Re: nfs weirdness

[Neil Brown]

On Monday July 23, [e-mail] wrote:
> 
> yesterday cleanly installed a system that is known to export three dirs
> without problems with NFS on 2.2.16.
> 
> The system is at 2.4.4 (SuSE) and while the same exports file is used, I
> get a problem that is identified with exportfs.
> 
> using knfsd, exports looks like :
> 
> /cdrom          192.168.1.2(ro)
> /windows        192.168.1.2(ro)
> /udf            192.168.1.2(ro)
> 
> exportfs -av with :
>  
> /dev/hdd on /udf type udf; /dev/hdc on /media/cdrom type iso9660 and
> /windows is NOT mounted.
> 
> exporting 192.168.1.2:/media/cdrom
> exporting 192.168.1.2:/windows
> exporting 192.168.1.2:/udf
> reexporting 192.168.1.2:/media/cdrom to kernel
> reexporting 192.168.1.2:/windows to kernel
> reexporting 192.168.1.2:/udf to kernel
> 
> looks good. remote mounts als shows no error :
> 
> toshiba:/cdrom          575506    575506         0 100% /tcdrom
> toshiba:/windows        976216    895820     30804  97% /windows
> toshiba:/udf            545984    530948     15036  98% /udf
> 
> except that /windows was NOT mounted so should be empty. it in fact
> reflects the fs size, free etc of the root FS of toshiba. ls /windows on
> 192.168.1.2 shows no files.
> 
> now, if I mount /windows on toshiba, there are files. all ok, however,
> exporting it causes :
> 
> reexporting 192.168.1.2:/windows to kernel
> 192.168.1.2:/windows: Invalid argument
> 
> (/dev/hda1 on /windows type vfat)
> 
> and mounting via nfs :
> 
> mount: toshiba:/windows failed, reason given by server: Permission
> denied
> 
> /v/l/m shows :
> 
> rpc.mountd: authenticated mount request from 192.168.1.2:670 for /windows (/windows)
> rpc.mountd: getfh failed: Operation not permitted
> 
> I'm lost, completely now. it worked with the old kernel/setup, now it
> doesn't.

What you describe is exactly what I would expect.  It might work
differently with unfsd - the user space nfs server, but this is all
correct for knfsd.

knfsd exports filesystems, or parts there-of.  It doesn't export
'parts of the visible namespace'.

If you ask to export "/windows" and nothing is mounted on "/windows",
then you are asking to export part of the root filesystem starting at
"/windows".  If you subsequently mount something on /windows, then you
haven't asked for that to be exported so it won't be, and mountd will
get confused.
You should always mount filesystems before trying to export them.

Now I appreciate that this could all be a bit more transparent, but
unfortunately it isn't and probably won't be for a while.

Possible changes to reduce the confusion:

1/ Don't allow a directory that is exported to be mounted on.
2/ Have a "mountpoint" export option which means "only export this
    directory if it is a mountpoint".

Patches welcome.....

NeilBrown