John Sellens - University of Waterloo
September 22, 1995
This paper was first published and presented at the Ninth USENIX Systems Administration Conference, LISA '95, in Monterey, California, September 17-22, 1995.
The University of Waterloo is in the process of converting its administrative systems from legacy implementations on IBM's VS/1 operating system under VM/CMS and the PICK database system on UNIX, to implementations based on the Oracle RDBMS on UNIX database and application servers. As part of the ongoing conversion, we have a need for regular (and irregular) file transfers between machines, primarily from the VS/1 and PICK systems to the Oracle systems, but also between UNIX systems.
To date, we have used a number of approaches, each of which seems to have had drawbacks, primarily security and reliability related. We needed a system to allow file transfers between our systems that allowed us to limit each programmer's sphere of control i.e. we wanted to avoid giving fullscale access to all systems and servers, just to enable file transfers. We also needed a file transfer system that was reasonably portable, and worked on UNIX and non-UNIX operating systems.
The end result was the filetsf family of programs, that provides the ability to send and receive files between cooperating machines, is implemented on top of the Berkeley lpr/lpd system, and which provides a suitable level of flexibility while retaining access controls and an acceptable level of security.
In the past, we have primarily used FTP for file transfer - this meant that most programmers and operators knew the passwords to any accounts that needed to receive files. This was not a good thing, since shared passwords are no longer secret, and because it allowed general use access to accounts that shouldn't have been used for anything except file transfer. And, in the absence of digital signatures, FTP does not allow the recipient to identify the sender once a file has been delivered, making it potentially easy for legitimate data to get replaced, intentionally or not.
We implemented a simple FTP-based system using a single recipient userid (named unifer, for UNIX transfer), where all files sent to a machine would end up in a single directory, in a particular group, with group write permission on everything. This meant, of course, that any file transferred to a machine was fair game for anyone else to access. Most file transfers were reasonably well-controlled, but left-over, unclaimed, files tended to accumulate, and it's hard to make a claim that this system provided a reasonable level of security and auditability.
We also used rdist(1) in a few instances, which either required setup and monitoring by the (limited number of) super-users, or which required .rhosts file access to the receiving account, with many of the same problems as with FTP. rdist also isn't available on our VM/CMS system, which meant that it could not be used for the bulk of our file transfers. rcp(1) suffers from many of the same problems as rdist, while providing a less-reliable copy to the recipient machine (rdist uses temporary files to ensure that the destination file is always complete, rcp just clobbers any existing file, and can leave files incomplete if interrupted or the destination disk fills up).
We haven't investigated track, because we have no experience with it on campus, and because, like rdist, it requires a level of pre-arrangement that we wanted to avoid. track, like rdist, is also geared towards the replication of a set of files on multiple machines - for file transfer, we need to be able to send a file and then remove the original.
None of these approaches allows for the queueing of file transfer requests if the recipient or network is down, although it is possible to periodically repeat the rdist or track commands. And they don't provide for an ordering of the files transferred, unless you name the files with indicative names, or use file modification times as an indicator. In our previous systems, if a series of files needed to be processed (e.g. batches of transactions), the files were typically sent one at a time, processed on the recipient machine, and an acknowledgement was returned to the sender (usually via a flag file) to indicate that it was safe to send the next file.
We identified the following features as desirable in a file transfer system. Many of these are convenience features, intended to help enable a more automated environment, with less manual intervention (e.g. through avoiding entering passwords, or having to manually re-initiate transfers in case of failure).
VM/CMS has the sendfile command, which allows file transfer between users on VM/CMS systems - the file shows up in the recipient's ``reader'', tagged with the sender, filename, and file attributes. It meets many of our requirements, but we could not use it - we had no way to send the files to our UNIX machines from our VM/CMS machine, and we didn't have a VM/CMS compatible sendfile command (or equivalent) on our UNIX machines.
We were not aware of any publically available file transfer systems, and were trying to avoid investing in (and investigating) commercial solutions that might exist. We felt that our best approach would be to implement our own file transfer system.
Past LISA proceedings have contained quite a few papers on software distribution (xhier, depot, lude, lfu, CMU depot, etc.), typically based on the idea of software ``packages'' that are replicated from master machines to slave machines, typically using rdist, NFS mounts, or FTP. These systems don't meet our needs, since they are based on the idea of replication of pre-defined sets, not the transfer of individual files. The tools used to transfer files (rdist and FTP) are not suited for our needs (as described above), and we were unwilling to investigate the use of NFS mounts, since that opens up a whole different can of worms, and since NFS is not available on our VM/CMS system.
Early in 1993, we started planning the construction of a complete file transfer system, with authentication, queueing, and so on, but we didn't make much headway, for a variety of reasons. That project never received much attention, and was eventually abandoned, and was set aside until a better opportunity presented itself.
It often seems to be the case that a problem left to its own devices eventually presents its own solution, and we finally realized, after much time had passed, that it would be relatively simple to implement a file transfer system on top of lpr/lpd. The lpr/lpd system has much of what we required, and it turned out to be relatively easy to layer the additional functions on top of lpr/lpd to obtain a file transfer system that met our needs.
Why not implement a file transfer system on top of SMTP mail and MIME file encapsulation? Mail doesn't meet some of our needs: it's too easy to forge, and it doesn't provide a reliable ordering of the messages. Digital signatures (such as those provided by PGP) could be used, but they add another level of complexity, and may be harder to port to our non-UNIX systems. In addition, the path a mail message takes can't always be controlled (in the presence of MX records in the domain name server), and some mailers implement limits on the size of a message.
It turned out that lpr/lpd and the LPD protocol were a pretty good basis for a file transfer system. The protocol is well-defined, and widely implemented, and lpr/lpd already has queuing, file ordering, sender identification, machine-based access control, and so on built-in. The destination machine can be indicated by the ``printer'' name used, an lpr/lpd ``output filter'' can be used to process the files at the destination machine, and the job and class options to lpr can be used to identify the name of the file, and the intended recipient.
Using lpr/lpd was appealing because it seemed that it would save us a lot of effort - most of the hard stuff was already done - and because we could claim that we were behaving in a manner that was consistent with the ``UNIX philosophy'': small, simple, reusable tools put together to do more interesting things.
It was fairly easy to implement filetsf on top of lpr/lpd. The sendfile command is just a cover for lpr, tsfif is an lpr/lpd output filter that deposits the files into the filetsf spool directory, and we just have to set up a ``sending'' print queue for each machine we want to send to, and a a single ``receiving'' queue on each ``destination'' machine. In our case, the number of machines that we want to be able to send files to is manageably small, and our printcap file maintenance tools help a lot. An lpr client program could be used to reduce the number of ``sending'' queues required, but this would mean that there would be no sender-side queueing, which is one of the features that we were interested in. One non-trivial program is required - acceptfile is a setgid program used to query and retrieve files from the filetsf spool directory.
One final indication that lpr/lpd and the LPD protocol were a good choice was the relative ease with which we could implement filetsf on our AIX systems (on top of IBM's queueing system, which is willing to use the LPD protocol to talk to other machines) and implement the sending side on our VM/CMS machine with a REXX script on top of the netprint command. We haven't investigated implementing the recipient side on VM/CMS yet, but we're optimistic that it shouldn't be too difficult, perhaps requiring just another REXX script running in a virtual machine invoking the VM/CMS sendfile command.
Filetsf works by accepting files to send to a recipient, submitting them to lpr/lpd, delivering them into a spool directory, and waiting for the intended recipient to come and pick them up. A separate filetsf spool directory is used for easier access and control, and to provide a level of abstraction above that of particular lpr/lpd implementations. It also makes permissions easier to control, and keeps our monitoring software from complaining that the print queues are stalled.
Files are stored in the filetsf spool directory in a hierarchy intended to make it easier to list and identify the files waiting to be retrieved. The structure is three levels deep below the spool directory, encoding the recipient, intended filename, and sender (user@host) in the directory names. The individual copies of files are named for the value of time() (the number of seconds since January 1, 1970) when the file was delivered to the spool directory by tsfif, with an optional prefix. This means that the pathnames under the spool directory are of the form recipient/name/sender/[prefix]time This structure allows multiple copies of the same file to be spooled from one or more senders for one or more recipients, providing a different pathname for each spooled file. Files are named for the time at which they were delivered using the integer returned by time() to allow for easy manipulation and sorting. Any name clashes discovered by tsfif can be resolved by waiting one second and re-generating the pathname.
Most special characters are prohibited in file, sender and recipient names, to avoid potential problems manipulating the files. We had considered making the spool hierarchy only one level deep, but we felt that that would have forced even more restrictions on the component names - we would have had to use some character or sequence of characters as a separator in the directory names.
A small number (currently three) of filename prefixes are used by filetsf to indicate the status of a particular version of a file. Temporary files, indicated by the # prefix, are used when writing a file into the spool directory, and are renamed to remove the prefix when they have been written correctly. A prefix of - is used to indicate a file that has been retrieved or deleted, but not yet purged from the system. And a prefix of + is used is to indicate that the newest version of the file should replace any older versions of the file still waiting to be retrieved. This is determined by the options used by the sender when initially sending the file, and can be useful for data such as a hosts file, or a telephone list, where only the most recent copy of the file is useful.
Files are transferred using lpr/lpd one file at a time, to keep things simple. The name (or desired name) of the file is sent as the LPD job name, and the LPD class is used to pass filetsf options and the name of the recipient. The options are encoded as an integer, and are separated from the recipient name by a colon. The LPD protocol defines limits on the length of the control values that can be passed to the destination machine, but they haven't turned out to be unduly restrictive in practice.
Filetsf currently supports three options that are passed to the destination machine: notify the sender by email on delivery to the spool directory, notify the recipient on delivery, and mark the file as replacing any older versions of the file.
The core of filetsf consists of only three commands, sendfile, to send a file, tsfif, which is used as the lpr/lpd print filter and which puts files into the filetsf spool area, and acceptfile, to list or retrieve files from the spool area. In addition to the core commands, there are two utility programs, and a tsfif cover for use on AIX. The commands are written in C and the Bourne shell, and are about 2,000 lines of code in total.
The following is a short description of each filetsf command.
tsfif usually ends up being run by daemon (lpd on AIX), and the spool directory is owned by daemon (or lpd) and is in group filetsf. If BSD-style group inheritance is in not in use, daemon (or lpd) must be added to the filetsf group. If the group of the files and directories is or can be set to filetsf, then tsfif sets them group readable and writable; otherwise it leaves the permissions conservatively set, on the expectation that ftcheck will complain to a human about the configuration.
The VM/CMS client implementation was fairly simple; it's a REXX script called filetsf (since there's already a sendfile command on VM/CMS) that invokes the netprint command to send the file to the destination machine. Our VM/CMS system also has a command called lpr, but netprint serves the same function and is easier to use. Both lpr and netprint seem to be client-only implementations - they don't provide any spooling if the remote machine is down, which means that we have to manually retry file transfers from VM/CMS if the network or the destination system is down. We have not yet determined the best solution to this problem.
We hope that it will be possible to implement a filetsf server on VM/CMS, but have not yet investigated what it would entail. We're hoping that it will turn out to be a another simple REXX script and a little bit of configuration.
Installation and configuration of filetsf is fairly straightforward. The programs require very little configuration to compile, and contain almost no machine-specific code, or non-standard library routines. The primary exception to this is, of course, aixbe. The programs can be installed just about anywhere in the filesystem.
There should be a sending queue named tsf-machine for each host machine that files will be sent to. Each destination machine needs a queue called filetsf with tsfif as the print filter. Sample entries are shown in figure 1. The filetsf-config(7) man page gives (hopefully) complete details on userids, groups, printcap file entries, and queue configuration for both lpr/lpd based systems and AIX.
Filetsf is intended to be easy to use. This is an very important feature for us, since we wish to convince the application programmers in our department to use this new file transfer system, rather than old familiar FTP. It's hard to convince anyone to use something that's more complicated than what they're already used to.
The simplest example of sending a file is
% sendfile -h otherhost filewhich sends file to the same user on host otherhost.
To distribute a file like /etc/hosts to a different machine, the command would be
% sendfile -h otherhost \ -r -u root /etc/hostsThe -r option indicates that this copy should replace any copy still in the filetsf spool directory on otherhost, and the -u option indicates that the file should be sent to root, regardless of who is sending the file (presumably root is the only user with permission to update the /etc/hosts file). The sendfile command only allows one -h option and one -u option, so multiple invocations of sendfile are required to send a file to more than one recipient. Multiple files can be sent by one sendfile command, so, for example, the /etc/networks file could be sent at the same time as the /etc/hosts file.
The acceptfile command is a little more complicated to use, since it may be necessary to select a particular version of a file. The -l option to acceptfile generates a list of the queued files, such as that shown in figure 2. The output is the recipient userid, file name, sender and the instance of the file, with any repeated information suppressed. The r preceding three of the file instances in the example indicates that those files were sent with the -r option to sendfile. The -c option to acceptfile can be used with -l to list the files in columns, for easier parsing by other programs.
To retrieve a file, acceptfile must be invoked with options that result in the selection of exactly one file. The selection options can also be used with -l in order to test the selection options. For example,
% acceptfile -n hostswould retrieve the newest version of the file hosts into the current directory. The -o option can be used to select the oldest version of a file, so that files can be processed in the order they were sent. A file labeled with a particular time can be selected with the -t option, as in
% acceptfile -t 805922175 hostsFiles can be selected by sender with the -s option, as in
% acceptfile -s \ firstname.lastname@example.org hostsThe retrieved file can be renamed or placed in a different directory with the -r option. Files can be deleted from the spool directory without being retrieved by using the -d option to acceptfile. When a file is retrieved or deleted, acceptfile renames the spooled copy of the file for later removal by ftcleanup. The -p (purge) option can be used to cause the files to be removed from the spool directory immediately.
As with any other business system, a file transfer system needs to be reviewed to ensure that it provides an appropriate level of security and reliability for the applications in which it will be used. We will discuss the security and reliability aspects of filetsf in a step by step review of the file transfer process.
In our environment, there are no new security implications in the act of sending a file. sendfile is not privileged, so the sender of a file must be able to read the file already. Since email and FTP are available on our systems, sendfile does not provide any confidentiality exposures that don't already exist. In the interest of reliability, sendfile does not send any files until it has tested all the files for existence and readability, so that the sending of a set of files is less likely to fail part way through.
The lpr/lpd system is reasonably secure and reasonably reliable - we have a reasonable expectation that we can control which machines have access to our print queues and the data flowing through them. Our administrative systems are on subnets separate from the rest of the campus, our routers can block LPD protocol traffic (if necessary), and we use a locally modified lpr/lpd (a precursor to PLP) that allows us to implement per-queue access controls, so that we can limit the machines that we are willing to trust to exchange files with. This allows us to trust the information in an LPD control file once it reaches the destination machine, so that we can label a file as being sent by a particular sender with a reasonable level of confidence. And lpr/lpd tries to deal correctly with full print spool directories and network interruptions, which allows us to believe that a file that has been passed to lpr/lpd will eventually show up in the right place, without being corrupted along the way.
tsfif is run on the destination machine by lpd, usually as the daemon user, and creates sub-directories in the spool directory and writes the transferred files into the appropriate directory. The primary concern with tsfif is that it needs to be careful that it does not deposit a file in the spool directory over an existing file or symbolic link. This risk is minimal, since the holding directory has no general permissions on it, making it unlikely that someone could create a dangling symbolic link in the spool directory. If tsfif runs into any problems, such as permission or ownership errors, or a full spool directory, it either writes the file with restrictive permissions, or fails with an exit code that causes lpd to keep trying to deliver the file. This means that, for some kinds of errors, the filetsf print queue will stop delivering files. We chose to keep retrying because the only alternative was to give up on a file, and have that particular transfer get lost, and we felt it was very important to protect the files being transferred.
acceptfile is a privileged program; it is setgid to the filetsf group. It is very careful about what it does since it has privileges that an ordinary user does not have, and it allows a user to make use of those privileges. acceptfile is very careful about identifying its invoker, and it renounces its privileges before writing a file to its final destination, so that a user should not be able to change a file with acceptfile that can't already be changed by the user. For reliability, acceptfile renames files in the spool directory, rather than deleting them after delivery. If something goes wrong in later processing, and the delivered file is lost or destroyed, there may be a copy that could be recovered from the filetsf system (by a super-user).
The filetsf system, through the use of the filetsf group, tries to limit its potential exposures. If the filetsf group is somehow compromised, there should be limited exposure to the rest of the machine, since nothing else should be using the filetsf group.
Filetsf isn't 100%secure or 100%reliable; as in any system, the risks have to be weighed against the potential consequences and the costs to eliminate the risks. The filetsf system provides, for us, an acceptable level of security and reliability.
Filetsf was planned to be a simple file transfer system, providing the necessary, basic functions, and providing appropriate building blocks on which to implement more enhanced functionality (we'll claim that this is the ``UNIX philosophy'' again). While filetsf seems to work just as it is, there are a number of potential extensions to filetsf that may be worth investigating.
Some file transfers will likely require greater security than that provided by filetsf, due to greater potential exposures. Since filetsf is just sending files, file encryption, acknowledgement of receipt, checksums, etc., could be built on top of filetsf.
We sometimes have the need to send sets of files, for processing as a unit. The simple approach to handling sets of files is to combine the separate files into a single file, using tar(1), or some other file archiving program. An alternative that might be useful would be to create a getset command, that would check to make sure that the complete set of files are available, retrieve them all, and run a command to process the files. This is something that is simple enough to do, and it may be a common enough operation to justify creating a command to do it.
There may be situations in which the lpr/lpd system or the LPD protocol is either not appropriate or not available. For example, there might be a need to transfer files over larger, public networks in a secure fashion, or through networks that do not allow LPD protocol traffic. It should be fairly easy to implement filetsf on top of a different method of transport, or to use multiple transports depending on the sender and recipient.
Filetsf seems to meet our needs for a basic file transfer system, and is an improvement over our past approaches to file transfer. While not providing extreme levels of security, it is a useful tool on our systems, and we expect to make fairly extensive use of it in the future. It was relatively easy to implement, and, if our needs change in the future, it should be a fairly straightforward task to replace lpr/lpd with some other transport mechanism providing greater reliability or security, while leaving the user interface unchanged.
The VM/CMS client implementation is due to the efforts and knowledge of Cameron McDonald.
John Sellens is a Chartered Accountant and holds a master's degree in Computer Science from the University of Waterloo. He is currently Project Leader, Technical Services in the Data Processing department at the University of Waterloo where he is responsible for system administration and planning, and tool development for the administrative computing systems. John likes to provide as long a list of references as possible in each paper that he writes, whether or not he has actually read whatever it is he's referring to. John can be reached by mail at Data Processing, University of Waterloo, Waterloo, ON N2L 3G1, Canada, or electronically as email@example.com.
The current version of filetsf is available through anonymous FTP to math.uwaterloo.ca as pub/filetsf/filetsf.tar.Z or through the author. Installation and configuration should be fairly straightforward at most sites. Like most freely available software, no support is available for filetsf, but bug fixes and suggested enhancements would be gratefully accepted.