nfsd server cache flooded, try to increase nfsrc_floodlevel

[Rick Macklem]

Berend de Boer wrote:
> Dear all,
> 
> I have a subversion repository on an ZFS file system exposed through
> NFS on FreeBSD 10.0.
> 
> This is mounted by linux clients running Apache.
> 
> When people check out the repository, I get:
> 
>  nfsd server cache flooded, try to increase nfsrc_floodlevel
> 
Increasing the value for the sysctl vfs.nfsd.tcphighwater increases
nfsrc_floodlevel.
You might also want to decrease vfs.nfsd.tcpcachetimeo since the
large default timeout can result in a large cache.

Hard to say what a correct value is, but Garrett Wollman runs
large NFS servers (using ZFS storage) and sets them to:
vfs.nfsd.tcphighwater=100000
vfs.nfsd.tcptimeo=300
I think.

There have been patches added to head and stable/10 that help
to keep the cache from growing too big (thanks to Alexander Motin),
but you'd need to upgrade to stable/10 to get those.

rick

> when this is mounted using nfs4, on the client you see then:
> 
>   E175002: REPORT of '/!svn/vcc/default': Could not read chunk size:
>   connection was closed by server
> 
I am not sure if this is caused by the flooded DRC cache or not, but
if you still get this after you have increased vfs.nfsd.tcphighwater
so that the "try to increase nfsrc_floodlevel" no longer occurs,
email again.

> If nfs3 mount is used, there is no such problem.
> 
> This is probably a bug report, but not sure filing this will help
> anyone. So that's why I throw it out here.
> 
> --
> All the best,
> 
> Berend de Boer
>