Linux Kernel Architecture

Chapter3:MemoryManagement

Registering Active Memory Regions

I noted above that initialization of the zone structures is an extensive task. Luckily this task is identical on

all architectures. While kernel versions before 2.6.19 had to set up the required data structures on a per-

architecture basis, the approach has become more modular in the meantime: The individual architectures

only need to register a very simple map of all active memory regions, and generic code then generates

the main data structures from this information.

Notice that individual architectures can still decide to set up all data structures on their own without

relying on the generic framework provided by the kernel. Since both IA-32 and AMD64 let the kernel do

the hard work, I will not discuss this possibility any further. Any architecture that wants to enjoy the pos-

sibilities offered by the generic framework must set the configuration optionARCH_POPULATES_NODE_MAP.

After all active memory regions are registered, the rest of the work is then performed by the generic

kernel code.

An active memory region is simply a memory region that does not contain any holes.add_active_range

must be used to register a region in the global variableearly_node_map.

mm/page_alloc.c

static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];

static int __meminitdata nr_nodemap_entries;

The number of currently registered regions is denoted bynr_nodemap_entries. The maximal number

of distinct regions is given byMAX_ACTIVE_REGIONS. The value can be set by the architecture-specific

code usingCONFIG_MAX_ACTIVE_REGIONS. If not, the kernel allows for registering 256 active regions per

default (or 50 regions per NUMA node if it is running on a system with more than 32 nodes). Each region

is described by the following data structure:

<mmzone.h>

struct node_active_region {

unsigned long start_pfn;

unsigned long end_pfn;

int nid;

};

start_pfnandend_pfndenote the first and last page frame in a continuous region, andnidis the NUMA

ID of the node to which the memory belongs. UMA systems naturally set this to 0.

An active memory region is registered withadd_active_range:

mm/page_alloc.c

void __init add_active_range(unsigned int nid, unsigned long start_pfn,

unsigned long end_pfn)

When two adjacent regions are registered, thenadd_active_regionsensures that they are merged to a

single one. Besides, the function does not present any surprises.

Recall from Figures 3-12 and 3-13 that the function is called fromzone_sizes_initon IA-32 systems,

and ine820_register_active_regionsonAMD64systems.ThusIwillbrieflydiscussthese

functions.