Linux Kernel Architecture

Chapter 18: Page Reclaim and Swapping

occupied by the page is returned to the buddy system. In contrast topagevec_release,this

function assumes that all pages in the vector are not on any LRU list.

All these functions expect apagevecstructure containing the pages to be processed as parameter. If the

vector is empty, all the functions return to the caller immediately.

There are also versions of the same functions with two preceding underscores (e.g.,_ _pagevec_release).

These donottest whether the vector passed contains pages or not.

What is still lacking is a function to add pages to a page vector:

<pagevec.h>

static inline unsigned pagevec_add(struct pagevec *pvec, struct page *page)

pagevec_addadds a new pagepageto a given page vectorpvec.

The implementation of the function is not considered in detail here as it is very straightforward and

reveals little of interest.

The LRU Cache

The kernel provides a further cache known as theLRU cacheto speed up the addition of pages to the

system’s LRU lists. It makes use of page vectors to collectpageinstances and place them block-by-block

on the system’sactiveandinactivelists. The list is a hotspot in the kernel, but must be protected by

a spinlock. To keep lock contention low, new pages are not immediately added to the list, but are first

buffered on a per-CPU list:

mm/swap.c

static DEFINE_PER_CPU(struct pagevec, lru_add_pvecs) = { 0, };

The function to add new elements via this buffer islru_cache_add. It provides a way of deferring the

addition of pages to the system’s LRU lists until a certain number of pages specified byPAGEVEC_SIZE

is reached:

mm/swap.c

void fastcall lru_cache_add(struct page *page)

{

struct pagevec *pvec = &get_cpu_var(lru_add_pvecs);

page_cache_get(page);

if (!pagevec_add(pvec, page))

__pagevec_lru_add(pvec);

put_cpu_var(lru_add_pvecs);

}

Since the function accesses a CPU-specific data structure, it must prevent the kernel from interrupting

execution and resuming later on another CPU. This form of protection is enabled implicitly by invoking

get_cpu_var, which not only disables preemption, but also returns the per-CPU variable.

lru_cache_addfirst increments thecountusage counter of thepageinstance as the page is now in the

page cache (and this is interpreted as usage). The page is then added to the CPU-specific page vector

usingpagevec_add.