Virtual Memory Overview

Virtual Memory Overview

• Virtual memory is a system mechanism that separates program address space from physical memory layout.
• Virtual addresses represent locations within a program's address space.
• Physical addresses represent locations in actual hardware RAM.
• Virtual memory allows these two address spaces to be different.

Process Layout

• Each process has unmapped memory near NULL.
• Processes can potentially access the entire address space.
• Each process is denied access to memory used by other processes.
• Virtual memory resolves the apparent contradiction.
• Every address in a process's address space is a virtual address.

Physical Layout

• Hardware RAM layout varies significantly between machines and platforms.
• Certain memory locations may be restricted.
• Devices may appear within the memory address space.
• Different RAM amounts may be present.
• Historically, programs needed to be aware of these restrictions.
• Today, virtual memory hides these details from programs.
• The kernel must still be aware of the physical layout.

Memory Management Unit (MMU)

• The MMU translates addresses.
• It uses a per-process mapping structure to change virtual addresses to physical addresses.
• The MMU is hardware between the CPU and memory bus.
• Address translation needs to be very fast, but occasional performance penalties are possible.
• Kernel and hardware need to tightly integrate for translation management.

Paging

• Multiple virtual memory models are possible.
• x86-64 architectures support several models.
• Linux on x86-64 uses paging.
• Paging divides memory into fixed-sized pages.
• Multiple page sizes might exist in a system.
• Page sizes are typically powers of two.
• x86-64 small pages are 4 KB.

Page Metadata

• Page table metadata describes page features like readability/writability, code execution permission, and presence in memory.
• A page fault occurs when a memory access violates this metadata (e.g., write to a non-writable page).
• The MMU notifies the processor.
• The processor jumps to a specific kernel routine for processing.
• The kernel tries fixing the problem or informing the offending process.

Page Backing

• Virtual pages can use files, physical pages, or both for backing.
• A backed page's contents depend on its backing.
• Backed pages don't always need to be in memory.
• Clean pages are identical to their backing; dirty pages are different.
• Clean pages can be recreated from their backing at any time.

Demand Paging

• A virtual page might have backing but not be present in memory.
• Such pages are marked "not present" in page tables.
• Attempts to access these pages notify the kernel. (This is a page fault.)
• The kernel pages in the page by: finding an unused physical page, locating the virtual page's backing, reading the backing data into the physical page.

Demand Paging Benefits

• Physical memory allocation is quicker, simply by updating page tables.
• Executable file loading is faster; pages are marked present but backed by files.
• Accessing these pages causes the file to be read into memory.
• Unused pages are never loaded.

Program Break

• Calling brk() or sbrk() changes a process's memory map.
• Pages adjacent to the old program break are marked as "Not present, Readable and writable".
• This only affects page table metadata; the pages themselves aren't allocated.
• On a new page request, the MMU notifies the processor.
• The kernel locates and clears an unused page.
• The kernel adds the new page to the process's page table.

Moving the Program Break (Page Fault)

• The process's access to new pages is handled normally after the program break is moved.
• The MMU notifies the kernel when the process tries to access the new page.
• The kernel allocates the requested page.

C Stack

• The kernel manages the program stack.
• The stack grows as needed (not explicitly sized by the program).
• The kernel configures the MMU to manage the program stack.
• Similar to newly allocated memory, the kernel determines the stack size.
• Stack pages are marked "not present" but "readable and writable"

Page Eviction

• When memory is low, the system can evict a page.
• Clean pages are removed from the map.
• Dirty pages are written to their backing before removal.
• Swap space can back un-backed dirty pages.
• This allows the system to simulate more memory than physically available.

Address Spaces Summary

• Virtual memory uses a memory management unit.
• The CPU operates in a virtual address space separate from physical memory.
• The MMU translates virtual addresses to physical addresses.
• Paging divides both virtual and physical address spaces into pages.
• Pages can be individually mapped between virtual and physical addresses.
• Page tables allow the MMU to translate addresses.
• Page faults handle situations where mapped pages may not be currently allocated in physical memory.