Plan 9 from Bell Labs
Murray Hill, New Jersey 07974
By the mid 1980’s, the trend in computing was away from large centralized time-shared computers towards networks of smaller, personal machines, typically UNIX ‘workstations’. People had grown weary of overloaded, bureaucratic timesharing machines and were eager to move to small, self-maintained systems, even if that meant a net loss in computing power. As microcomputers became faster, even that loss was recovered, and this style of computing remains popular today.
In the rush to personal workstations, though, some of their weaknesses were overlooked. First, the operating system they run, UNIX, is itself an old timesharing system and has had trouble adapting to ideas born after it. Graphics and networking were added to UNIX well into its lifetime and remain poorly integrated and difficult to administer. More important, the early focus on having private machines made it difficult for networks of machines to serve as seamlessly as the old monolithic timesharing systems. Timesharing centralized the management and amortization of costs and resources; personal computing fractured, democratized, and ultimately amplified administrative problems. The choice of an old timesharing operating system to run those personal machines made it difficult to bind things together smoothly.
Plan 9 began in the late 1980’s as an attempt to have it both ways: to build a system that was centrally administered and cost-effective using cheap modern microcomputers as its computing elements. The idea was to build a time-sharing system out of workstations, but in a novel way. Different computers would handle different tasks: small, cheap machines in people’s offices would serve as terminals providing access to large, central, shared resources such as computing servers and file servers. For the central machines, the coming wave of shared-memory multiprocessors seemed obvious candidates. The philosophy is much like that of the Cambridge Distributed System [NeHe82]. The early catch phrase was to build a UNIX out of a lot of little systems, not a system out of a lot of little UNIXes.
The problems with UNIX were too deep to fix, but some of its ideas could be brought along. The best was its use of the file system to coordinate naming of and access to resources, even those, such as devices, not traditionally treated as files. For Plan 9, we adopted this idea by designing a network-level protocol, called 9P, to enable machines to access files on remote systems. Above this, we built a naming system that lets people and their computing agents build customized views of the resources in the network. This is where Plan 9 first began to look different: a Plan 9 user builds a private computing environment and recreates it wherever desired, rather than doing all computing on a private machine. It soon became clear that this model was richer than we had foreseen, and the ideas of per-process name spaces and file-system-like resources were extended throughout the system—to processes, graphics, even the network itself.
By 1989 the system had become solid enough that some of us began using it as our exclusive computing environment. This meant bringing along many of the services and applications we had used on UNIX. We used this opportunity to revisit many issues, not just kernel-resident ones, that we felt UNIX addressed badly. Plan 9 has new compilers, languages, libraries, window systems, and many new applications. Many of the old tools were dropped, while those brought along have been polished or rewritten.
Why be so all-encompassing? The distinction between operating system, library, and application is important to the operating system researcher but uninteresting to the user. What matters is clean functionality. By building a complete new system, we were able to solve problems where we thought they should be solved. For example, there is no real ‘tty driver’ in the kernel; that is the job of the window system. In the modern world, multi-vendor and multi-architecture computing are essential, yet the usual compilers and tools assume the program is being built to run locally; we needed to rethink these issues. Most important, though, the test of a system is the computing environment it provides. Producing a more efficient way to run the old UNIX warhorses is empty engineering; we were more interested in whether the new ideas suggested by the architecture of the underlying system encourage a more effective way of working. Thus, although Plan 9 provides an emulation environment for running POSIX commands, it is a backwater of the system. The vast majority of system software is developed in the ‘native’ Plan 9 environment.
There are benefits to having an all-new system. First, our laboratory has a history of building experimental peripheral boards. To make it easy to write device drivers, we want a system that is available in source form (no longer guaranteed with UNIX, even in the laboratory in which it was born). Also, we want to redistribute our work, which means the software must be locally produced. For example, we could have used some vendors’ C compilers for our system, but even had we overcome the problems with cross-compilation, we would have difficulty redistributing the result.
This paper serves as an overview of the system. It discusses the architecture from the lowest building blocks to the computing environment seen by users. It also serves as an introduction to the rest of the Plan 9 Programmer’s Manual, which it accompanies. More detail about topics in this paper can be found elsewhere in the manual.
The view of the system is built upon three principles. First, resources are named and accessed like files in a hierarchical file system. Second, there is a standard protocol, called 9P, for accessing these resources. Third, the disjoint hierarchies provided by different services are joined together into a single private hierarchical file name space. The unusual properties of Plan 9 stem from the consistent, aggressive application of these principles.
A large Plan 9 installation has a number of computers networked together, each providing a particular class of service. Shared multiprocessor servers provide computing cycles; other large machines offer file storage. These machines are located in an air-conditioned machine room and are connected by high-performance networks. Lower bandwidth networks such as Ethernet or ISDN connect these servers to office- and home-resident workstations or PCs, called terminals in Plan 9 terminology. Figure 1 shows the arrangement.
The modern style of computing offers each user a dedicated workstation or PC. Plan 9’s approach is different. The various machines with screens, keyboards, and mice all provide access to the resources of the network, so they are functionally equivalent, in the manner of the terminals attached to old timesharing systems. When someone uses the system, though, the terminal is temporarily personalized by that user. Instead of customizing the hardware, Plan 9 offers the ability to customize one’s view of the system provided by the software. That customization is accomplished by giving local, personal names for the publicly visible resources in the network. Plan 9 provides the mechanism to assemble a personal view of the public space with local names for globally accessible resources. Since the most important resources of the network are files, the model of that view is file-oriented.
The client’s local name space provides a way to customize the user’s view of the network. The services available in the network all export file hierarchies. Those important to the user are gathered together into a custom name space; those of no immediate interest are ignored. This is a different style of use from the idea of a ‘uniform global name space’. In Plan 9, there are known names for services and uniform names for files exported by those services, but the view is entirely local. As an analogy, consider the difference between the phrase ‘my house’ and the precise address of the speaker’s home. The latter may be used by anyone but the former is easier to say and makes sense when spoken. It also changes meaning depending on who says it, yet that does not cause confusion. Similarly, in Plan 9 the name /dev/cons always refers to the user’s terminal and /bin/date the correct version of the date command to run, but which files those names represent depends on circumstances such as the architecture of the machine executing date. Plan 9, then, has local name spaces that obey globally understood conventions; it is the conventions that guarantee sane behavior in the presence of local names.
The 9P protocol is structured as a set of transactions that send a request from a client to a (local or remote) server and return the result. 9P controls file systems, not just files: it includes procedures to resolve file names and traverse the name hierarchy of the file system provided by the server. On the other hand, the client’s name space is held by the client system alone, not on or with the server, a distinction from systems such as Sprite [OCDNW88]. Also, file access is at the level of bytes, not blocks, which distinguishes 9P from protocols like NFS and RFS. A paper by Welch compares Sprite, NFS, and Plan 9’s network file system structures [Welc94].
This approach was designed with traditional files in mind, but can be extended to many other resources. Plan 9 services that export file hierarchies include I/O devices, backup services, the window system, network interfaces, and many others. One example is the process file system, /proc, which provides a clean way to examine and control running processes. Precursor systems had a similar idea [Kill84], but Plan 9 pushes the file metaphor much further [PPTTW93]. The file system model is well-understood, both by system builders and general users, so services that present file-like interfaces are easy to build, easy to understand, and easy to use. Files come with agreed-upon rules for protection, naming, and access both local and remote, so services built this way are ready-made for a distributed system. (This is a distinction from ‘object-oriented’ models, where these issues must be faced anew for every class of object.) Examples in the sections that follow illustrate these ideas in action.
The Command-level View
Plan 9 is meant to be used from a machine with a screen running the window system. It has no notion of ‘teletype’ in the UNIX sense. The keyboard handling of the bare system is rudimentary, but once the window system, 8½ [Pike91], is running, text can be edited with ‘cut and paste’ operations from a pop-up menu, copied between windows, and so on. 8½ permits editing text from the past, not just on the current input line. The text-editing capabilities of 8½ are strong enough to displace special features such as history in the shell, paging and scrolling, and mail editors. 8½ windows do not support cursor addressing and, except for one terminal emulator to simplify connecting to traditional systems, there is no cursor-addressing software in Plan 9.
Each window is created in a separate name space. Adjustments made to the name space in a window do not affect other windows or programs, making it safe to experiment with local modifications to the name space, for example to substitute files from the dump file system when debugging. Once the debugging is done, the window can be deleted and all trace of the experimental apparatus is gone. Similar arguments apply to the private space each window has for environment variables, notes (analogous to UNIX signals), etc.
Each window is created running an application, such as the shell, with standard input and output connected to the editable text of the window. Each window also has a private bitmap and multiplexed access to the keyboard, mouse, and other graphical resources through files like /dev/mouse, /dev/bitblt, and /dev/cons (analogous to UNIX’s /dev/tty). These files are provided by 8½, which is implemented as a file server. Unlike X windows, where a new application typically creates a new window to run in, an 8½ graphics application usually runs in the window where it starts. It is possible and efficient for an application to create a new window, but that is not the style of the system. Again contrasting to X, in which a remote application makes a network call to the X server to start running, a remote 8½ application sees the mouse, bitblt, and cons files for the window as usual in /dev; it does not know whether the files are local. It just reads and writes them to control the window; the network connection is already there and multiplexed.
The intended style of use is to run interactive applications such as the window system and text editor on the terminal and to run computation- or file-intensive applications on remote servers. Different windows may be running programs on different machines over different networks, but by making the name space equivalent in all windows, this is transparent: the same commands and resources are available, with the same names, wherever the computation is performed.
The command set of Plan 9 is similar to that of UNIX. The commands fall into several broad classes. Some are new programs for old jobs: programs like ls, cat, and who have familiar names and functions but are new, simpler implementations. Who, for example, is a shell script, while ps is just 95 lines of C code. Some commands are essentially the same as their UNIX ancestors: awk, troff, and others have been converted to ANSI C and extended to handle Unicode, but are still the familiar tools. Some are entirely new programs for old niches: the shell rc, text editor sam, debugger acid, and others displace the better-known UNIX tools with similar jobs. Finally, about half the commands are new.
Compatibility was not a requirement for the system. Where the old commands or notation seemed good enough, we kept them. When they didn’t, we replaced them.
The File Server
A central file server stores permanent files and presents them to the network as a file hierarchy exported using 9P. The server is a stand-alone system, accessible only over the network, designed to do its one job well. It runs no user processes, only a fixed set of routines compiled into the boot image. Rather than a set of disks or separate file systems, the main hierarchy exported by the server is a single tree, representing files on many disks. That hierarchy is shared by many users over a wide area on a variety of networks. Other file trees exported by the server include special-purpose systems such as temporary storage and, as explained below, a backup service.
The file server has three levels of storage. The central server in our installation has about 100 megabytes of memory buffers, 27 gigabytes of magnetic disks, and 350 gigabytes of bulk storage in a write-once-read-many (WORM) jukebox. The disk is a cache for the WORM and the memory is a cache for the disk; each is much faster, and sees about an order of magnitude more traffic, than the level it caches. The addressable data in the file system can be larger than the size of the magnetic disks, because they are only a cache; our main file server has about 40 gigabytes of active storage.
The most unusual feature of the file server comes from its use of a WORM device for stable storage. Every morning at 5 o’clock, a dump of the file system occurs automatically. The file system is frozen and all blocks modified since the last dump are queued to be written to the WORM. Once the blocks are queued, service is restored and the read-only root of the dumped file system appears in a hierarchy of all dumps ever taken, named by its date. For example, the directory /n/dump/1995/0315 is the root directory of an image of the file system as it appeared in the early morning of March 15, 1995. It takes a few minutes to queue the blocks, but the process to copy blocks to the WORM, which runs in the background, may take hours.
There are two ways the dump file system is used. The first is by the users themselves, who can browse the dump file system directly or attach pieces of it to their name space. For example, to track down a bug, it is straightforward to try the compiler from three months ago or to link a program with yesterday’s library. With daily snapshots of all files, it is easy to find when a particular change was made or what changes were made on a particular date. People feel free to make large speculative changes to files in the knowledge that they can be backed out with a single copy command. There is no backup system as such; instead, because the dump is in the file name space, backup problems can be solved with standard tools such as cp, ls, grep, and diff.
The other (very rare) use is complete system backup. In the event of disaster, the active file system can be initialized from any dump by clearing the disk cache and setting the root of the active file system to be a copy of the dumped root. Although easy to do, this is not to be taken lightly: besides losing any change made after the date of the dump, this recovery method results in a very slow system. The cache must be reloaded from WORM, which is much slower than magnetic disks. The file system takes a few days to reload the working set and regain its full performance.
Access permissions of files in the dump are the same as they were when the dump was made. Normal utilities have normal permissions in the dump without any special arrangement. The dump file system is read-only, though, which means that files in the dump cannot be written regardless of their permission bits; in fact, since directories are part of the read-only structure, even the permissions cannot be changed.
Once a file is written to WORM, it cannot be removed, so our users never see ‘‘please clean up your files’’ messages and there is no df command. We regard the WORM jukebox as an unlimited resource. The only issue is how long it will take to fill. Our WORM has served a community of about 50 users for five years and has absorbed daily dumps, consuming a total of 65% of the storage in the jukebox. In that time, the manufacturer has improved the technology, doubling the capacity of the individual disks. If we were to upgrade to the new media, we would have more free space than in the original empty jukebox. Technology has created storage faster than we can use it.
Unusual file servers
Plan 9 is characterized by a variety of servers that offer a file-like interface to unusual services. Many of these are implemented by user-level processes, although the distinction is unimportant to their clients; whether a service is provided by the kernel, a user process, or a remote server is irrelevant to the way it is used. There are dozens of such servers; in this section we present three representative ones.
Perhaps the most remarkable file server in Plan 9 is 8½, the window system. It is discussed at length elsewhere [Pike91], but deserves a brief explanation here. 8½ provides two interfaces: to the user seated at the terminal, it offers a traditional style of interaction with multiple windows, each running an application, all controlled by a mouse and keyboard. To the client programs, the view is also fairly traditional: programs running in a window see a set of files in /dev with names like mouse, screen, and cons. Programs that want to print text to their window write to /dev/cons; to read the mouse, they read /dev/mouse. In the Plan 9 style, bitmap graphics is implemented by providing a file /dev/bitblt on which clients write encoded messages to execute graphical operations such as bitblt (RasterOp). What is unusual is how this is done: 8½ is a file server, serving the files in /dev to the clients running in each window. Although every window looks the same to its client, each window has a distinct set of files in /dev. 8½ multiplexes its clients’ access to the resources of the terminal by serving multiple sets of files. Each client is given a private name space with a different set of files that behave the same as in all other windows. There are many advantages to this structure. One is that 8½ serves the same files it needs for its own implementation—it multiplexes its own interface—so it may be run, recursively, as a client of itself. Also, consider the implementation of /dev/tty in UNIX, which requires special code in the kernel to redirect open calls to the appropriate device. Instead, in 8½ the equivalent service falls out automatically: 8½ serves /dev/cons as its basic function; there is nothing extra to do. When a program wants to read from the keyboard, it opens /dev/cons, but it is a private file, not a shared one with special properties. Again, local name spaces make this possible; conventions about the consistency of the files within them make it natural.
8½ has a unique feature made possible by its design. Because it is implemented as a file server, it has the power to postpone answering read requests for a particular window. This behavior is toggled by a reserved key on the keyboard. Toggling once suspends client reads from the window; toggling again resumes normal reads, which absorb whatever text has been prepared, one line at a time. This allows the user to edit multi-line input text on the screen before the application sees it, obviating the need to invoke a separate editor to prepare text such as mail messages. A related property is that reads are answered directly from the data structure defining the text on the display: text may be edited until its final newline makes the prepared line of text readable by the client. Even then, until the line is read, the text the client will read can be changed. For example, after typing
to the shell, the user can backspace over the final newline at any time until make finishes, holding off execution of the rm command, or even point with the mouse before the rm and type another command to be executed first.
There is no ftp command in Plan 9. Instead, a user-level file server called ftpfs dials the FTP site, logs in on behalf of the user, and uses the FTP protocol to examine files in the remote directory. To the local user, it offers a file hierarchy, attached to /n/ftp in the local name space, mirroring the contents of the FTP site. In other words, it translates the FTP protocol into 9P to offer Plan 9 access to FTP sites. The implementation is tricky; ftpfs must do some sophisticated caching for efficiency and use heuristics to decode remote directory information. But the result is worthwhile: all the local file management tools such as cp, grep, diff, and of course ls are available to FTP-served files exactly as if they were local files. Other systems such as Jade and Prospero have exploited the same opportunity [Rao81, Neu92], but because of local name spaces and the simplicity of implementing 9P, this approach fits more naturally into Plan 9 than into other environments.
One server, exportfs, is a user process that takes a portion of its own name space and makes it available to other processes by translating 9P requests into system calls to the Plan 9 kernel. The file hierarchy it exports may contain files from multiple servers. Exportfs is usually run as a remote server started by a local program, eith