Virtual memory is a computer system technique which gives an application program the impression that it has contiguous working memory, while in fact it may be physically fragmented and may even overflow on to disk storage. Systems that use this technique make programming of large applications easier and use real physical memory (e.g. RAM) more efficiently than those without virtual memory.
Note that "virtual memory" is not just "using disk space to extend physical memory size". Extending memory is a normal consequence of using virtual memory techniques, but can be done by other means such as overlays or swapping programs and their data completely out to disk while they are inactive. The definition of "virtual memory" is based on tricking programs into thinking they are using large blocks of contiguous addresses.
All modern general-purpose computer operating systems use virtual memory techniques for ordinary applications, such as word processors, spreadsheets, multimedia players, accounting, etc. Few older operating systems, such as DOS of the 1980s, or those for the mainframes of the 1960s, had virtual memory functionality - notable exceptions being the Atlas and B5000.
Embedded systems and other special-purpose computer systems which require very fast, very consistent response time do not generally use virtual memory.
Systems can have one page table for the whole system or a separate page table for each application. If there is only one, different applications which are running at the same time share a single virtual address space, i.e. they use different parts of a single range of virtual addresses. Systems which use multiple page tables provide multiple virtual address spaces - concurrent applications think they are using the same range of virtual addresses, but their separate page tables redirect to different real addresses.
Most virtual memory systems enable programs to use virtual address ranges which in total exceed the amount of real memory (e.g. RAM). To do this they use disk files to save virtual memory pages which are not currently active, and restore them to real memory when they are needed. Pages are not necessarily restored to the same real addresses from which they were saved - applications are aware only of virtual addresses. Usually when a page is going to be restored to real memory, the real memory already contains another virtual memory page which will be saved to disk before the restore takes place.
In IBM's early virtual memory operating systems virtual=real mode was the only way to "pin down" pages. z/OS has 3 modes, V=V (virtual=virtual; fully pageable), V=R and V=F (virtual = fixed, i.e. "pinned down" but with DAT operating).
Memory is still physically addressed with a single number (called absolute or linear address). To obtain it, the processor looks up the segment number in a segment table to find a segment descriptor. The segment descriptor contains a flag indicating whether the segment is present in main memory and, if it is, the address in main memory of the beginning of the segment (segment's base address) and the length of the segment. It checks whether the offset within the segment is less than the length of the segment and, if it isn't, an interrupt is generated. If a segment is not present in main memory, a hardware interrupt is raised to the operating system, which may try to read the segment into main memory, or to swap in. The operating system might have to remove other segments (swap out) from main memory in order to make room in main memory for the segment to be read in.
Notably, the Intel 80286 supported a similar segmentation scheme as an option, but it was unused by most operating systems.
It is possible to combine segmentation and paging, usually dividing each segment into pages. In systems that combine them, such as Multics and the IBM System/38 and IBM System i machines, virtual memory is usually implemented with paging, with segmentation used to provide memory protection. With the Intel 80386 and later IA-32 processors, the segments reside in a 32-bit linear paged address space, so segments can be moved into and out of that linear address space, and pages in that linear address space can be moved in and out of main memory, providing two levels of virtual memory; however, few if any operating systems do so. Instead, they only use paging.
The difference between virtual memory implementations using pages and using segments is not only about the memory division with fixed and variable sizes, respectively. In some systems, e.g. Multics, or later System/38 and Prime machines, the segmentation was actually visible to the user processes, as part of the semantics of a memory model. In other words, instead of a process just having a memory which looked like a single large vector of bytes or words, it was more structured. This is different from using pages, which doesn't change the model visible to the process. This had important consequences.
A segment wasn't just a "page with a variable length", or a simple way to lengthen the address space (as in Intel 80286). In Multics, the segmentation was a very powerful mechanism that was used to provide a single-level virtual memory model, in which there was no differentiation between "process memory" and "file system" - a process' active address space consisted only a list of segments (files) which were mapped into its potential address space, both code and data. It is not the same as the later mmap function in Unix, because inter-file pointers don't work when mapping files into semi-arbitrary places. Multics had such addressing mode built into most instructions. In other words it could perform relocated inter-segment references, thus eliminating the need for a linker completely. This also worked when different processes mapped the same file into different places in their private address spaces.
The main reason for introducing virtual memory was therefore not simply to extend primary memory, but to make such an extension as easy to use for programmers as possible.
Many systems already had the ability to divide the memory between multiple programs (required for multiprogramming and multiprocessing), provided for example by "base and bounds registers" on early models of the PDP-10, without providing virtual memory. That gave each application a private address space starting at an address of 0, with an address in the private address space being checked against a bounds register to make sure it's within the section of memory allocated to the application and, if it is, having the contents of the corresponding base register being added to it to give an address in main memory. This is a simple form of segmentation without virtual memory.
Virtual memory was developed in approximately 1959–1962, at the University of Manchester for the Atlas Computer, completed in 1962. However, Fritz-Rudolf Güntsch, one of Germany's pioneering computer scientists and later the developer of the Telefunken TR 440 mainframe, claims to have invented the concept in 1957 in his doctoral dissertation Logischer Entwurf eines digitalen Rechengerätes mit mehreren asynchron laufenden Trommeln und automatischem Schnellspeicherbetrieb (Logic Concept of a Digital Computing Device with Multiple Asynchronous Drum Storage and Automatic Fast Memory Mode).
Like many technologies in the history of computing, virtual memory was not accepted without challenge. Before it could be implemented in mainstream operating systems, many models, experiments, and theories had to be developed to overcome the numerous problems. Dynamic address translation required a specialized, expensive, and hard to build hardware, moreover initially it slightly slowed down the access to memory. There were also worries that new system-wide algorithms of utilizing secondary storage would be far less effective than previously used application-specific ones.
By 1969 the debate over virtual memory for commercial computers was over. An IBM research team led by David Sayre showed that the virtual memory overlay system consistently worked better than the best manually controlled systems.
Virtual memory was introduced to the x86 architecture with the protected mode of the Intel 80286 processor. At first it was done with segment swapping, which became inefficient with larger segments. The Intel 80386 introduced support for paging underneath the existing segmentation layer. The page fault exception could be chained with other exceptions without causing a double fault.