Virtual Memory

Agenda

• Address Spaces
• Segmentation
• Paging
• Swapping
• User space memory management
• Case Study: Linux + x86 VM

Address Spaces

Binaries have addresses hardcoded into them

When a program is executed:

• Text & Data Sections are initialized with contents of binary
• BSS section zero initialized
• Starting addresses for initially empty stack & heap are also established

If different binaries have conflicting addresses, how is this possible? They should not be able to simultaneously access the same physical memory…

• Not all text sections can begin at 0x400000
• Not all data sections can begin at 0x600000
• Not all heaps can begin at 0x18f0000
• Not all stacks can begin at 0x7fff00000000

Uniform process address spaces are an illusion created by the kernel

• To simplify program loading and execution (among other reasons)

Hardware Address Translation

• All process memory requests on CPU are for virtual addresses
• Hardware translates these to physical addresses
• Kernel sets up mapping for virtual -> physical
• Translation hardware is the memory management unit (MMU)

Primary goals

• “Transparency” (not really)
  • Processes aren’t involved in (or aware of) the translation process
• Efficiency
  • Time and Space
• Protection
  • Process can’t access data out of their address space
  • Isolated from each other and the kernel

A simple implementation: Relocation

Kernel maintains the address of each process in PCB

Load base address into into MMU on each context switch Relocation = register access + addition

What about protection?

• Incorporate a limit register to enforce memory protection
• Assert address is valid