--- Yo! Emacs! This is -*- Text -*- ! --- PART 1 -- What we are doing TOWARDS A NEW STRATEGY OF OPERATING SYSTEM DESIGN The fundamental purpose of an operating system is to enable a variety of programs to share a single computer efficiently. Efficiency demands things such as memory protection, preemptively scheduled timesharing, and coordinated access to I/O peripherals. In addition, operating systems can allow a variety of users to share a computer efficiently. The same general areas apply here as well: preventing users from harming each other, enabling them to share without prior arrangement, and mediating access to physical devices. On today's computer systems, implementing these goals usually involves a large program, called the kernel, which mediates all these exchanges. Since this program must be accessible to all user programs, it becomes a natural place to add functionality to the system. As time goes by, more and more gets added to the kernel, because the only model for interaction between processes is that of specific, individual services provided by the kernel. The traditional system allows users to add components to the mammoth shared library "vmunix" only if they understand most of that library, and only if they have a privileged status within the system. Testing these new components requires a much more painful edit-compile-debug cycle than other programs, and cannot be done while others are using the system. Bugs usually cause fatal system crashes, further disrupting others' use of the system. All of the kernel is usually non-pageable, [fn: There are systems with pageable kernels. However, deciding what can be pageable is difficult and error prone. As well, the mechanisms are usually quite complex, making them difficult to use by people wanting to add simple extensions.] further adding to the cost of adding functions to it. Because of these restrictions, functionality which properly belongs "behind" the wall of the traditional kernel is usually left out of systems unless it is absolutely mandatory. Many good ideas, best implemented with an open/read/write interface cannot be implemented because of the problems inherent in the monolithic nature of traditional systems. Further, even for those with the endurance to implement such ideas, only those who are privileged users of their computers can do so. The software copyright system, preventing unlicensed people from even reading the kernel source, only compounds the problem still more. Some systems have addressed these difficulties. Smalltalk-80 and the Lisp Machine both represent one method of getting around the problem. System code is not distinguished from user code; all of the system is accessible to the user, and can be changed as need be. Both systems were built around languages that facilitated such easy replacement and extension, and were moderately successful. These systems, however, were fairly poor at insulating users and programs from each other, thus failing one of the principle goals of operating system design. Most projects using the Mach 3.0 kernel today carry on the same tradition of operating system design. The internal structure of the system has changed, but the same heavy barrier between user and system remains. The single-servers, while fairly easy to construct, inherit quite naturally all the deficiencies of the monolithic kernels from which they came. Most multi-server projects do somewhat better. Much more of the system is pageable. The system is more easily debugged, and testing of new system components need not interfere with other users. But there is still a wall between the user and the system, and no user can cross that wall without special privilege. The GNU Hurd, by contrast, is designed to make the area of "system" code as limited as possible. Users are only required to communicate with the kernel; the rest of the system is replacable dynamically. Users can use whatever parts of the remainder of the system they want, and can add components themselves, easily, for other users to take advantage of. No mutual trust need exist in advance for users to use each other's services, nor does the system become vulnerable by trusting the services of arbitrary users. We have done this by identifying those components of the system which users must use in order to communicate with each other. One of these is a mechanism responsible for identifying users' identities. This program is called the authentication server. Programs must communicate, each with an authentication server they trust, in order to establish each other's identities. The other is a mechanism for establishing control over system components by the superuser, and global bookkeeping operations. This server is called the process server. No user program needs to communicate with the process server at all; it is only necessary for users which require its services. The authentication server is only necessary for programs which wish to establish their identity to another agent. All remaining services in the system carry no special status. This includes the network implementation, the filesystems, the exec mechanism (including setuid), and so forth. THE TRANSLATOR MECHANISM The Hurd uses Mach ports primarily as methods for communicating between users and servers. Each port implements a particular set of protocols, representing operations that can be undertaken on the underlying object represented by the port. Some of the protocols specified by the Hurd are the I/O protocol, used for generic I/O operations; the file protocol, used for filesystem operations; the socket protocol, used for network operations; the process protocol, used for manipulating processes; and so forth. Most servers are accessed by opening files. Normally, opening a file results in a port associated with that file, owned by the same server as the directory containing the file. For example, a disk based filesystem would normally be serving a large number of ports, each representing an open file or directory. When a new file is opened the server creates a new port, associates it with the file, and returns a send right to the user. However, a file can have a "translator" associated with it. In this case, rather than returning its own port referring to the contents of the file, the server executes a program, called the translator. This program is provided a port to the actual contents of the file, and is then asked to return a port to the original user to complete the open operation. This mechanism is used quite obviously for mount. A mount point would have a translator associated with it. When a user opens the mount point, the translator, in this case, a program which understands the disk format of the mounted filesystem, is executed, and returns a port to the user. Once the translator is started, it need not be run again, unless it dies; the parent filesystem retains a port to the translator to use in further requests. Translators are guaranteed that new programs will not be started for further open operations. Translators can be associated with files by the owner of the file, needing no additional permission. As a result, any program can be specified as a translator. Obviously the system will not work properly if the translator does not implement the file protocol correctly, or at all. The system is so constructed, however, to cause an interruptible hang to be the worst possible result. One way to use translators, then, is to access hierarchically structured data using the file protocol. All the complexity of the user interface to the ftp program, for example, is then removed. Users need only know that a particular directory represents FTP, and can use all the standard file manipulation commands, ls, cp, mv, and so forth, to access the remote system, rather than learning a new set. Similarly, the complexity of tar can be eased by a simple translator. [fn: Such transparent access to tar would not necessarily be cheap, though it would be convenient.] GENERIC SERVICES Another way to use translators is to use the filesystem as a rendezvous for interfaces which are not similar to files. Consider a service which implements some version of the X protocol using Mach messages as an underlying transport. For each X display, a file can be created with the appropriate program as its translator. X clients would open that file. At that point, most file operations (read and write, for example) would not be useful, but new operations (XCreateWindow or XDrawText) might become meaningful. In this kind of case, the filesystem protocol is used further only to manipulate characteristics of the node used for the rendezvous. The node need not support I/O operations, though it should reply to any such messages with a "message not understood" return code. (If MiG stubs are used to demultiplex messages, this will happen automatically.) This technique is used to contact most of the services in the Hurd which are not structured like hierarchical filesystems. For example, the password server, which hands out authorization tags in exchange for passwords, is contacted this way. Network protocol servers as well are contacted in this fashion. Roland McGrath thought up this use of translators. CLEVER FILESYSTEM PICTURES The third common method of using translators in the Hurd is to present a filesystem-like view of another part of the filesystem, with some semantics changed. For example, it might be nice to have a filesystem which cannot be changed, but records changed versions of its files elsewhere. (This might be useful for source code management.) A translator is available which presents "that directory" with all changes going "over there instead". Similarly, a translator is available which creates a directory which is a conceptual union of a number of other directories, with collision resolution rules of various sorts. A variety of other ideas have been presented which do similar things. WHAT THE USER CAN DO None of these translators gain extra privilege by virtue of being hooked into the filesystem. Translators run with the uid of the owner of the file being translated, and can only be set or changed by that owner. The I/O and filesystem protocols are carefully designed to allow their use by mutually untrusting clients and servers. Translators are just ordinary programs. There are a variety of facilities in the GNU C library to make common sorts of translators easier to write. Some translators which might appear to need special privilege are those which allow setuid exec or the password server referred to above. In fact, these translators could be run by anyone. Only if they are set on a root-owned node, however, would they be able to successfully provide all their services. This is analogous to letting any user call the reboot system call, but only honoring it if that user is root. WHY THIS IS SO DIFFERENT What this system organization lets users do is completely novel in the Unixoid world. To this point, operating systems have kept huge portions of this functionality in the realm of system code, thus preventing its modification and extension, except in cases of extreme need. Individual users cannot replace parts of the system in their programs no matter how much easier that would make their task, and system managers are loath to install random tweaks off the net into their kernels. In our system, users can change almost all of the things which are decided for them in advance by traditional systems. In combination with the tremendous control given by the Mach kernel over task address spaces and so forth, for the first time we have a system in which users will be able to replace parts of the system they dislike, without disrupting other users of the same system. Most Mach-based operating systems to date have concentrated, instead, on implementing a wider set of the "same old" Unix semantics in a new environment. By contrast, we are extending those semantics in ways that allow users to bypass or replace them, virtually arbitrarily. PART 2 -- What we are doing, in detail THE AUTHENTICATION SERVER One of the most central servers in the Hurd is the authentication server. Each port to the authentication server identifies a user, and is associated by the authentication server with an "id block". Each id block contains a set of user ids and a set of group ids. Either set may be empty. This server is not the same as the "password server" referred to above. There are three services exported by the authentication server. First, the authentication server provides simple boolean operations on authentication ports. Given a two authentication ports, the authentication server will provide a third port representing the union of the two sets of uids and gids. Second, the authentication server allows any user with a uid of zero to create an arbitrary authentication port. Finally, the authentication server provides RPC's which allow mutually untrusting clients and servers to establish identity and pass initial information on each other. This is crucial to the security of the filesystem and I/O protocols, for example. Any user could write a program which implements the authentication protocol. This does not, however, violate the security of the system. When a given service needs to authenticate a user, it communicates with its trusted authentication server. If that user is using a different authentication server, the transaction will fail, and the server can refuse to communicate further. Because, in effect, this forces all programs on the system to use the same authentication server, we have designed its interface to make any safe operation possible, and to include no extraneous operations. (This is why passwords are implemented in a different server.) THE PROCESS SERVER The process server has undergone much change in the system design. Originally it was to be responsible for much of the mechanics of signal delivery, process creation and destruction, and so forth. We realized, however, that virtually all of these features did not need to be in a global system server. The process server, in the final design, acts as an information categorization repository. There are four main services supported by the process server. First, the process server keeps track of generic host-level information not handled by the Mach kernel. The hostname, the hostid, the system version, and so forth are maintained by the process server. Second, the process server maintains the Posix notions of sessions and process groups, for the convenience of programs which wish to use Posix features. Third, the process server maintains a one-to-one mapping between tasks and processes. Every task is assigned a pid. Processes can register a message port with the process server, which will then be given out to any other program which requests it. The process server makes no attempt to keep these message ports private, so user programs need to implement whatever security they need themselves. (The C Library, which normally implements reception of messages on the message port, provides convenient functions for doing so.) Processes can tell the process server their current argv and envp values; the process server will then provide to those requesting it a vector of the arguments and environment for any process. This is useful in writing ps-like programs, as well as making it easier to hide or change this information. None of these features is mandatory. Programs are free to disregard all of this if they wish, and never register themselves with the process server at all. They will still have a pid assigned, however. Finally, the process server implements "process collections", which are used to collect a number of process message ports at the same time. Also, facilities are provided for converting between pids, process server ports, and Mach task ports, all the while ensuring the security of the ports managed. It is important to stress that that the process server is optional. Because of restrictions in Mach, programs must run as root in order to identify all the tasks in the system, but given that, multiple process servers could co-exist, each with their own clients, giving their own model of the universe. Those process server features which do not require root privileges to be implemented could be done as per-user servers, for those user programs which wish to do recordkeeping in that fashion. The user's hands are not tied. TRANSPARENT FTP Transparent FTP is an intriguing idea whose time has come. The popular ange-ftp package available for GNU Emacs makes access to FTP files virtually transparent to all the emacs file manipulation functions. Transparent FTP does the same thing, but in a system wide fashion. This server is not yet written; the details of its access remain to be fleshed out, and will doubtless change when we have experience using the server. In a BSD kernel, a transparent FTP filesystem would be no harder to write than it is in the Hurd. But mention the idea to a BSD kernel hacker, and the response is that "such a thing doesn't belong in the kernel". In a sense, this is correct. It violates all the layering principles of such a system to place such a thing in the kernel. The unfortunate side effect, however, is that the design methodology (which is based on preventing users from changing things they don't like) is being used to prevent system designers from making things better. In the Hurd, there are no obstacles to doing transparent FTP. A translator will be provided for the node /ftp. The contents of /ftp will not be directly listable, though further subdirectories will. The will have a variety of possible formats. If I want to access some files on uunet, for example, I might do `cd /ftp/ftp.uu.net:anonymous:mib@gnu'. If I want to access some files on an account I have elsewhere, I might do `cd /ftp/unmvax.cs.unm.edu:mike:my-password-here'. Parts of this could be left out, and the transparent FTP program would read them from my .netrc file. In the last case, this would allow me to do simply `cd /ftp/unmvax.cs.unm.edu'; the rest of the information is present in my .netrc file already. There is no need to do a `cd' first--I can use any file commands I want. If I want to implement RFC 1097 (the Telnet Subliminal Message Option), I can just type `more /ftp/ftp.uu.net/inet/rfc/rfc1097'. I can use a copy command if I will need it frequently, or just directly load the file into my emacs. FILESYSTEMS We are implementing ordinary filesystems as well. The initial release of the system will contain a filesystem upward compatible with the Fast File System as found in BSD 4.4. In addition to the ordinary semantics, we will provide the recording of translators, thirty-two bit user ids and group ids, and a new id per file, called the "author" of the file, which can be arbitrarily set by the owner. In addition, because users in the Hurd can have multiple uids, or even none, there is an additional set of permission bits providing access control for "unknown user" (no uids) as distinct from "known but arbitrary user" (some uids). The latter category is the existing "world" category of file permissions. We plan to implement the Network File System protocol, using the 4.4 BSD implementation as a starting point. We will also implement a log-structured filesystem, using the same ideas as the work at Sprite, but probably not the same format. We may design our own network file protocol as well, or we may just extend NFS to rid it of deficiencies. We will also implement various "little" filesystems, such as the MSDOS filesystem, to help people move files between GNU and other operating systems on the same hardware. TERMINALS We will have an I/O server to implement the terminal semantics of Posix. The C Library has features for keeping track of the controlling terminal and arranging to have the proper job control signals sent at the proper times, as well as obeying keyboard and hangup signals from the terminal driver. Programs will be able to insert the terminal driver into communications channels in a variety of ways. Servers like rlogind, for example, will be able to insert the terminal protocol onto their network communication port. Pseudo-terminals will not be necessary, though the will be provided for backward compatibility with older programs--no programs in GNU will depend on them. Nothing about the terminal driver is forced upon users. The terminal driver allows a user to get at the underlying communications channel easily, and to either bypass the terminal driver on an as-needed basis, or entirely, or even substitute a different terminal driver-like program. In the latter case, provided the alternate program implements the necessary interfaces, it will be used by the C Library exactly as if it were the ordinary terminal driver. Because of this flexibility, the original terminal driver will not provide complex line editing features, restricting itself to the behavior found in Posix and BSD. We plan, eventually, to have a readline-like terminal driver, which will provide complex line-editing features for those users who wish to use it. The terminal driver will probably not support high-volume rapid data transmission (such as is required by UUCP or slip) very well. Those programs, however, do not need any of its features, and will be modified to use the underlying Mach device ports for terminals, which do efficiently support moving large amounts of data. EXECUTING PROGRAMS The mechanics of implementing the execve call are distributed between three programs. The library handles marshalling the argument and environment vectors. It then sends a message to the file server holding the file to be executed. The file server checks execute permissions, making whatever changes it desires in the exec call. For example, if the file is marked setuid and the fileserver has the ability, it will change the user identification going to the new image appropriately. The file server also decides if programs which had access to the old task should continue to have access to the new task. If the file server is augmenting permissions, or executing an unreadable image, for example, then the exec needs to take place in a new Mach task to maintain security. (The process server contains some special features to allow the new task to keep the pid of the old task in this case, thus preserving Posix semantics.) After deciding the policy associated with the new image, the filesystem calls the exec server to load the task. The exec server, written using the GNU BFD (Binary File Descriptor) library, loads the image. BFD supports a large number of object file formats; almost any object file format it supports will be executable. The exec server also recognizes scripts starting with `#!' and does the right thing for them. It was thought that it would be nice to make it easy for users to have their own exec servers. Roland McGrath thought of the following technique. The standard exec server also looks at the environment of the new image; if it contains a variable EXEC_SERVERS then it uses the programs specified there as exec servers instead of the system default. (This is, of course, not done for execs that the file server has requested be kept secure.) The new image starts running in the C Library, which sends a message to the exec server to get the arguments, environment, and other state (umask, current directory, etc.). None of this additional state is special to the file server or the exec server; if programs wish, they can use it in a different manner than the C Library. NEW PROCESSES The fork call is implemented almost entirely in the C Library. The new task is created by Mach kernel calls. The C Library arranges to have its image inherited properly. The new task is registered with the process server (though this is not mandatory). The C Library provides vectors of functions to be called at fork time: one vector to be called before the fork, one after in the parent, and one after in the child. This should not be used to replace the normal sequence of calling fork; it is intended for libraries which need to close ports or clean up lock state before a fork occurs. The library will implement both fork calls specified by the draft Posix.4a (the threads extension to the real-time extension). Nothing forces the user to create new tasks this way. If a program wants to use almost the normal fork, but with some special characteristics, then it can do so. Hooks will be provided by the C Library; in addition, the function can even be completely replaced if desired. None of this is possible in a Unix system. ASYNCHRONOUS MESSAGES As mentioned above, the process server maintains a "message port" for each task registered with it. These ports are public, and are used to send asynchronous messages to the task. Signals, for example, are sent to the message port. The signal message also provides a port as an indication that the sender should be trusted to send the signal. The C Library maintains a variety of ports in a table, each of which identifies a set of signals that can be sent by anyone who posesses that port. For example, if the user possess the task's kernel port, it is allowed to send any signal. If the user possesses a special "terminal id" port, it is allowed to send the keyboard and hangup signals. Users can add arbitrary new entries into the C Libraries signal permissions table. When a process's process group changes, the process server will send it a message indicating the new process group. In this case, the process server proves its authority by providing the task's kernel port. The library also implements messages to add and delete uids currently used by the process. If new uids are sent to the program, the library adds them to its current set, and then exchanges message with all the I/O servers it knows about, proving to them its new authorization. Similarly, uids can be deleted with a message. In the latter case, the caller must provide the process's task port. (You can't harm a process by giving it extra permission, but you can harm it by taking permission away.) We will provide user programs to send these messages to processes. This will enable the `su' command to cause all the programs in your current login session to gain a new uid, rather than spawn a subshell. The library will allow programs to add asynchronous messages they wish to recognize, as well as prevent recognition of the standard set. MAKING IT LOOK LIKE UNIX The C Library will implement all the calls from BSD and Posix, as well as some obvious extensions to them. This enables users to replace those calls they dislike, or bypass them entirely, whereas in Unix, the calls must be used "as they come", with no alternatives possible. On some environments, we will support binary compatibility as well. This works by building a special version of the library. This is then loaded somewhere in the address space of the process. (On, a Vax, for example, it would be tucked in above the stack.) A feature of Mach, called system call redirection is then used to trap Unix system calls and turn them into jumps into this special version of the library. (On almost all machines, the cost of such a redirection is very small; this is a highly optimized path of Mach. On a 386 it's about two dozen instructions.) This is only slightly worse than a simple procedure call. Many features of Unix, such as signal masks and vectors, are handled completely by the library. This makes such calls significantly cheaper than in Unix. It is now reasonable to use sigblock extensively to protect critical sections, rather than seeking out some other, less expensive method. NETWORK PROTOCOLS We are writing a library that will make it very easy to port 4.4 BSD protocol stacks into the Hurd. This will enable operation, virtually for free, of all the protocols supported by BSD. Currently, this includes the CCITT protocols, the TCP/IP protocols, the Xerox NS protocols, and the ISO protocols. For maximal performace, some work would be necessary to take advantage of Hurd features that provide for very high speed I/O. Generally, for most protocols, this would require some thought, but not too much time. We intend to spend effort making the TCP/IP protocols run as efficiently as possible. As an interesting example of the flexibility of the Hurd design, consider the case of IP trailers, used extensively in BSD for performance. While the Hurd will be willing to send and receive trailers, it will gain fairly little advantage in doing so, because there is no requirement that data be copied *ever*, so avoiding copies for page-aligned data is not important.