Linux Kernel Architecture

Chapter3:MemoryManagement

in the cache and allocates it again as quickly as possible (in response to the next request). Only when the

per-CPU caches are empty are free objects from the slabs used to refill them.

This results in a three-level hierarchy for object allocation within which both the allocation cost and the

negative impact of the operation on caches and TLBs rise from level to level:

1. Per-CPU objects in the CPU cache.


2. Unused objects from an existing slab.


3. Unused objects from a new slab just reserved using the buddy system.

Fine Structureof Slabs

Objects are not listed continuously in slabs but are distributed according to a rather complicated scheme.

Figure 3-46 illustrates the details.

Manage-

ment

head

...

BUFCTL_END

Management

of free objects

Color

Space

Unused

Memory Mappings

Free Used Free ...

Fill bytes for

alignment

Object Size

Figure 3-46: Fine structure of a slab.

The size used for each object does not reflect its exact size. Instead, the size is rounded to fulfill certain

alignment criteria. Two alternatives are possible:

❑ Using the flagSLAB_HWCACHE_ALIGNat slab creation time, the slab user can request that objects

are aligned to hardware cache lines. The alignment is then performed along the value returned

bycache_line_size, which returns the processor-specific size of the L1 cache.

If objects are smaller than half of the cache line size, then more than one object is fit into one

cache line.

❑ If alignment along hardware cache lines is not requested, then the kernel ensures that objects are

aligned withBYTES_PER_WORD— the number of bytes needed to represent avoidpointer.

On 32-bit processors, 4 bytes are required for avoidpointer. Consequently, for an object with 6 bytes,

8 = 2 ×4 bytes are needed, and objects with 15 bytes require 16= 4 ×4 bytes. The superfluous bytes are

referred to asfill bytes.