[syslinux] PXELinux Kernel NFS Root Errors

[Manu Bhardwaj]

Hi

This is a PXELinux problem that I'm facing. The problem, in brief, is that I
have a set of diskless nodes (AMD Athlon XP 1800+ with 256MB RAM) connected via
an internal 172.20.0.0/24 network to a dual AMD Opteron running SuSE Linux 8.3
for AMD64. I want the remote machines to diskless boot off the main server and
run independently. The diskless machines have a PXE-compliant BIOS for booting
off the network.

Details of absolutely everything I did has been written down at the end of this
email. I will briefly summarize what I did here, and ask the questions here. For
further details, please read the rest of this email and/or ask me specific
questions.

Summary: I got the right linux kernel, I set up DHCP and the right TFTP (without
blksize enabled) and NFSD on the server. I switched on the remote diskless
machine, and everytime I switch it on, I get the same RPC error which has
something to do with being unable to mount NFS as root. I have even tried using
an intermediate ramdisk, but it hasn't worked (possibly because my ramdisk
wasn't ideal). My questions thus are:

- What specifically does one do to remote-mount a root partition over NFS while
using such a setup?
- Would I even require a ramdisk in such a case, or is there a way to get the
remote machines working after bypassing the ramdisk?
- I do not have much experience with creating ramdisks. If one is required, what
does it need to contain to work properly?
- Will an option of using a floppy to mount a very basic root partition,
followed by an NFS mount of root and a 'chroot' command, be an effective
solution? Or will it be too slow/a stopgap approach?

ANY suggestions will help.

Thanks in advance,
Manu

-- 
Manu Bhardwaj   <http://manubhardwaj.net>
                <PGP: 0x7EF46A88>
--

Details:

I have searched the web and spent some days trying to get the system up and
running. (Server address = 172.20.0.1, netmask 255.255.255.0). The server has
been different at different points of time (sometimes TurboLinux for the AMD64,
sometimes Debian Woody 32bit, and now SuSE for the AMD64. The errors are ALWAYS
THE SAME.)

These are the steps I followed:

- Get ISC's DHCPD up and running off the server. It broadcasts addresses in the
range 172.20.0.11 to 172.20.0.19. The relevant entries in /etc/dhcpd.conf for
the subnet are:

range 172.20.0.11 172.20.0.13;
filename "/tftpboot/pxelinux.0";
#root-path "/";

- The diskless nodes, on startup, immediately obtain an address from the range.
Typically, the first machine ALWAYS chooses the address 172.20.0.13. It also
shows the correct netmask 255.255.255.0.

- TFTP then starts up on the client machine. (I use the inetd tftpd standalone
server with the command line

# in.tftpd -l -s /tftpboot -v -r blksize
)

- TFTP downloads the 11kb pxelinux.0 kernel, and then reads the file
/tftpboot/pxelinux.cfg/default (which is the only file in that directory). The
contents of this file are:

DEFAULT vmlinuz_32bit_remote root=/dev/nfs nfsroot=172.20.0.1:/ ip=dhcp
#ipappend 1

The kernel vmlinuz_32bit_remote is a linux 2.4.18 kernel compiled on an Athlon
XP machine. It has IP: kernel level autoconfiguration (DHCP, BOOTP, RARP)
enabled; it also has NFS server and NFS as root options enabled. It also has
ramdisk option enabled (4096kb). None of them are modules.

Sometimes, I have tried passing another static IP address to 

- I have enabled NFS on the server. I tried using Knoppix and Gentoo Live on the
diskless node, and tried remote mounting. It works perfectly and flawlessly. The
contents of /etc/exports have been, at different points in time, one of these:

/tftpboot/remote    172.20.0.11(rw,no_root_squash)
/tftpboot/remote    172.20.0.12(rw,no_root_squash)
/tftpboot/remote    172.20.0.13(rw,no_root_squash)

or

/tftpboot/remote    172.20.0.0/24(rw,no_root_squash)

or

/tftpboot/remote    172.20.0.0/255.255.255.0(rw,no_root_squash)

or

/tftpboot/remote    (rw,no_root_squash)

but have made no difference at any point in time. /tftpboot/remote has a
complete Debian Woody 3.0r0 32-bit tree on it (got by using a hard drive on the
remote machine, compiling a kernel for it and scp'ing it to the server before
removing the hard drive from the remote machine).

- The contents of hosts.allow have been:

ALL : ALL at ALL : ALLOW

or

ALL : ALL at 172.20.0.0/24 : ALLOW

etc. etc. Basically, everybody gets every permission. At other times, I have
tried manual selection, such as

portmap : ALL at 172.20.0.13 : ALLOW
rpc.mountd : ALL at 172.20.0.13 : ALLOW
rpc.statd : ALL at 172.20.0.13 : ALLOW

and all combinations of variations of these lines for the network, and for other
IP addresses.

(When this is the case, the remote machine searches every ASCII file from
01-00-00-0a-0b-0c and from A0ABCDEF (eg.) to A, before finally settling on the
file 'default'. This takes 30 seconds when hosts.allow is set to "all all all
all", but takes about 1 second when hosts.allow only allows portmap, mountd,
etc. etc. selectively. Why? But that's beside the point.) 

-The remote machine then easily downloads the linux kernel image via the network
using TFTP, and then seems to work for about 10 seconds. I cannot actually type
out all the contents of what is displayed (I grepped all files on the server but
couldn't find where the remote machine's kernel logs go), but here is the jist
of the output -

1. The kernel seems to identify hda and hdb from the SERVER rather than from
itself. Obviously, it should not identify more than one device (the CDROM on its
own machine - hdc - Secondary Master), but it identifies three drives - clearly
the server's drives.

2. Most importantly, it stops at this point:

Looking up port of RPC 100003/2 on 172.20.0.1
RPC: sendmsg returned error 101
Root-NFS: Unable to get nfsd port number from server, using default
Lookingup port of RPC 100005/1 on 172.20.0.1
RPC: sendmsg returned error 101
Root-NFS: Unable to get mountd port number from server, using default
RPC sendmsg returned error 101
Mount: RPC call returned error 101
Root-NFS: Server returned error -101 while mounting /
VFS: Unable to mount root fs via NFS, trying floppy.
VFS: Insert root floppy and press ENTER

So I assumed it was an NFS error (though, as I mentioned before, a clean NFS
mount using Knoppix works perfectly). So when I try

# rpcinfo -p

on the server, I get references to 
100005/1 and 100005/2 (tcp and udp) mountd 
as also references to 100003/something (tcp and udp) portmapper, along with some
others. Obviously, they are working on the server.

I then created a ramdisk image, and modified pxelinux.cfg/default and appended
the lines

root=/dev/ram0 initrd=image.gz 

to the kernel. But no difference - RPC errors still thrown up.

End.