Linux Kernel Architecture

Chapter 4: Virtual Process Memory

4.11.3 Copy on Write

Copy on write is handled indo_wp_page, whose code flow diagram is shown in Figure 4-20.

do_wp_page

Insert page into page tables

vm_normal_page

anon_vma_prepare

alloc_page_vma

cow_user_vma

page_remove_rmap

lru_cache_add_active

page_add_new_anon_rmap

Figure 4-20: Code flow diagram for

do_wp_page.

Let’s examine a slightly simplified version in which I have omitted potential interference with the swap

cache as well as some corner cases, since this would complicate the situation without revealing anything

insightful about the mechanism itself.

The kernel first invokesvm_normal_pageto find thestruct pageinstance of the page by reference to

the page table entry — essentially, this function builds onpte_pfnandpfn_to_page, which must be

defined on all architectures. The former finds the page number for an associated page table entry, and

the latter determines thepageinstance associated with the page number. This is possible because the

COW mechanism is invoked only for pages that actually reside in memory (otherwise, they are first

automatically loaded by one of the other page fault mechanisms).

After obtaining a reference on the page withpage_cache_get,anon_vma_preparethen prepares the

data structures of the reverse mapping mechanism to accept a new anonymous area. Since the fault

originates from a page filled with useful data that must be copied to a new page, the kernel invokes

alloc_page_vmato allocate a fresh page.cow_user_pagethen copies the data of the faulted page into the

new page to which the process may subsequently write.

The reverse mapping to the original read-only page is then removed usingpage_remove_rmap. The new

page is added to the page tables, at which point the CPU caches must also be updated.

The final actions involve placing the newly allocated pages on the active list of the LRU cache using

lru_cache_add_activeand inserting them in the reverse mapping data structures by means of

page_add_anon_rmap. Thereafter, the userspace process can write to the page to its heart’s content.