Re: [Nbd] [PATCH] nbd: improve request timeouts handling (rebased)

[Goswin von Brederlow <goswin-v-b@...186...>]

On Wed, Jul 10, 2013 at 09:40:42PM +0200, Michal Belczyk wrote:
> [rebased against the master, no other changes]
> 
> The main idea behind it is to be able to quickly detect broken replica
> and switch over to another when used with any sort of mirror type device
> built on top of any number of nbd devices.

It is good to see someone checking and fixing timeout issues.

But I would also like to have the opposite behaviour and delay
detecting failures. Insane? read on...


I have a small server that runs 24/7 with minimal power consumption. I
also have a storage node that exports a number of disks via NBD. This
should be powered down when not in use. The problem is that "not in
use" is incompatible with NBD as is. The storage node must be running
and must have an active connection to the server.

What I'm suggesting is that the nbd-client should be able to create a
network block device in disconnected state (without attached socket)
and the kernel would return on the next access. The nbd client can
then actually connect to the NBD server and restart the device with a
socket to fullfill the request. In my case that would involve a
wake-on-lan call to first bring up the storage node.

Secondly it would be nice to have an idle timeout in the kernel. If no
access happened to the NBD for some time the kernel should return so
the client can switch to disconnected mode. In my case that would
involve sending the storage node to sleep.

Interested in trying to implement this?

MfG
	Goswin