VMEM(9) Kernel Concepts VMEM(9)

vmem
virtual memory allocator

An address space is divided into a number of logically distinct pieces, or arenas: text, data, heap, stack, and so on. Within these arenas we often subdivide further; for example, we use heap addresses not only for the kernel heap (kmem_alloc() space), but also for DVMA, bp_mapin(), /dev/kmem, and even some device mappings.

The kernel address space, therefore, is most accurately described as a tree of arenas in which each node of the tree imports some subset of its parent. The virtual memory allocator manages these arenas and supports their natural hierarchical structure.

An arena is nothing more than a set of integers. These integers most commonly represent virtual addresses, but in fact they can represent anything at all. For example, we could use an arena containing the integers minpid through maxpid to allocate process IDs. For uses of this nature, prefer id_space(9F) instead.

vmem_create() and vmem_destroy() create and destroy vmem arenas. In order to differentiate between arenas used for addresses and arenas used for identifiers, the VMC_IDENTIFIER flag is passed to vmem_create(). This prevents identifier exhaustion from being diagnosed as general memory failure.

We represent the integers in an arena as a collection of spans, or contiguous ranges of integers. For example, the kernel heap consists of just one span: [kernelheap, ekernelheap). Spans can be added to an arena in two ways: explicitly, by vmem_add(); or implicitly, by importing, as described in Imported Memory below.

Spans are subdivided into segments, each of which is either allocated or free. A segment, like a span, is a contiguous range of integers. Each allocated segment [addr, addr + size) represents exactly one vmem_alloc(size) that returned addr. Free segments represent the space between allocated segments. If two free segments are adjacent, we coalesce them into one larger segment; that is, if segments [a, b) and [b, c) are both free, we merge them into a single segment [a, c). The segments within a span are linked together in increasing-address order so we can easily determine whether coalescing is possible.

Segments never cross span boundaries. When all segments within an imported span become free, we return the span to its source.

As mentioned in the overview, some arenas are logical subsets of other arenas. For example, kmem_va_arena (a virtual address cache that satisfies most kmem_slab_create() requests) is just a subset of heap_arena (the kernel heap) that provides caching for the most common slab sizes. When kmem_va_arena runs out of virtual memory, it imports more from the heap; we say that heap_arena is the vmem source for kmem_va_arena. vmem_create() allows you to specify any existing vmem arena as the source for your new arena. Topologically, since every arena is a child of at most one source, the set of all arenas forms a collection of trees.

Some vmem clients are quite picky about the kind of address they want. For example, the DVMA code may need an address that is at a particular phase with respect to some alignment (to get good cache coloring), or that lies within certain limits (the addressable range of a device), or that doesn't cross some boundary (a DMA counter restriction) — or all of the above. vmem_xalloc() allows the client to specify any or all of these constraints.

Every arena has a notion of ‘quantum’, specified at vmem_create() time, that defines the arena's minimum unit of currency. Most commonly the quantum is either 1 or PAGESIZE, but any power of 2 is legal. All vmem allocations are guaranteed to be quantum-aligned.

Every kmem cache has a vmem arena as its slab supplier. The kernel memory allocator uses vmem_alloc() and vmem_free() to create and destroy slabs.

id_space(9F), vmem_add(9F), vmem_alloc(9F), vmem_contains(9F), vmem_create(9F), vmem_walk(9F)

Jeff Bonwick and Jonathan Adams, Magazines and vmem: Extending the Slab Allocator to Many CPUs and Arbitrary Resources., Proceedings of the 2001 Usenix Conference, http://www.usenix.org/event/usenix01/bonwick.html.

January 18, 2017 OmniOS