The Linux memory manager implements demand paging with a copy-on-write strategy relying on the 386's paging support. A process acquires its page tables from its parent (during a fork()) with the entries marked as read-only or swapped. Then, if the process tries to write to that memory space, and the page is a copy-on-write page, it is copied, and the page is marked read-write. An exec() results in the reading in of a page or so from the executable. The process then faults in any other pages it needs.
Each process has a page directory which means it can access 1 KB of page tables pointing to 1 MB of 4 KB pages which is 4 GB of memory. A process' page directory is initialized during a fork by copy_page_tables(). The idle process has its page directory initialized during the initialization sequence.
Each user process has a local descriptor table that contains a code segment and data-stack segment. These user segments extend from 0 to 3 GB (0xc0000000). In user space linear addresses and logical addresses are identical.
The kernel code and data segments are priveleged segments defined in the global descriptor table and extend from 3 GB to 4 GB. The swapper page directory (swapper_page_dir is set up so that logical addresses and physical addresses are identical in kernel space.
The space above 3 GB appears in a process' page directory as pointers to kernel page tables. This space is invisible to the process in user mode but the mapping becomes relevant when privileged mode is entered, for example, to handle a system call.
Supervisor mode is entered within the context of the current process so address translation occurs with respect to the process' page directory but using kernel segments. This is identically the mapping produced by using the swapper_pg_dir and kernel segments as both page directories use the same page tables in this space. Only task (the idle task [This should be documented earlier in this guide...]) uses the swapper_pg_dir directly.
The user stack sits at the top of the user data segment and grows down. The kernel stack is not a pretty data structure or segment that I can point to with a ``yon lies the kernel stack.'' A kernel_stack_frame (a page) is associated with each newly created process and is used whenever the kernel operates within the context of that process. Bad things would happen if the kernel stack were to grow below its current stack frame. [Where is the kernel stack put? I know that there is one for every process, but where is it stored when it's not being used?]
User pages can be stolen or swapped. A user page is one that is mapped below 3 GB in a user page table. This region does not contain page directories or page tables. Only dirty pages are swapped.
Minor alterations are needed in some places (tests for process memory limits comes to mind) to provide support for programmer defined segments.