client/server loopback hang

[Trond Myklebust]

On Mon, 2006-08-21 at 10:53 -0400, Dean Hildebrand wrote:
> In the 2.6.17 with citi patches (but it also seems to happen in the 
> latest 2.6.18+citi and 2.6.17-1.2142_FC4smp), my computer will hang in 
> the following situation:
> 1) Export a local ext3 directory
> rpc.mountd
> rpc.nfsd 5
> exportfs -r
> 
> 2) Mount the directory locally (loopback) via nfsv4 (I don't know about 
> nfsv3)
> 
> 3) Attempt to write data to the mountpoint to a file equal to the amount 
> of memory (RAM) on the machine.  I used iozone on one computer to do 
> this, but dd also worked on another computer.  The important thing is to 
> exhaust the available memory while writing the file.  Using 'top', I can 
> see that once the memory is fully utilized, the iozone process stops and 
> the computer/file system hangs. 
> I used 'iozone -aec -i 0 -+n -f /mnt/nfs4/io8 -r 2m -s 500m' on a FC4 
> i386 machine with 1GB ram.
> (or dd if=/dev/urandom of=/mnt/nfs4/d2 bs=1M count=500)\
> 

Known issue. The VM can deadlock the machine in this case since its
response to knfsd's requests for more memory is to ask NFS to write out
more pages. That is what the calls to nfs_release_page() are all about.

Cheers,
  Trond