Multics - Novel Ideas

Novel Ideas

Multics implemented a single level store for data access, discarding the clear distinction between files (called segments in Multics) and process memory. The memory of a process consisted solely of segments which were mapped into its address space. To read or write to them, the process simply used normal CPU instructions, and the operating system took care of making sure that all the modifications were saved to disk. In POSIX terminology, it was as if every file was mmaped; however, in Multics there was no concept of process memory, separate from the memory used to hold mapped-in files, as Unix has. All memory in the system was part of some segment, which appeared in the file system; this included the temporary scratch memory of the process, its kernel stack, etc.

One disadvantage of this was that the size of segments was limited to 256 kilowords, giving just over 1 MiB. This was due to the particular hardware architecture of the machines on which Multics ran, having a 36-bit word size (not using the 8-bit byte). Extra code had to be used to work on files larger than this, called multisegment files. In the days when one megabyte of memory was prohibitively expensive, and before large databases and later huge bitmap graphics, this limit was rarely encountered.

Another major new idea of Multics was dynamic linking, in which a running process could request that other segments be added to its address space, segments which could contain code that it could then execute. This allowed applications to automatically use the latest version of any external routine they called, since those routines were kept in other segments, which were dynamically linked only when a process first tried to begin execution in them. Since different processes could use different search rules, different users could end up using different versions of external routines automatically. Equally importantly, with the appropriate settings on the Multics security facilities, the code in the other segment could then gain access to data structures maintained in a different process.

Thus, to interact with an application running in part as a daemon (in another process), a user's process simply performed a normal procedure-call instruction, to a code segment to which it had dynamically linked (a code segment which implemented some operation associated with the daemon). The code in that segment could then modify data maintained and used in the daemon. When the action necessary to commence the request was completed, a simple procedure return instruction returned control of the user's process to the user's code.

The single-level store and dynamic linking are still not available to their full power in other widely used operating systems, despite the rapid and enormous advance in the computer field since the 1960s. They are becoming more widely accepted and available in more limited forms, e.g., dynamic linking.

Multics also supported extremely aggressive on-line reconfiguration; central processing units, memory banks, disk drives, etc. could be added and removed while the system continued operating. At the MIT system, where most early software development was done, it was common practice to split the multiprocessor system into two separate systems during off-hours by incrementally removing enough components to form a second working system, leaving the rest still running the original logged-in users. System software development testing could be done on the second machine, then the components of the second system were added back onto the main user system, without ever having shut it down. Multics supported multiple CPUs: It was one of the earliest multiprocessor systems.

Multics was the first major operating system to be designed as a secure system from the outset. Despite this, early versions of Multics were broken into repeatedly. This led to further work that made the system much more secure and prefigured modern security engineering techniques. Break-ins became very rare once the second-generation hardware base was adopted; it had hardware support for ring-oriented security, a multilevel refinement of the concept of master mode.

Multics was the first operating system to provide a hierarchical file system, and file names could be of almost arbitrary length and syntax. A given file or directory could have multiple names (typically a long and short form), and symbolic links between directories were also supported. Multics was the first to use the now-standard concept of per-process stacks in the kernel, with a separate stack for each security ring. It was also the first to have a command processor implemented as ordinary user code - an idea later used in the Unix shell (although the details are different, since Multics possessed powerful mechanisms which Unix then lacked). It was also one of the first written in a high level language (Multics PL/1), just after the Burroughs MCP system written in ALGOL.