Linux Kernel Architecture

Chapter 18: Page Reclaim and Swapping

this point, the kernel must nevertheless consider which area it wants to select for the page and into which

slot it will be inserted. This decision must then be saved in the data structures of the swap cache.

Page fault management

Page Cache Management Reverse Mapping

do_swap_page

Architecture-specific

page fault handler

swapin_readahead

read_swap_cache_async

_ _add_to_swap_cache

add_to_swap

Page Reclaim

Tranfer swap data

to swap space and

handle page tables

shrink_page_list

Zone scanning

try_to_unmap

try_to_unmap_one

swap_writepage

Figure 18-7: Overview of the most important functions that implement swapping and page reclaim. The

figure is not a proper code flow diagram and skips some intermediate functions.

Two kernel methods add pages to the swap cache but serve different purposes:

1. add_to_swapis invoked when the kernel wants to swap out a pageactively;thatis,whenthe
   policy algorithm has determined that insufficient memory is available. The routine not only
   adds the page to the swap cache (where it remains until its data are written to disk), but also
   reserves a slot in one of the swap areas.


2. When a page shared by several processes (this can be determined by referring to the usage
   counter in the swap area) is read in from the swap area, the page is retained in both the
   swap area and the swap cache until it is either swapped out again or until it is swapped
   in by all the processes that share it. The kernel implements this behavior by means of the
   add_to_swap_cachefunction, which adds a page to the swap cachewithoutperforming oper-
   ations on the swap areas themselves.

Reserving Page Slots

Before dealing with the implementation details of these two functions, we should examine how page

slots are reserved in swap areas. The kernel delegates this task toget_swap_page, which — when called

without parameters — returns the number of the page slot to be used next.

The function must first ensure that the system does, in fact, have swap areas — if so, the global variable

nr_swap_pageshas a value greater than 0.

swap_list.nextalways yields the number of the swap area currently in use (if there is only one swap

area, it obviously always returns the same number). Logically, the kernel begins the search for a free page

slot in this area.scan_swap_mapscans the page bitmap, making use of swap clustering — a technique

examined below.

Ifnofree slot is found in the current swap area, the kernel checks the alternate slots. To do this, it runs

through the list of all swap areas until it finds a free slot. Naturally, searching is performed in line with

the priorities defined for each individual area by means of thenextelement of eachswap_info[]entry.