introduction to kernel memory allocators

The NetBSD kernel provides several memory allocators, each with different characteristics and purpose. This document summarizes the main differences between them.
You should use the kmem(9) allocator for all allocations unless you have special needs that it does not provide, such as:

The kmem(9) allocator is main general purpose allocator in the kernel. It was modelled after an interface of the same name implemented in Solaris.
kmem(9) is fast and requires no setup. It cannot be used from interrupt context.
Internally, kmem(9) is implemented using a collection of pool caches for common small allocation sizes, so there is no performance benefit to using a pool cache if you have no other needs.

The pool(9) allocator is a fixed-size memory allocator which requires setup to initialize a shared pool.
A pool can be configured with a low-water mark to reserve a minimum number of objects available, a high-water mark to bound the maximum number of objects in reserve, and a hard limit to bound on the maximum number of objects in use.
pool_get(9) can be used to allocate memory in interrupt context for objects that have been reserved in advance, with the possibility of failure if there are none.
By default, pool(9) allocates pageable RAM-backed kernel virtual address space from the same backing store as kmem(9), but it can be configured to allocate any kind of resource with a custom allocator.

The pool cache allocator is a per-CPU cache on top of pool(9) for fixed-size memory allocations that may occur in interrupt context requiring setup beforehand.
The per-CPU cache makes allocation much cheaper — no interprocessor synchronization in the fast case — at the cost of potentially caching some extra resources on one CPU that cannot be used by another.
In addition to all the features of a pool like a custom backing allocator, a pool cache also supports a constructor and destructor routine for when objects are drawn from the shared pool in case the per-CPU cache is empty, or returned to it when the cache is full. This can reduce the cost of reusable initialization and finalization, or associate objects with CPU-local resources.

The uvm_km(9) API is a low-level memory allocator for page-aligned kernel virtual address space in multiples of PAGE_SIZE, with wired RAM backing, pageable RAM backing, or backing to be supplied by the caller with pmap(9).

The vmem(9) API is a general address space allocator. It is used internally by kmem(9), pool(9), uvm(9), and other kernel subsystems and device drivers to allocate regions of various kinds of address spaces. Internally, it allocates large chunks of the address space and uses a pool_cache(9) to draw small allocations out of them.

The extent(9) API manages and allocates constrained regions of an address space. The extent manager is optimized for simplicity, not speed, and is available early at boot. NetBSD uses extent(9) to reserve regions of I/O port and memory spaces to prevent drivers from using the same device registers or bus memory.

bus_space(9), extent(9), intro(9), kmem(9), pool(9), pool_cache(9), uvm(9), uvm_km(9), vmem(9)

Elad Efrat <>
YAMAMOTO Takashi <>
Taylor R Campbell <>
October 28, 2017 NetBSD-current