Linux Kernel Architecture

Chapter3:MemoryManagement

❑ If__GFP_ZEROis set,prep_zero_pagefills the page with zero bytes using an efficient,

architecture-specific function.

❑ If__GFP_COMPis set and more than one page has been requested, the kernel must group the

pages intocompound pages. The first page is called thehead page, while all other pages are called

tail pages. The structure of compound pages is shown in Figure 3-33.

All pages are identified as compound pages by thePG_compoundbit. Theprivateelements of

thepageinstance ofallpages — even the head page itself — point to the head page. Besides,

the kernel needs to store information on how to free the compound page. This requires both a

function to free the page and information on how many pages compose the compound page.

The LRU list element of the first tail page is abused for this purpose: A pointer to a destructor

function is thus kept inlru.next, while the allocation order is stored inlru.prev.Noticethat

thelruelement cannot be used for this purpose because it is required if the compound page is

to be kept on a kernel list.

Why is this information required? The kernel can combine multiple adjacent physical pages to a

so-called huge-TLB page. When a userland application works with large chunks of data, many

processors allow using huge-TLB pages to keep the data in memory. Since the page size of a

huge-TLB page is larger than the regular page size, this reduces the amount of information that

must be stored in the translation lookaside buffer (TLB), that, in turn, reduces the probability of

a TLB cache miss — and thus speeds things up.^22 However, huge-TLB pages need to be freed

differently than compound pages composed of multiple regular pages, so an explicit destructor

is required.free_compound_pagesis used for this purpose. The function essentially determines

the page order stored inlru.prevand frees the pages one after another when the compound

page is freed.

The auxiliary functionprep_compound_pageis used to arrange the described structure.

PG_compound

private

Iru.next

Iru.prev

PG_compound

private

Iru.next

2 n pages

Iru.prev

PG_compound

private

PG_compound

private

struct

page

free_compound_page

n

Figure 3-33: Higher-order allocations generate compound pages in which the

individual pages are linked.

The__rmqueueHelper Function

The kernel uses the__rmqueuefunction (whose purpose is evidentfrom the preceding description),

which acts as a gatekeeper to penetrate into the innermost core of the buddy system:

mm/page_alloc.c

static struct page *__rmqueue(struct zone *zone, unsigned int order,

int migratetype)

(^22) Huge-TLB pages are created at boot time and kept in a special cache. The kernel parameterhugepagesallows for specifying
how many huge-TLB pages are to be created, and applications can request them via the special filesystem hugetlbfs. The library
libhugetlbfsallows userland applications to use huge-TLB pageswithout direct interference with this filesystem.