XDR, RPC, NFS

1
XDR, RPC, NFS

• Internetworking with UNIX TCP/IP
• Winter 2002-2003

2
XDR

• Asymmetric data conversion
• Convert to what other entity understands!
• Symmetric data conversion
• Convert to a standard format!
• Network standard byte order
• XDR has implicit types
• 32 bit big-endian integer has exactly 32 bits
• No explicit information about the type of data

3
How to convert to XDR?

• XDR library routines are available
• These routines are machine dependent
• Convert data from native format to XDR and vice
  versa
• A buffer paradigm is mostly used
• Complete message is converted to XDR
• Code example
• Create a buffer to hold XDR data
• Append data to the buffer (or XDR stream)

4
XDR code example
include ltrpc/xdr.hgt define BUFFER_SIZE
4000 int i XDR xdr_stream char
bufBUFFER_SIZE xdrmem_create(xdr_stream, buf,
BUFFER_SIZE, XDR_ENCODE) i
35 xdr_int(xdr_stream, i) // convert integer
and append to stream // why are we
passing the address of i?
See figure 19.4, page 237 for other data
conversion routines
5
XDR code example, receiver
include ltrpc/xdr.hgt define BUFFER_SIZE
4000 int j XDR xdr_stream char
bufBUFFER_SIZE xdrmem_create(xdr_stream, buf,
BUFFER_SIZE, XDR_DECODE) // someone filled in
the XDR stream xdr_stream here! xdr_int(xdr_stre
am, j) // this is why we were passing the
address of i!
6
Using XDR with TCP and UDP

• TCP is stream oriented
• Can directly be used with XDR
• Open TCP socket and get a descriptor fd
• Use FILE fdopen(int fd, char mode) to associate
  a standard I/O stream to fd (socket fd)
• Use xdrstdio_create(XDR xdrs, FILE
  stdio_stream, enum xdr_op) to connect a new XDR
  stream to the I/O descriptor
• Subsequent calls to XDR conversion routines
  automatically send XDR data over TCP
• For use with UDP, records and datagram I/O are
  described in XDR (see xdrrec_create( ))

7
Remote Procedure Call

• An Extension of Procedural Call Model
• In many programming languages (not object
  oriented), complete program flow is described by
  procedure (or function) calls
• If the procedures cross the boundary of a
  computer, it becomes RPC
• A communication protocol is required to carry out
  the remote call to the procedure

8
Conventional Procedure Call
A single thread of execution Control transfer
from one procedure to another Variables of the
calling procedure are frozen Procedures calls
are usually iterative
main
proc_A
proc_B
9
Distributed Procedure Call
Imagine a different client-server model Server
implements a procedure called by client Client
interacts with server as a procedure
caller Server represents an implementer of the
procedure
Machine 1
Machine 2
Machine 3
main
proc_A
proc_B
10
Local calls vs. RPC

• For local calls, a procedure is inactive until it
  is called. For RPCs, a server process must exist
  waiting for a call
• Local procedures generally use a few arguments.
  RPCs can return arbitrary stream of data
• RPCs incur higher cost due to network delays
• Local procedures are generally in the same
  address space, so pointers are passable
• Remote procedures do not have access to callers
  I/O descriptors
• Remote procedures do not write data to the error
  file of a calling computer

11
Shared Data, Remote Programs
A single remote program
Procedure 1
Procedure 2
Procedure 3
Shared global data
12
RPC Number of Arguments

• Calls with large number of arguments are hard to
  read
• Data aggregation
• In a struct, for example
• XDR tells how to encode structs for sending these
  across to another machine
• Typically, all the arguments in RPC are collected
  in one single struct
• RPC Identification (prog, vers, proc)
• Prog 32 bits vers, proc integers

13
More on RPCs

• Mutual Exclusions
• At most one procedure in a remote program can be
  running at a time
• Ensures integrity of shared data
• Retransmissions
• Default timeout mechanism specifies a fixed and
  non-adaptive timeout
• A fixed number of retries is also specified
• Some applications may adapt the timeout on their
  own, but RPC does not adapt to network delays

14
RPC Call Message Format
16
0
31
MESSAGE ID
MESSAGE TYPE (0 FOR CALL)
RPC VERSION NUMBER (2)
REMOTE PROGRAM (0x000186a3 for NFS)
REMOTE PROGRAM VERSION (2)
REMOTE PROCEDURE (1 for GETATTR)
UNIX AUTHENTICATION
ARGUMENTS FOR REMOTE PROCEDURE (IF APPLICABLE)
15
Distributed File Systems

• Network File System (generic name) or Distributed
  File System
• Files are distributed over a network of
  computers
• Examples
• NFS
• AFS
• Coda

16
Design Goals

• Uses same/similar file system interfaces as used
  for local access
• Facilities
• Independent of user location
• Transparent to the end user
• Backups
• Caching
• Consideration of user mobility (?)
• Data sharing for a multitude of users

17
NFS

• Very similar to UNIX file system
• open( )
• read( )
• write( )
• close( )
• Provides file access rather than transfer
• One copy of the file is kept
• Independent of machine architecture
• Uses RPC for machine-to-machine communication
• Provides common login environment
• It is transparent to the end user

18
stat( ) call for file information

• stat(string filename, stat stat_block)
• Provides information on a UNIX file
• S_ISREG(stat_block-gtst_mode) //regular file
• S_ISDIR(stat_block-gtst_mode) //directory
• Symbolic links
• Created using ln s realfile linkfile
• Traversing a directory tree will not work using
  stat( ) call
• Use lstat( ) to traverse directories

19
File System

• Everything you can logically see, not necessarily
  on the same disk
• Data is organized in files files are organized
  in directories
• Each file (directory is a special file) is
  referenced by an i-node which contains
• stat( ) info
• Reference to the data block
• i-node numbers on a physical device are unique
• To differentiate between files and directories,
  looking at data block is not enough need to go
  to i-node

20
File System

• If i-numbers associated with the two files on a
  UNIX file system are same
• Data block for the two must be the same
• This is defined as hard link
• To create hard link
• ln realfile linkfile
• How to ensure that i-numbers are unique across
  machines in a distributed system?
• Cant ensure i-number duplication is allowed
• Hard links not allowed across physical devices

21
Functions of NFS

• Directory mount
• From other computers on the local network
• Directory export
• To other computers off the local network
• Has client-server architecture
• An NFS client is able to mount directories from
  multiple NFS servers
• Servers are stateless
• No information at the server about files open at
  the client

22
NFS File Attributes

• Just like stat provides information for UNIX
  files, file attributes for NFS are described in
  the fattr structure
• Instead of a stat( ) or lstat( ) call as in case
  of UNIX file system, a remote procedure call
  fills in the fattr structure

23
RPC for NFS

• Runs on top of UDP
• Clients specify block size, retransmit attempts,
  and timeout values
• rsize, wsize, retrans, timeo
• Default values are 8KB block, 5 retransmits, and
  1s timeout
• timeo is for the first request the value is
  doubled with every retransmit request
• Generally, processes requesting server files will
  block until server responds