xen/docs/interface.tex

\documentclass[11pt,twoside,final,openright]{xenstyle}
\usepackage{a4,graphicx,setspace}
\setstretch{1.15}
\input{style.tex}

\begin{document}

% TITLE PAGE
\pagestyle{empty}
\begin{center}
\vspace*{\fill}
\includegraphics{eps/xenlogo.eps}
\vfill
\vfill
\vfill
\begin{tabular}{l}
{\Huge \bf Interface manual} \\[4mm]
{\huge Xen v1.3 for x86} \\[80mm]

{\Large Xen is Copyright (c) 2004, The Xen Team} \\[3mm]
{\Large University of Cambridge, UK} \\[20mm]
{\large Last updated on 11th March, 2004}
\end{tabular}
\vfill
\end{center}
\cleardoublepage

% TABLE OF CONTENTS
\pagestyle{plain}
\pagenumbering{roman}
{ \parskip 0pt plus 1pt
  \tableofcontents }
\cleardoublepage

% PREPARE FOR MAIN TEXT
\pagenumbering{arabic}
\raggedbottom
\widowpenalty=10000
\clubpenalty=10000
\parindent=0pt
\renewcommand{\topfraction}{.8}
\renewcommand{\bottomfraction}{.8}
\renewcommand{\textfraction}{.2}
\renewcommand{\floatpagefraction}{.8}
\setstretch{1.15}

\chapter{Introduction}
Xen allows the hardware resouces of a machine to be virtualized and
dynamically partitioned such as to allow multiple different 'guest'
operating system images to be run simultaneously.

Virtualizing the machine in this manner provides flexibility allowing
different users to choose their preferred operating system (Windows,
Linux, FreeBSD, or a custom operating system). Furthermore, Xen provides
secure partitioning between these 'domains', and enables better resource
accounting and QoS isolation than can be achieved with a conventional
operating system.

The hypervisor runs directly on server hardware and dynamically partitions
it between a number of {\it domains}, each of which hosts an instance
of a {\it guest operating system}. The hypervisor provides just enough
abstraction of the machine to allow effective isolation and resource 
management between these domains.

Xen essentially takes a virtual machine approach as pioneered by IBM VM/370.
However, unlike VM/370 or more recent efforts such as VMWare and Virtual PC,
Xen doesn not attempt to completely virtualize the underlying hardware. Instead
parts of the hosted guest operating systems to work with the hypervisor; the
operating system is effectively ported to a new target architecture, typically
requiring changes in just the machine-dependent code. The user-level API is
unchanged, thus existing binaries and operating system distributions can work
unmodified.

In addition to exporting virtualized instances of CPU, memory, network and
block devicees, Xen exposes a control interface to set how these resources
are shared between the running domains. The control interface is privileged
and may only be accessed by one particular virtual machine: {\it domain0}.
This domain is a required part of any Xen-base server and runs the application
software that manages the control-plane aspects of the platform. Running the
control software in {\it domain0}, distinct from the hypervisor itself, allows
the Xen framework to separate the notions of {\it mechanism} and {\it policy}
within the system.


\chapter{CPU state}

All privileged state must be handled by Xen. The guest OS has no direct access
to CR3 and is not permitted to update privileged bits in EFLAGS.

\chapter{Exceptions}
The IDT is virtualised by submitting a virtual 'trap
table' to Xen. Most trap handlers are identical to native x86
handlers. The page-fault handler is a noteable exception.

\chapter{Interrupts and events}
Interrupts are virtualized by mapping them to events, which are delivered 
asynchronously to the target domain. A guest OS can map these events onto
its standard interrupt dispatch mechanisms, such as a simple vectoring 
scheme. Each physical interrupt source controlled by the hypervisor, including
network devices, disks, or the timer subsystem, is responsible for identifying
the target for an incoming interrupt and sending an event to that domain.

This demultiplexing mechanism also provides a device-specific mechanism for 
event coalescing or hold-off. For example, a guest OS may request to only 
actually receive an event after {\it n} packets are queued ready for delivery
to it, {\it t} nanoseconds after the first packet arrived (which ever is true
first). This allows latency and throughput requirements to be addressed on a
domain-specific basis.

\chapter{Time}
Guest operating systems need to be aware of the passage of real time and their
own ``virtual time'', i.e. the time they have been executing. Furthermore, a
notion of time is required in the hypervisor itself for scheduling and the
activities that relate to it. To this end the hypervisor provides for notions
of time: cycle counter time, system time, wall clock time, domain virtual 
time.


\section{Cycle counter time}
This provides the finest-grained, free-running time reference, with the
approximate frequency being publicly accessible. The cycle counter time is
used to accurately extrapolate the other time references. On SMP machines
it is currently assumed that the cycle counter time is synchronised between
CPUs. The current x86-based implementation achieves this within inter-CPU
communication latencies.

\section{System time}
This is a 64-bit value containing the nanoseconds elapsed since boot
time. Unlike cycle counter time, system time accurately reflects the
passage of real time, i.e.  it is adjusted several times a second for timer
drift. This is done by running an NTP client in {\it domain0} on behalf of
the machine, feeding updates to the hypervisor. Intermediate values can be
extrapolated using the cycle counter.

\section{Wall clock time}
This is the actual ``time of day'' Unix style struct timeval (i.e. seconds and
microseconds since 1 January 1970, adjusted by leap seconds etc.). Again, an 
NTP client hosted by {\it domain0} can help maintain this value. To guest 
operating systems this value will be reported instead of the hardware RTC
clock value and they can use the system time and cycle counter times to start
and remain perfectly in time.


\section{Domain virtual time}
This progresses at the same pace as cycle counter time, but only while a
domain is executing. It stops while a domain is de-scheduled. Therefore the
share of the CPU that a domain receives is indicated by the rate at which
its domain virtual time increases, relative to the rate at which cycle
counter time does so.

\section{Time interface}
Xen exports some timestamps to guest operating systems through their shared
info page. Timestamps are provided for system time and wall-clock time. Xen
also provides the cycle counter values at the time of the last update
allowing guests to calculate the current values. The cpu frequency and a
scaling factor are provided for guests to convert cycle counter values to
real time. Since all time stamps need to be updated and read
\emph{atomically} two version numbers are also stored in the shared info
page.

Xen will ensure that the time stamps are updated frequently enough to avoid
an overflow of the cycle counter values. Guest can check if its notion of
time is up-to-date by comparing the version numbers.

\section{Timer events}

Xen maintains a periodic timer (currently with a 10ms period) which sends a
timer event to the currently executing domain. This allows Guest OSes to
keep track of the passing of time when executing. The scheduler also
arranges for a newly activated domain to receive a timer event when
scheduled so that the Guest OS can adjust to the passage of time while it
has been inactive.

In addition, Xen exports a hypercall interface to each domain which allows
them to request a timer event send to them at the specified system
time. Guest OSes may use this timer to implemented timeout values when they
block.

\chapter{Memory}

The hypervisor is responsible for providing memory to each of the domains running 
over it. However, the Xen hypervisor's duty is restricted to managing physical
memory and to policing page table updates. All other memory management functions
are handly externally. Start-of-day issues such as building initial page tables
for a domain, loading its kernel image and so on are done by the {\it domain builder}
running in user-space with {\it domain0}. Paging to disk and swapping is handled
by the guest operating systems themselves, if they need it.

On a Xen-based system, the hypervisor itself runs in {\it ring 0}. It has full
access to the physical memory available in the system and is responsible for 
allocating portions of it to the domains. Guest operating systems run in and use
{\it rings 1}, {\it 2} and {\it 3} as they see fit, aside from the fact that
segmentation is used to prevent the guest OS from accessing a portion of the 
linear address space that is reserved for use by the hypervisor. This approach
allows transitions between the guest OS and hypervisor without flushing the TLB.
We expect most guest operating systems will use ring 1 for their own operation
and place applications (if they support such a notion) in ring 3.

\section{Physical Memory Allocation}
The hypervisor reserves a small fixed portion of physical memory at system boot
time. This special memory region is located at the beginning of physical memory
and is mapped at the very top of every virtual address space. 

Any physical memory that is not used directly by the hypervisor is divided into
pages and is available for allocation to domains. The hypervisor tracks which
pages are free and which pages have been allocated to each domain. When a new
domain is initialized, the hypervisor allocates it pages drawn from the free 
list. The amount of memory required by the domain is passed to the hypervisor
as one of the parameters for new domain initialization by the domain builder.

Domains can never be allocated further memory beyond that which was requested
for them on initialization. However, a domain can return pages to the hypervisor
if it discovers that its memory requirements have diminished.

% put reasons for why pages might be returned here.
\section{Page Table Updates}
In addition to managing physical memory allocation, the hypervisor is also in
charge of performing page table updates on behalf of the domains. This is 
neccessary to prevent domains from adding arbitrary mappings to their page
tables or introducing mappings to other's page tables.

\section{Segment Descriptor Tables}

On boot a guest is supplied with a default GDT, which is {\em not}
taken from its own memory allocation. If the guest wishes to use other
than the default `flat' ring-1 and ring-3 segments that this default
table provides, it must register a custom GDT and/or LDT with Xen,
allocated from its own memory.

int {\bf set\_gdt}(unsigned long *{\em frame\_list}, int {\em entries})

{\em frame\_list}: An array of up to 16 page frames within which the GDT
resides. Any frame registered as a GDT frame may only be mapped
read-only within the guest's address space (e.g., no writeable
mappings, no use as a page-table page, and so on).

{\em entries}: The number of descriptor-entry slots in the GDT. Note that
the table must be large enough to contain Xen's reserved entries; thus
we must have '{\em entries $>$ LAST\_RESERVED\_GDT\_ENTRY}'. Note also that,
after registering the GDT, slots {\em FIRST\_} through
{\em LAST\_RESERVED\_GDT\_ENTRY} are no longer usable by the guest and may be
overwritten by Xen.

\section{Pseudo-Physical Memory}
The usual problem of external fragmentation means that a domain is unlikely to
receive a contiguous stretch of physical memory. However, most guest operating
systems do not have built-in support for operating in a fragmented physical
address space e.g. Linux has to have a one-to-one mapping for it physical
memory. There a notion of {\it pseudo physical memory} is introdouced. 
Once a domain is allocated a number of pages, at its start of the day, one of
the first things it needs to do is build its own {\it real physical} to 
{\it pseudo physical} mapping. From that moment onwards {\it pseudo physical}
address are used instead of discontiguous {\it real physical} addresses. Thus,
the rest of the guest OS code has an impression of operating in a contiguous
address space. Guest OS page tables contain real physical addresses. Mapping
{\it pseudo physical} to {\it real physical} addresses is need on page
table updates and also on remapping memory regions with the guest OS.



\chapter{Network I/O}
Since the hypervisor must multiplex network resources, its network subsystem
may be viewed as a virtual network switching element with each domain having
one or more virtual network interfaces to this network.

The hypervisor acts conceptually as an IP router, forwarding each domain's
traffic according to a set of rules.

\section{Hypervisor Packet Handling}
The hypervisor is responsible primarily for {\it data-path} operations.
In terms of networking this means packet transmission and reception.

On the transmission side, the hypervisor needs to perform two key actions:
\begin{itemize}
\item {\tt Validation:} A domain is only allowed to emit packets matching a certain
specification; for example, ones in which the source IP address matches
one assigned to the virtual interface over which it is sent. The hypervisor
is responsible for ensuring any such requirements are met, either by checking
or by stamping outgoing packets with prescribed values for certain fields.

\item {\tt Scheduling:} Since a number of domains can share a single ``real'' network 
interface, the hypervisor must mediate access when several domains each 
have packets queued for transmission. Of course, this general scheduling
function subsumes basic shaping or rate-limiting schemes.

\item {\tt Logging and Accounting:} The hypervisor can be configured with classifier 
rules that control how packets are accounted or logged. For example, 
{\it domain0} could request that it receives a log message or copy of the
packet whenever another domain attempts to send a TCP packet containg a 
SYN.
\end{itemize}
On the recive side, the hypervisor's role is relatively straightforward:
once a packet is received, it just needs to determine the virtual interface(s)
to which it must be delivered and deliver it via page-flipping. 


\section{Data Transfer}

Each virtual interface uses two ``descriptor rings'', one for transmit,
the other for receive. Each descriptor identifies a block of contiguous
physical memory allocated to the domain. There are four cases:

\begin{itemize}

\item The transmit ring carries packets to transmit from the domain to the
hypervisor.

\item The return path of the transmit ring carries ``empty'' descriptors
indicating that the contents have been transmitted and the memory can be
re-used.

\item The receive ring carries empty descriptors from the domain to the 
hypervisor; these provide storage space for that domain's received packets.

\item The return path of the receive ring carries packets that have been
received.
\end{itemize}

Real physical addresses are used throughout, with the domain performing 
translation from pseudo-physical addresses if that is necessary.

If a domain does not keep its receive ring stocked with empty buffers then 
packets destined to it may be dropped. This provides some defense against 
receiver-livelock problems because an overload domain will cease to receive
further data. Similarly, on the transmit path, it provides the application
with feedback on the rate at which packets are able to leave the system.

Synchronization between the hypervisor and the domain is achieved using 
counters held in shared memory that is accessible to both. Each ring has
associated producer and consumer indices indicating the area in the ring
that holds descriptors that contain data. After receiving {\it n} packets
or {\t nanoseconds} after receiving the first packet, the hypervisor sends
an event to the domain. 

\chapter{Block I/O}

\section{Virtual Block Devices (VBDs)}

All guest OS disk access goes through the VBD interface. The VBD interface
provides the administrator with the ability to selectively grant domains 
access to portions of block storage devices visible to the system.

A VBD can also be comprised of a set of extents from multiple storage devices.
This provides the same functionality as a concatenated disk driver.

\section{Virtual Disks (VDs)}

VDs are an abstraction built on top of the VBD interface. One can reserve disk
space for use by the VD layer. This space is then managed as a pool of free extents.
The VD tools can automatically allocate collections of extents from this pool to
create ``virtual disks'' on demand. 

\subsection{Virtual Disk Management}
The VD management code consists of a set of python libraries. It can therefore
be accessed by custom scripts as well as the convenience scripts provided. The
VD database is a SQLite database in /var/db/xen\_vdisk.sqlite.

The VD scripts and general VD usage are documented in the VBD-HOWTO.txt.

\subsection{Data Transfer}
Domains which have been granted access to a logical block device are permitted
to read and write it directly through the hypervisor, rather than requiring
{\it domain0} to mediate every data access. 

In overview, the same style of descriptor-ring that is used for network
packets is used here. Each domain has one ring that carries operation requests to the 
hypervisor and carries the results back again. 

Rather than copying data in and out of the hypervisor, we use page pinning to
enable DMA transfers directly between the physical device and the domain's 
buffers. Disk read operations are straightforward; the hypervisor just needs
to know which pages have pending DMA transfers, and prevent the guest OS from
giving the page back to the hypervisor, or to use them for storing page tables.

%block API here 

\chapter{Privileged operations}
{\it Domain0} is responsible for building all other domains on the server
and providing control interfaces for managing scheduling, networking, and
blocks.


\chapter{Debugging}

Xen provides tools for debugging both Xen and guest OSes.  Currently, the
Pervasive Debugger provides a GDB stub, which provides facilities for symbolic
debugging of Xen itself and of OS kernels running on top of Xen.  The Trace
Buffer provides a lightweight means to log data about Xen's internal state and
behaviour at runtime, for later analysis.

\section{Pervasive Debugger}

Information on using the pervasive debugger is available in pdb.txt.


\section{Trace Buffer}

The trace buffer provides a means to observe Xen's operation from domain 0.
Trace events, inserted at key points in Xen's code, record data that can be
read by the {\tt xentrace} tool.  Recording these events has a low overhead
and hence the trace buffer may be useful for debugging timing-sensitive
behaviours.

\subsection{Internal API}

To use the trace buffer functionality from within Xen, you must {\tt \#include
<xeno/trace.h>}, which contains definitions related to the trace buffer.  Trace
events are inserted into the buffer using the {\tt TRACE\_xD} ({\tt x} = 0, 1,
2, 3, 4 or 5) macros.  These all take an event number, plus {\tt x} additional
(32-bit) data as their arguments.  For trace buffer-enabled builds of Xen these
will insert the event ID and data into the trace buffer, along with the current
value of the CPU cycle-counter.  For builds without the trace buffer enabled,
the macros expand to no-ops and thus can be left in place without incurring
overheads.

\subsection{Enabling tracing}

By default, the trace buffer is enabled only in debug builds (i.e. {\tt NDEBUG}
is not defined).  It can be enabled separately by defining {\tt TRACE\_BUFFER},
either in {\tt <xeno/config.h>} or on the gcc command line.

\subsection{Dumping trace data}

When running a trace buffer build of Xen, trace data are written continuously
into the buffer data areas, with newer data overwriting older data.  This data
can be captured using the {\tt xentrace} program in Domain 0.

The {\tt xentrace} tool uses {\tt /dev/mem} in domain 0 to map the trace
buffers into its address space.  It then periodically polls all the buffers for
new data, dumping out any new records from each buffer in turn.  As a result,
for machines with multiple (logical) CPUs, the trace buffer output will not be
in overall chronological order.

The output from {\tt xentrace} can be post-processed using {\tt
xentrace\_cpusplit} (used to split trace data out into per-cpu log files) and
{\tt xentrace\_format} (used to pretty-print trace data).

For more information, see the manual pages for {\tt xentrace},
{\tt xentrace\_format} and {\tt xentrace\_cpusplit}.


\chapter{Hypervisor calls}

\section{ set\_trap\_table(trap\_info\_t *table)} 

Install trap handler table.

\section{ mmu\_update(mmu\_update\_t *req, int count)} 
Update the page table for the domain. Updates can be batched.
The update types are: 

{\it MMU\_NORMAL\_PT\_UPDATE}:

{\it MMU\_UNCHECKED\_PT\_UPDATE}:

{\it MMU\_MACHPHYS\_UPDATE}:

{\it MMU\_EXTENDED\_COMMAND}:

\section{ console\_write(const char *str, int count)}
Output buffer str to the console.

\section{ set\_gdt(unsigned long *frame\_list, int entries)} 
Set the global descriptor table - virtualization for lgdt.

\section{ stack\_switch(unsigned long ss, unsigned long esp)} 
Request context switch from hypervisor.

\section{ set\_callbacks(unsigned long event\_selector, unsigned long event\_address,
                        unsigned long failsafe\_selector, unsigned long failsafe\_address) } 
 Register OS event processing routine. In Linux both the event\_selector and 
failsafe\_selector are the kernel's CS. The value event\_address specifies the address for
an interrupt handler dispatch routine and failsafe\_address specifies a handler for 
application faults.

\section{ net\_io\_op(netop\_t *op)}  
Notify hypervisor of updates to transmit and/or receive descriptor rings.

\section{ fpu\_taskswitch(void)} 
Notify hypervisor that fpu registers needed to be save on context switch.

\section{ sched\_op(unsigned long op)} 
Request scheduling operation from hypervisor. The options are: {\it yield},
{\it block}, {\it stop}, and {\it exit}. {\it yield} keeps the calling
domain run-able but may cause a reschedule if other domains are
run-able. {\it block} removes the calling domain from the run queue and the
domains sleeps until an event is delivered to it. {\it stop} and {\it exit}
should be self-explanatory.

\section{ set\_dom\_timer(dom\_timer\_arg\_t *timer\_arg)} 
Request a timer event to be sent at the specified system time.

\section{ dom0\_op(dom0\_op\_t *op)} 
Administrative domain operations for domain management. The options are:

{\it DOM0\_CREATEDOMAIN}: create new domain, specifying the name and memory usage
in kilobytes.

{\it DOM0\_STARTDOMAIN}: make domain schedulable

{\it DOM0\_STOPDOMAIN}: mark domain as unschedulable

{\it DOM0\_DESTROYDOMAIN}: deallocate resources associated with the domain

{\it DOM0\_GETMEMLIST}: get list of pages used by the domain

{\it DOM0\_BUILDDOMAIN}: do final guest OS setup for domain

{\it DOM0\_BVTCTL}: adjust scheduler context switch time

{\it DOM0\_ADJUSTDOM}: adjust scheduling priorities for domain

{\it DOM0\_GETDOMAINFO}: get statistics about the domain

{\it DOM0\_GETPAGEFRAMEINFO}:

{\it DOM0\_IOPL}:

{\it DOM0\_MSR}:

{\it DOM0\_DEBUG}: interactively call pervasive debugger

{\it DOM0\_SETTIME}: set system time

{\it DOM0\_READCONSOLE}: read console content from hypervisor buffer ring

{\it DOM0\_PINCPUDOMAIN}: pin domain to a particular CPU

{\it DOM0\_GETTBUFS}: get information about the size and location of
                      the trace buffers (only on trace-buffer enabled builds)

{\it DOM0\_PHYSINFO}: get information about the host machine

{\it DOM0\_PCIDEV\_ACCESS}: modify PCI device access permissions


\section{network\_op(network\_op\_t *op)} 
update network ruleset

\section{ block\_io\_op(block\_io\_op\_t *op)}

\section{ set\_debugreg(int reg, unsigned long value)}
set debug register reg to value

\section{ get\_debugreg(int reg)}
 get the debug register reg

\section{ update\_descriptor(unsigned long pa, unsigned long word1, unsigned long word2)} 

\section{ set\_fast\_trap(int idx)}
 install traps to allow guest OS to bypass hypervisor

\section{ dom\_mem\_op(dom\_mem\_op\_t *op)}
 increase or decrease memory reservations for guest OS

\section{ multicall(multicall\_entry\_t *call\_list, int nr\_calls)}
 execute a series of hypervisor calls

\section{ kbd\_op(unsigned char op, unsigned char val)}

\section{update\_va\_mapping(unsigned long page\_nr, unsigned long val, unsigned long flags)}

\section{ event\_channel\_op(unsigned int cmd, unsigned int id)} 
inter-domain event-channel management, options are: open, close, send, and status.

\end{document}