Cpu Virtualization

Summary

• CPU Virtualization
  • Delivering Events to Parent Guest OS
• Device Virtualization
  • Control Transfer
  • Data Transfer
  • Control and Data Transfer in Action
  • Disk IO Virtualization
• Xen and Guests

CPU Virtualization

Hypervisor has to give illusion of ownership of cpu for each VM

• This includes the “ready queue”

The hypervisor schedules execution of VMs, just like guest schedules execution of its processes

• proportional share or fair share

Delivering Events to Parent Guest OS

Interrupts from currently executing process (owned by guest OS) need to get back to that guest

Program discontinuities (PD) need to be passed up from hypervisor to guest OS in full and para virtualization:

• Syscalls
• Page faults
• Throw exception
• External Interrupt

Some PDs needs elevated privileges, but guest OS has normal user-level privileges (especially in fully virtualized)

• Sends a trap back to hypervisor

Device Virtualization

• Full virtualization
  • “trap and emulate”
• Para virtualization
  • more opportunity for innovation
  • guest OS can see devices directly
    • Shared buffers can optimize communication

Control Transfer

• Full virtualization
  • implicit (traps) guest -> hypervisor
  • software interrupts (events) hypervisor -> guest
• Para virtualization
  • explicit (hypercalls) guest -> hypervisor
  • software interrupts (events) hypervisor -> guest
    • guest has control via hypercalls on when event notifications need to be delivered

Data Transfer

• Full
  • implicit
• Para
  • explicit
    • opportunity to innovate

Xen provides async IO rings

• Shared between guest and Xen for communication
  • Ring buffer

Control and Data Transfer in Action

Network virtualization

2 IO Rings

• One for transmit
• One for receive

If you want to transmit data, you place file descriptors on transmit IO ring

• via hypercalls

Data pointed to by file descriptors are NOT copied into Xen, instead they’re guest OS buffers

• No copying
• page pinned

Receiving data works the same way

• Before receiving packet, Xen allocates guest OS buffer and puts data directly there

Disk IO Virtualization

No copying into Xen

Requests from competing domains may be reordered

• Reorder barrier supported by Xen for guest OS semantics
  • useful for things like Write-Ahead Logging

Xen and Guests

Different between full vs para is a focus on protection and flexibility