[Nbd] Re: nbd status report!

[Wouter Verhelst <wouter@...3...>]

On Wed, Jul 07, 2004 at 10:02:14AM +0200, Auke Kok wrote:
> 
> Hi,
> 
> I thought I'd let you know what the stability was since I had some 
> problems with it.
> 
> I've been stresstesting the nbd setup a bit and it turns out to be 
> horribly unstable. Doing relatively easy copies from my nfs server to 
> the setup I have with nbd cause many 'connecttion reset by peer's 
> causing my nbd-clients to end. This of course results in the raid system 
> losing a disc. I ended up doing a cp from an nfs mount to the raided-nbd 
> volume and received 5 disconnects within the first 2 minutes. This 
> rendered my raid disc offline and end-of-exercise

Ugh.

> points learned: nbd-client should 'hold on' to the /dev/nbd/X device 
> node and try reconnecting. In true nfs spirit this should even be a 
> 'hard' mount, so it will try until eternity. Only in special cases you 
> might want to use 'soft-mount' like disconnects in case of timeouts. 
> These timeouts would really have to be customizeable and set to long 
> periods (think going away for a weekend and your network hub dies).

Yeah, indeed.

I have to point out, though, that everything the user-space nbd-client
program does, is
* set up a socket
* do the initial handshake with the server
* pass the connected socket to the kernel, using an ioctl() call

after that, the kernel takes over; and since I'm not into kernel-space
coding (yet?), I'm not maintaining that part of the software.

You may want to contact Paul Clements (see the MAINTAINERS file in the
kernel source tree) with your results; I'm sure he'll be happy to learn
about them.

> I hope you can do something with this! I get the feeling that you're not 
> getting that much feedback so I'll try to keep you posted.

That's true, and I appreciate it :-)

-- 
         EARTH
     smog  |   bricks
 AIR  --  mud  -- FIRE
soda water |   tequila
         WATER
 -- with thanks to fortune