The support of large sparse address spaces, mapped files, and shared memory was a requirement for 4.2BSD. An interface was specified, called mmap, that allowed unrelated processes to request a shared mapping of a file into their address spaces. If multiple processes mapped the same file into their address spaces, changes to the file's portion of an address space by one process would be reflected in the area mapped by the other processes, as well as in the file itself. Ultimately, 4.2BSD was shipped without the mmap interface, because of pressure to make other features, such as networking, available.

Further development of the mmap interface continued during the work on 4.3BSD. Over 40 companies and research groups participated in the discussions leading to the revised architecture that was described in the Berkeley Software Architecture Manual [McKusick et al, 1994]. Several of the companies have implemented the revised interface [Gingell et al, 1987].

Once again, time pressure prevented 4.3BSD from providing an implementation of the interface. Although the latter could have been built into the existing 4.3BSD virtual-memory system, the developers decided not to put it in because that implementation was nearly 10 years old. Furthermore, the original virtual-memory design was based on the assumption that computer memories were small and expensive, whereas disks were locally connected, fast, large, and inexpensive. Thus, the virtual-memory system was designed to be frugal with its use of memory at the expense of generating extra disk traffic. In addition, the 4.3BSD implementation was riddled with VAX memory-management hardware dependencies that impeded its portability to other computer architectures. Finally, the virtual-memory system was not designed to support the tightly coupled multiprocessors that are becoming increasingly common and important today.

Attempts to improve the old implementation incrementally seemed doomed to failure. A completely new design, on the other hand, could take advantage of large memories, conserve disk transfers, and have the potential to run on multiprocessors. Consequently, the virtual-memory system was completely replaced in 4.4BSD. The 4.4BSD virtual-memory system is based on the Mach 2.0 VM system [Tevanian, 1987]. with updates from Mach 2.5 and Mach 3.0. It features efficient support for sharing, a clean separation of machine-independent and machine-dependent features, as well as (currently unused) multiprocessor support. Processes can map files anywhere in their address space. They can share parts of their address space by doing a shared mapping of the same file. Changes made by one process are visible in the address space of the other process, and also are written back to the file itself. Processes can also request private mappings of a file, which prevents any changes that they make from being visible to other processes mapping the file or being written back to the file itself.

Another issue with the virtual-memory system is the way that information is passed into the kernel when a system call is made. 4.4BSD always copies data from the process address space into a buffer in the kernel. For read or write operations that are transferring large quantities of data, doing the copy can be time consuming. An alternative to doing the copying is to remap the process memory into the kernel. The 4.4BSD kernel always copies the data for several reasons:

The biggest incentives for memory mapping are the needs for accessing big files and for passing large quantities of data between processes. The mmap interface provides a way for both of these tasks to be done without copying.

The kernel often does allocations of memory that are needed for only the duration of a single system call. In a user process, such short-term memory would be allocated on the run-time stack. Because the kernel has a limited run-time stack, it is not feasible to allocate even moderate-sized blocks of memory on it. Consequently, such memory must be allocated through a more dynamic mechanism. For example, when the system must translate a pathname, it must allocate a 1-Kbyte buffer to hold the name. Other blocks of memory must be more persistent than a single system call, and thus could not be allocated on the stack even if there was space. An example is protocol-control blocks that remain throughout the duration of a network connection.

Demands for dynamic memory allocation in the kernel have increased as more services have been added. A generalized memory allocator reduces the complexity of writing code inside the kernel. Thus, the 4.4BSD kernel has a single memory allocator that can be used by any part of the system. It has an interface similar to the C library routines malloc and free that provide memory allocation to application programs [McKusick & Karels, 1988]. Like the C library interface, the allocation routine takes a parameter specifying the size of memory that is needed. The range of sizes for memory requests is not constrained; however, physical memory is allocated and is not paged. The free routine takes a pointer to the storage being freed, but does not require the size of the piece of memory being freed.