6.824 - Spring 2005

Lab 4 : MKDIR, REMOVE, and Locking

Last change: $Date: 2005/03/14 11:53:35 $

Due: Thursday March 3rd, 1:00pm.


In this lab, you'll continue implementing your file server. You'll start with the code you wrote for Lab 3, add handlers for the MKDIR and REMOVE RPCs, add support for the mtime attribute, and add simple locking. The mtime attribute is neccessary to help the NFS client code check whether locally-cached file and directory contents are up to date. The locking is required to ensure that concurrent modifications to the same file or directory occur one at a time.

Getting Started

Your first task is to add the lock client module to your file server, and your lock server module to the block server. The latter is only for convenience, the two servers are logically separate.
anguish% wget -nc http://pdos.lcs.mit.edu/6.824/labs/lab-4.tgz
anguish% tar xzf lab-4.tgz
anguish% cd lab-4
anguish% cp ../lab-3/fs.[Ch] .
anguish% cp ../lab-1/lock_server.[Ch] .
Edit fs.h (in the lab-4 directory) and add the following early in the file:
#include "blockdbc.h"
#define LOCKS 1 // ADD
#include "lock_client.h" // ADD
Now change the fs constructor to take a lock client argument, and add a new class member to hold it:
class fs {
  fs(blockdbc *db, lock_client *lc); // CHANGE
  lock_client *lc; // ADD
Now change fs.C:
fs::fs(blockdbc *xdb, lock_client *xlc) // CHANGE
  db = xdb;
  this->lc = xlc; // ADD
At this point you should be able to run gmake and successfully compile your file server and a block server. Operations that worked in Lab 3 should still work. You have to run this new block server with your new ccfs, since now ccfs expects a lock server to be resident in the block server process.

The rest of this lab has two parts.

Part 1: Your Job

Your job in part 1 is to handle the MKDIR and REMOVE RPCs, and to add support for the mtime attribute.

Your server should update a file's mtime when SETATTR changes its size, during a WRITE, and when the file is created. The server should update a directory's mtime whenever an entry is added or removed from the directory. In all cases the server should set the mtime to the current time with nfstime().

If your server passes the tester (see below), then you are done. Please modify only fs.C and fs.h. You can make any changes you like to these files.

When you're done with Part 1, the following should work:

suffering% mkdir /classfs/dir1/newdir
suffering% echo hi > /classfs/dir1/newdir/newfile
suffering% ls /classfs/dir1/newdir
suffering% rm /classfs/dir1/newdir/newfile
suffering% ls /classfs/dir1/newdir

Part 1: Testing

You can test your file server using the test-lab-4-a.pl script. The script creates a directory, creates and deletes lots of files in the directory, and checks file and directory mtimes. If the tester is happy, it will say "Passed all tests!". Here's a successful run:
suffering% ./test-lab-4-a.pl /classfs/dir1
mkdir /classfs/dir1/d3319
create x-0
delete x-0
create x-1
checkmtime x-1
delete x-33
Passed all tests!

Part 2: Your Job

Your server's CREATE handler probably reads the directory's contents from the block server, makes some changes or additions, and writes the new contents back. If two clients were to issue simultaneous CREATEs for different file names in the same directory to two ccfs servers sharing the same file system, chances are that only one file would be exist in the end. The correct answer, however, is for both files to exist.

A convenient way to fix this "race condition" is for the file servers to use locks to force the two CREATE operations to happen one at a time. That is, a server would acquire a lock before starting the CREATE, and only release the lock after finishing the write of the new information back to the block server. If there are concurrent operations, the locks force one of the two operations to delay until the other one has completed.

You must choose what the locks refer to. At one extreme you could have a single lock for the whole file system, so that operations never proceed in parallel. At the other extreme you could lock each entry in a directory, or each field in the attributes structure. Neither of these is a good idea! A single global lock prevents concurrency that would have been OK, for example CREATEs in different directories. Fine-grained locks have high overhead and make deadlock likely, since you often need to hold more than one fine-grained lock.

Your best bet is to associate one lock with each file handle. Use the file handle as the name of the lock (i.e. pass the file handle to acquire and release). The convention should be that any NFS operation should acquire the lock on the file or directory it uses, perform the operation, finish updating the block server (if the operation has side-effects), and then release the lock on the file handle. One reason this is convenient is that you can add locking relatively simply by calling acquire in fs::get_fh(); thus all of your RPC implementations will acquire file handle locks without you having to modify your RPC handler code. Similarly, you can modify fs::put_fh() to release the file handle lock. You'll still need to add explicit releases for RPC handlers such as GETATTR that don't write the block store, and for error cases.

You'll use your lock server from Lab 1. lab-4.tgz includes a blockdbd.C that incorporates your lock server code, so that you don't have to start a separate server. blockdbd looks at the RPC program number to figure out whether it's a locking RPC or a block RPC. We've also included a new ccfs.C that instantiates a lock_client and connects it to your lock server. You can call acquire() and release() from fs.C as in this example:

fs::acquire(nfs_fh3 fh, callback<void,void>::ref cb)
  lc->acquire(fh2s(fh), cb);

fs::release(nfs_fh3 fh)

Part 2: Testing

Test your locking with test-lab-4-b. The tester takes two directories as arguments. It issues concurrent operations in the two directories and checks that the results are consistent with the operations executing in some serial order. Here's a successful execution of the tester:
suffering% ./ccfs dir1 blood 5566 &
root file handle: f7dbbbadfd082018
suffering% ./ccfs dir2 blood 5566  f7dbbbadfd082018 &
root file handle: f7dbbbadfd082018
suffering% ./test-lab-4-b /classfs/dir1 /classfs/dir2
Create then read: OK
Unlink: OK
Append: OK
Readdir: OK
Many sequential creates: OK
Write 20000 bytes: OK
Concurrent creates: OK
Concurrent creates of the same file: OK
Concurrent create/delete: OK
Concurrent creates, same file, same server: OK
test-lab-4-b: Passed all tests.
If you try this before you add locking, your server will probably fail the "Write 20000 bytes" test or the "Concurrent creates" test. The reason the "Write 20000 bytes" test might fail is that it produces three concurrent WRITE RPCs (of 8192, 8192, and 3616 bytes).

You might want to test your solution to Part 2 with test-lab-4-a first, to make sure you didn't break anything. You might then test with test-lab-4-b, but giving it the same directory twice, to make sure you handle concurrent operations in one server before you go on to concurrent operations in two servers.


You should consult the NFS v3 specification for descriptions of what the NFS RPCs should do. You may also want to consult the book "NFS Illustrated" by Brent Callaghan.

You can learn more about NFS loopback servers and asynchronous programming in the loop-back NFS paper. You can find the sources for this software at www.fs.net or in /u/6.824/sfs-0.7.2 and /u/6.824/classfs-0.0. You can see the NFS RPC definitions in /u/6.824/sfs-0.7.2/svc/nfs3_prot.x. The output of the RPC compiler is in /u/6.824/sfs-0.7.2-build/svc/nfs3_prot.h. You can learn more about the asynchronous programming library (wrap, callbacks, str, and strbuf) by reading the Libasync Tutorial.

Don't use a nfscall structure after you send a reply or error. The RPC library de-allocates the structure when it sends the reply. You may be tempted to use the arguments in the nfscall structure after replying in order to release locks: don't. Instead, make a copy of the arguments, or release the locks before replying.

Here's how to extract the arguments from a MKDIR RPC:

  mkdir3args *a = nc->template getarg<mkdir3args> ();
  // a->where.dir : existing directory
  // a->where.name : name of new directory
  // a->attributes.mode.set : if non-zero, set new dir's fattr3.mode to...
  // *a->attributes.mode.val : the new directory's mode
Here's how to reply to a MKDIR RPC. before_attrs and after_attrs are the attributes of the containing directory before and after the MKDIR.
  diropres3 *res = nc->template getres<diropres3> ();
  *res->resok->obj.handle = fh;
  res->resok->dir_wcc = make_wcc(before_attrs, after_attrs);
You may need to create, or fake, a "." entry in each new directory, referring to the directory itself.

Here's how to extract the arguments from a REMOVE RPC:

  diropargs3 *a = nc->template getarg<diropargs3> ();
  // a->dir : the directory
  // a->name : the file name

Here's how to reply to a REMOVE RPC:

  wccstat3 *res = nc->template getres<wccstat3> ();
  *res->wcc = make_wcc(before_attrs, after_attrs);

Collaboration policy

You must write all the code you hand in for the programming assignments, except for code that we give you as part of the assigment. You are not allowed to look at anyone else's solution (and you're not allowed to look at solutions from previous years). You may discuss the assignments with other students, but you may not look at or copy each others' code.

Handin procedure

You should hand in the gzipped tar file lab-4-handin.tgz produced by running these commands in your ~/lab-4 directory.
% tar czf lab-4-handin.tgz *.[Cchx] Makefile
% chmod og= lab-4-handin.tgz
Copy this file to ~/handin/lab-4-handin.tgz. We will use the first copy of the file that we find after the deadline.