NFS proc entry

NFS proc entry

[Fabian Frederick]

Hi,

	Do we have some /proc entry for realtime NFS access point report /
client like showmount does with RPC ?

Regards,
Fabian

Re: NFS proc entry

[Trond Myklebust]

På to , 15/04/2004 klokka 13:25, skreiv Fabian Frederick:
> Hi,
> 
> 	Do we have some /proc entry for realtime NFS access point report /
> client like showmount does with RPC ?

Exactly what possible justification would there be for putting something
like that into the kernel?

Cheers,
  Trond

Re: NFS proc entry

[Chris Friesen]

Trond Myklebust wrote:
> På to , 15/04/2004 klokka 13:25, skreiv Fabian Frederick:

>>	Do we have some /proc entry for realtime NFS access point report /
>>client like showmount does with RPC ?

> Exactly what possible justification would there be for putting something
> like that into the kernel?

I agree with you that it's kind of messy to put this in the kernel.

However, with the current setup filesystem monitoring deamons must fork 
off a child for each mount, since statfs() can block for many seconds if 
the server has gone away.


Chirs

Re: NFS proc entry

[Trond Myklebust]

På to , 15/04/2004 klokka 15:03, skreiv Chris Friesen:
> However, with the current setup filesystem monitoring deamons must fork 
> off a child for each mount, since statfs() can block for many seconds if 
> the server has gone away.

So exactly how would moving that monitoring into the kernel change the
parameters of the above problem? The kernel has exactly the same issues:
The only way it can tell if the server is down is by probing the
connection and timing out. That's the reason why those userland child
processes end up hanging in the first place.

Cheers,
  Trond

Re: NFS proc entry

[Chris Friesen]

Trond Myklebust wrote:
> På to , 15/04/2004 klokka 15:03, skreiv Chris Friesen:
> 
>>However, with the current setup filesystem monitoring deamons must fork 
>>off a child for each mount, since statfs() can block for many seconds if 
>>the server has gone away.

> So exactly how would moving that monitoring into the kernel change the
> parameters of the above problem?

I guess I was thinking that if the kernel knew the status of the mounts, 
it could speed things up for userspace apps that don't properly handle 
network filesystems.  Probably not practical though.

We had an interesting time converting some apps that were originally 
using a ramdisk to run with NFS-mounted files.  Since reading from the 
ramdisk never blocked for significant amounts of time, some of the apps 
didn't design for IO delay and behaved poorly the first time we pulled 
the links to the NFS server.

Chris