Lecture: memory virtualization

administrivia

• proposal feedback
• don’t make your solutions public
• exercise: VM-exit handling
• QEMU’s BIOS

overview

• read chapter 4 of the textbook
• review page tables
  • %cr3, page walking, TLB
  • x86 examples
    • two levels
    • four levels
    • five levels
• goal
  • vmm needs to run multiple guests
  • a guest shouldn’t be able to trash the memory of vmm/other guests
  • addresses
    • guest virtual/physical address
    • host virtual/physical address
  • example
    • guests 1 & 2, each taking 1GB memory
    • one possible layout
      • vmm: host physical 0-1G
      • guest 1: guest physical 0-1G to host physical 1-2G
      • guest 2: guest physical 0-1G to host physical 2-3G
• shadow page tables
  • recall from last lecture: performance issues w/o memory virtualization
  • guest virtual → host physical
• two-dimensional page tables
  • guest physical → host physical
  • also known as
    • second level address translation (SLAT)
    • extended page tables (EPT)
    • nested page tables (NPT)

shadow paging

• “A Comparison of Software and Hardware Techniques for x86 Virtualization”, ASPLOS 2006
• strawman
  • allocate smaller memory regions for guests
  • intercept %cr3 write
  • scan & construct a separate, “shadow” page table
    • what to ensure in the shadow page table?
    • what if the guest constructs a page table that writes to the vmm or other guests?
  • problems
    • performance
    • what if the guest modifies its own page table - is this okay
    • what if the guest modifies the shadow page table
• ideas
  • lazy construction
    • start from empty shadow page table
    • add entries during page faults
  • cache shadow page tables
  • make guest page-table pages read-only
    • trap on writes page-table pages
    • emulate
  • TLB flush
  • performance trade-offs

two-dimensional paging

• use EPT as an example
  • vmm sets EPT root in VMCS
  • vmm construct EPT
• two-dimensional page walking
  • “Accelerating Two-Dimensional Page Walks for Virtualized Systems”, ASPLOS 2008, figure 1
  • also figure 5.1 of the textbook
  • consider n-level guest page table and m-level EPT
  • quadratic lookup: (m + 1) * n + m
  • 24 for n = 4 and m = 4
• performance trade-offs
  • use 2M/1G pages
  • when are shadow paga tables faster?