Most recent kernel where this bug did not occur: 18.104.22.168
Distribution: Debian/unstable, i386
Hardware Environment: AMD 2600+, 2GB RAM, see http://nerdbynature.de/bits/2.6.24-rc2/nfsd/dmesg.2.gz
Software Environment: nfs-kernel-server 1.1.1-8
Exported NFS shares can be mounted (client: 2.6-git/powerpc32,
nfs-common-1.1.1~git-20070709-3ubuntu1), but running "ls /mountpoint" (even without "-l") on the client is enough to get the [nfsd] processes in "D" state.
Doing this again will put more [nfsd] processes in "D" state and the loadavg is incremented +1. Restarting the rpc.nfsd process on the does not help much, the new rpc.nfsd processes get stuck quickly.
This has been reported by Chris Wedgwood too:
More details: http://www.nerdbynature.de/bits/2.6.24-rc2/nfsd/
Steps to reproduce:
* mount -o vers=2 (tested, works)
* mount -o vers=3,nordirplus
* set CONFIG_NFSD to "depends on !XFS && ..."
*** Bug 9377 has been marked as a duplicate of this bug. ***
*** Bug 9369 has been marked as a duplicate of this bug. ***
This is clearly a regression. (But I don't seem to have the rights to change it.)
Patch is available: http://lkml.org/lkml/2007/11/25/39
fix is still not in 2.6.24-rc4, due to undergoing more QA:
Author: Christoph Hellwig <firstname.lastname@example.org>
Date: Fri Dec 7 14:07:53 2007 +1100
[XFS] revert to double-buffering readdir