Linux Kernel Architecture

Chapter 18: Page Reclaim and Swapping

Starting at the current position, the kernel checks whether all subsequent slots —SWAPFILE_CLUSTERin

number — are free; this check is performed by theforloop within theifquery. If the kernel finds an

allocated entry whoseswap_mapentry is greater than 0, the search for a free cluster is resumed at the

next slot position. This is repeated until a position is finally reached at which there is insufficient space to

create a cluster.

If the kernel is successful in its search for a new cluster, it jumps to thelowestlabel, as above.

Allocating SwapSpace

After the policy routine has decided that a particular page needs to be swapped out,add_to_swapfrom

mm/filemap.ccomes into play. This function accepts astruct pageinstance as parameter and forwards

the swap-out request to the technical part of swapping implementation.

As the code flow diagram in Figure 18-8 shows, this is not a very difficult task and consists basically of

three steps. After theget_swap_pageroutine mentioned above has reserved a page slot in one of the

swap areas, all that needs to be done is move the page into the swap area. This is the responsibility of

the__add_to_swap_cachefunction, which is very similar to the standardadd_to_page_cachefunction

described in Chapter 16. The primary difference is that thePG_swapcacheflag is set and the swap iden-

tifierswp_entry_tis stored in theprivateelement of the page — it will be required to construct an

architecture-dependent page table entry when the page is actually swapped out. Additionally, the global

variabletotal_swapcache_pagesis incremented to update the statistics. Nevertheless, as we would

expect ofadd_to_page_cache, the page is inserted in the radix tree set up byswapper_space.

add_to_swap

get_swap_page

SetPageUptodate

SetPageDirty

_ _add_to_swap_page_cache

Figure 18-8: Code flow diagram for

add_to_swap.

Finally,SetPageUpdateandSetPageDirtymodify the page flags appropriately. Dirtying the page is

essential because the contents of the page are not yet contained in the swap area. Recall from Chapter 17

that pages in the page cache are synchronized with their underlying block device when they are dirty.

For a swap page, the underlying block deviceisthe swap space, and synchronization is therefore (nearly)

equivalent to swapping the page out! What remains to be done after the page has been written to a swap

slot is updating the page table to reflect this.

But, otherwise, that’s it. Nothing more is required of the policy routines when pages are swapped out.

The rest of the work — particularly the transferring of data from memory to the swap area — is per-

formed by the address-space-specific operations associated withswapper_space. The implementation of

the routines is discussed below — as far as the policy is concerned, it is enough to know that the kernel

actually writes the data to the swap area and thus releases a page afteradd_to_swaphas been invoked.

More details follow in the discussion of theshrink_page_listfunction.