NFS : soft mount, but operations still hang

contremaitre · 01-27-2010, 07:34 AM

Hi,

I mounted a nfs server with :
-o soft,intr,nolock,proto=udp,timeo=2,retrans=1
(tried both tcp and udp)

and then :
dd if=/dev/zero of=/mnt/mynfs/test.dd

Then I unplug my network cable, and the dd command never terminates.

Why ?

Thanks.

I tried a 2.6.27 and 2.6.32 kernel, for the NFS client. My NFS server is a NAS.

phlyer · 01-27-2010, 11:58 PM

did you try "hard" in combination w/ "proto=tcp" ?

did you try nfs4 yet?

HTH

contremaitre · 01-28-2010, 01:28 AM

Quote:

Originally Posted by phlyer

did you try "hard" in combination w/ "proto=tcp" ?

I just did, and nfs calls never terminate as well. But it doesn't surprise me, because man nfs says :

Code:

if the hard option is specified, NFS requests are retried indefinitely

What surprises me, is that man nfs explicitly says :

Code:

If the soft option is specified, then the NFS  client  fails  an  NFS request 
after retrans retransmissions have been sent, causing the NFS client to return 
an error to the calling application.

Quote:

did you try nfs4 yet?

No, because my NFS server doesn't support it.

phlyer · 01-28-2010, 01:38 AM

ok, that sheds some light
what NFS are you using?

on what distro?

you have two physical machines? I assume

are you using multi-homing on the exporting machine?

what's your /etc/exports entry for that NFS client?

contremaitre · 01-28-2010, 02:10 AM

I am using nfs on ubuntu 8.10 for the client.
But my NFS server is a NAS. (QNAP) with

Code:

[~] # cat /etc/exports 
"/share/MD0_DATA/Public" *(rw,async,no_root_squash) 
"/share/MD0_DATA/disk1" *(rw,async,no_root_squash)

phlyer · 01-28-2010, 02:55 AM

nice one on the QNAP
is it some SOHO (single NIC) or some TS-809U-RP model?

maybe you could restrict access, and also drop the tuning?
just for testing

Code:

"/share/MD0_DATA/disk1" 192.168.1.0/255.255.255.0(rw,sync,no_root_squash)

mount your device

Code:

mount -t nfs qnap:/share/MD0_DATA/disk1 /whatever -o rsize=BLKSIZ,wsize=BLKSIZ,intr,soft,proto=tcp

then try again using with

Code:

dd if=/dev/zero of=/mnt/mynfs/test.dd bs=BLKSIZ

where BLKSIZ is whatever is defined in /usr/include/nfsd/const.h
if in doubt, set BLKSIZ=8192
thing is the TCP block size in Linux Kernels >= 2.4 is 32K, NFS version 2
has 8K defined and in NFS version 3 it is up to the server.

You may consider to drop the block size altogether (rzise,wsize &dd).

I've heard about some issues w/ Ubuntu mount.nfs but I am no Ubuntu expert. Sorry.

contremaitre · 01-28-2010, 03:27 AM

I have the TS-439 Pro

I tried to modify the options in /etc/exports, but when I disable/enable nfs my modifications are removed. And if I don't, I don't know if they are taken into account.

But I tried it anyway, with your blocksize suggestions, and it's still blocking.

contremaitre · 01-29-2010, 01:59 AM

I found that with sync writes (client side, ie dd -oflags=sync), nfs is not blocking indefinitely anymore.
But there is no mention of this anywhere in nfs documentations.
Also, the error is reported long after I unplug the cable (up to 15 minutes), but if I unplug the cable 10s, then plug it again, the error is reported just there.
And I would prefer to use async writes so I am still open to suggestions.

fsogumo · 02-12-2010, 11:43 AM

We had the same problem here but after two days investigating it it seems we found an explanation for this. The default protocol for NFS clients is TCP which seems to set its own timeout value. If you use UDP as protocol NFS clients will consider the "timeo" parameter timeout value. After unplugging the network cable the client program will stop after exhausting the local buffer and the timeout time.

contremaitre · 02-14-2010, 12:24 PM

I am using UDP but nfs is still blocking.
It seems it depends if the local buffer is full or not.
If I open my files with the SYNC option, then the local buffer will never grow, and the timeout value will be respected.
Otherwise, it never times out.