An Example:
-----------
This example shows (very briefly!) how to install and run the NFS
server to export the directory /home/testnfs, and then to mount it on
the same machine as /mnt/nfs. It is NOT intended to be complete,
rather it reflects a simple test case that may help you understand
what this enhancement permits.
Volunteers to write real documentation are very welcome!
Installation
Unpack the tgz file using tar:
$ tar -xvzf ClusterNFS_1.0.0.tgz
change into the directory
$ cd ClusterNFS_1.0.0
and run "BUILD" and answer the questions
$ ./BUILD
now run make to compile the package
$ make
and then (as root) install the binaries and help pages using
$ make install
This last step is not necessary to run the examples below.
Setup:
Create the directory /home/testnfs:
$ mkdir /home/testnfs
and create some sample files
$ echo "SAMPLE" > /home/testnfs/sample.txt
$ echo "SAMPLE FILE FOR 127.0.0.1" > /home/testnfs/sample.txt\$\$IP=127.0.0.1\$\$
$ ls -l /home/testnfs
total 2
-rw-r--r-- 1 root staff 7 Mar 9 16:32 sample.txt
-rw-r--r-- 1 root staff 23 Mar 10 13:51 sample.txt$$IP=127.0.0.1$$
Now, enable NFS mounting of this directory by adding an appropriate line
to /etc/exports:
$ echo /home/testnfs localhost(rw,no_root_squash) >> /etc/exports
Now, start the NFS daemon with the command line option "--translate-names":
$ ./rpc.nfsd --translate-names
Start the mount daemon
$ ./rpc.mountd
Create (if necessary) the mount point /mnt/nfs
$ mkdir /mnt/nfs
This should let you mount the directory at /mnt/nfs
$ mount -t nfs localhost:/home/testnfs /mnt/nfs
(IF you get errors here, try starting nfsd and mountd using the parameters
-F -d auth -d call -d fhcache --translate-names
this will make them run in the foreground and display debugging information.)
Once you have mounted the directory, you can change to it
Usage:
Ask for a directory listing of the mounted directory:
Note that you only see one file, even though the directory actually
contains 2 files.
$ ls -l /mnt/nfs
total 1
-rw-r--r-- 1 root staff 26 Mar 10 15:01 sample.txt
The original sample.txt file has been "shadowed" by
the sample.txt$$IP=127.0.0.1$$ file.
(If you don't like this behavior, modify Makefile.in to comment out
the definition of READDIR_MASQUERADE. This will restore the full
directory listing.)
When we try to read or write "sample.txt" we get the contents of
"sample.txt$$IP=127.0.0.1$$ "instead:
$ cat /mnt/nfs/sample.txt
SAMPLE FILE FOR 127.0.0.1
If we need the contents of "sample.txt" rather than
"sample.txt$$IP=127.0.0.1$$", we can use the file name "sample.txt$$$$":
$ cat /mnt/nfs/sample.txt\$\$\$\$
SAMPLE
The same things happens for file writes. On the mounted filesystem,
$ echo Message from 127.0.0.1 > /mnt/nfs/sample.txt
actually modifies /home/testnfs/sample.txt\$\$IP=127.0.0.1\$\$ .
$ cat /home/testnfs/sample.txt\$\$IP=127.0.0.1\$\$
Message from 127.0.0.1
This also happens for directory access.
On the server, create two directories. On "base" directory and one "client"
directory:
$ mkdir /home/testnfs/test/
$ mkdir /home/testnfs/test\$\$IP=127.0.0.1\$\$
Now, on the client, creating the file "/mnt/nfs/test/Test"
$ echo Test > /mnt/nfs/test/Test
results in a file being created in the "/home/testnfs/test\$\$IP=127.0.0.1\$\$"
directory.
$ ls -la /home/testnfs/test\$\$IP=127.0.0.1\$\$
total 3
drwxr-sr-x 2 root staff 1024 Mar 10 15:20 .
drwxr-sr-x 4 root staff 1024 Mar 10 15:20 ..
-rw-r--r-- 1 root staff 5 Mar 10 15:20 Test
leaving the "plain" test directory is unchanged:
$ ls -la /home/testnfs/test
total 2
drwxr-sr-x 2 root staff 1024 Mar 10 15:20 .
drwxr-sr-x 4 root staff 1024 Mar 10 15:20 ..
!!! NEW !!!: ClusterNFS now supports filename translation on mount requests.
First, unmount the NFS filesystem
$ umount /mnt/nfs
Kill and restart the mountd daemon:
$ killall rpc.mountd
$ rpc.mountd --translate-names
Now, try mounting the 'test' directoy
$ mount localhost:/home/testnfs/test /mnt/nfs
Notice that the directory contains the file 'Test',
$ ls -l /mnt/nfs
total 4
-rw-r--r-- 1 root staff 5 Sep 24 18:31 Test
This demonstrats that the mounted directory is actually
/home/testnfs/test$$127.0.0.1$$.
Wish/Bug list:
1. Currently the name translation applies only to nfsd. Mount
requests, which are serviced by mountd, do not have the name
translation. This could easily be implemented.
[ Fixed in version 1.0.0-pre1, 2001-09-24 ]
2. Currently, there is no method for causing file creation to occur
with the an extended file name. EG, there is no way to ensure
that when a client creates a file named 'muck' that it will get
created on the server with the name 'muck$$IP=xxx.yyy.zzz".
One idea for implementing this is allowing the user to create a
file named "muck$$CREATE=IP$$". This would mean 'Append the IP
number to files created with the name "muck". It would also seem
useful to allow wildcards like "*$$CREATE=IP$$" meaning 'On any
file creation, append the IP number.'
3. The current implementation, searching sequentially along a fixed
list of possible translations, has a number of inefficiencies and
limitations. A more efficient and more flexible approach would
be to 'reverse' the process.
When a file request arrives,
A) do a directory listing to find all files that match the
base name
if the request is for 'my.file', search for
'my.file\$\$*\$\$' plus 'my.file'
B) use the discovered entries as patterns to match against,
possibly allowing wildcards.
this will, of course, require definition of rules for
resolving multiple matches. A reasonable approach is that
files which match on the largest number of fields match first.
I.E., if the list of matching files is
my.file$$IP=10.0.0.14$$
my.file$$GID=57,HOSTNAME=bee*$$
my.file
we would first examine "my.file$$GID=57,HOSTNAME=bee*$$",
since it has the most fields (2). Parsing it, we arrive at
two tests, first testing if GID=57. If not, we can fall
through to the next match. If this matches, we then have
another test "HOSTNAME=bee*". This time we have a wildcard so
we need to do pattern machine.
If the file matched and is accessible, (readable) return it,
otherwise, try the next filename in the list.