requires human intervention for each installation, in order to execute an optional installation script, and to modify system files (such as inetd.conf) if necessary; xhier uses the xh-install program and the package's Install script. The depot uses redundant servers for the packages, but distributing a package to a redundant server is done manually; xhier has automatic distribution of the packages. The depot is geared primarily to the use of server machines that contain all the files for all architectures. This makes it less useful for standalone machines (such as a home workstation), since such a machine would have to have the binaries for every architecture, requiring more disk space (unless manual pruning was used).
Maintenance activities are less automated in the depot as well. In order to recompile or update a package, a maintainer has to go to the depot administration machine of the appropriate architecture, and compile and install the package there. Xhier, on the other hand, uses the xh-sdist command to rebuild and install a package on all subsidiary architectures, which helps ensure that packages are kept up to date on all architectures.
Since the depot uses automount mount points, as listed in an NIS database, to make packages available to client machines, it is harder for a system administrator to tailor the set of packages available on a particular machine or set of machines. Xhier makes it easy for an administrator to pick and choose what packages are available on his or her machine(s). Admittedly, in a primarily client-server, workstation environment, this is less of a problem.
With xhier, a system administrator can control when packages get updated on his or her machine, which is useful in a student lab environment, or on a personal workstation when trying to complete a project. Since the depot is based on the client-server model, it is much more difficult for a client to ``refuse'' an update.
Xhier provides a set of tools for modifying, or replacing vendor provided files, which makes it possible to correct flaws in stock systems, or customize them to local requirements. Xhier provides simple interfaces to system services such as /etc/rc and cron. This means that a package maintainer does not usually have to worry about system-specific details.
Xhier makes extensive use of search paths in order to hide the details of an implementation from the users. With the depot, a user must know where programs and man pages are located in the hierarchy in order to make use of them.
Xhier is still under development, and still contains many transitional provisions to correct past mistakes, and it is currently difficult to bootstrap xhier into a new environment (such as another campus). It is also a large collection of source, which relies in part on modifications to licensed software.
As such, xhier is not currently available for distribution, though we hope to be able to make all or part of it and its tools available at some point in the future.