Xen-BartMiller

Report
Bart Miller
Outline
 Definition and goals
 Paravirtualization
 System Architecture
 The Virtual Machine Interface
 Memory Management
 CPU
 Device I/O

Network, Disk
 Xen Timeline
Definition and Goals
 An x86 Virtual Machine Monitor (aka Hypervisor)
 Developed in 2003
 Approximately 60k lines of code
 Goals:
 100 VMs per system
 Support full multi-application OSes
 No modifications to guest applications
 Negligible sacrifice in performance
 Fully isolate guests
Paravirtualization
 Full virtualization on x86 (ca. 2003) is complex and
not efficient
 Create a unique interface to the hardware
 Let the Guest OS access the hardware directly when
appropriate
 Prevent the Guest OS from accessing functionality
which could affect other guests or the VMM.
 Must modify the Guest OS
 For Linux, 2995 lines (1.36% of code base)
 For Windows XP, 4620 lines (0.04% of code base)
System Architecture
Memory Management
 Problem
 x86 has a hardware managed TLB
 Assumes single OS, does not support tagging nor
managing in software
 Context switch requires TLB flush
 Solution
 Guest OSes manage hardware page table

Direct read access; updates batched and validated by Xen
 Xen resides in 64Mbyte section at the top of every
address space
CPU
 Protection
 x86 has 4 privilege levels, known as Rings.
 Ring O is highest and Ring 3 lowest privilege
 For Xen, the VMM executes in Ring O, the Guest OS
executes in Ring 1, and the user programs execute in
Ring 3.
 Exceptions
 The Guest OS registers a table of exception handlers
with Xen
 All are unmodified except the Page Fault handler, since
it normally requires access to a privileged register (CR2)
CPU (2)
 System Calls
 The Guest OS can register a “fast” exception handler

Executes without indirection
 Xen verifies that the handler does not specify execution
in Ring O
 Interrupts
 Interrupts are replaced by a lightweight event system
 Asynchronous, relies on Ring buffer
 Time
 Guest OSes have access to “real” and “virtual” timers
Device I/O
 Network
 VIF, VFR
 Transmit and receive Ring I/O buffers
 Domain O manages and enforces the firewall rules
 Transmit:


A guest enqueues a request to the transmit ring
Xen validates the request against the firewall rules and forwards to
the device
 Receive:
 A guest enqueues a receive request to the receive ring
 Xen determines the appropriate recipient
 The packet buffer is exchanged for a sacrificial page frame on the
receiver’s ring
Device I/O (2)
 Disk access
 Only Domain O can directly access physical disks
 All DomUs communicate through Virtual Block Devices
(VBD)
 Channels are comprised of Ring buffers
 Requests can be reordered by the Guest OS and Xen

Unless the Guest OS issues a reorder barrier
Ring Structure
Ring Structure (2)
Xen Timeline
 2003: Initial release of Xen
 2005 was a significant year for Virtualization
 Intel introduces VT-x, quickly utilized by Xen
 Narrows performance gap between HVM and PVM
 2006: Amazon opens up public beta of EC2
 2007: Live migration for HVM guests
 2008: PCI pass-through (VT-d) and ACPI S3 support
 2011: Xen support for DomO and DomU is added to
the Linux kernel
Questions?

similar documents