Re: [Nbd] NBD wishlist items?

[Mike S. <sn...@gm...>]

On 6/26/07, Paul Clements <pau...@st...> wrote:
> Wouter Verhelst wrote:
>
> > A protocol extension that lets both the server and the client initiate
> > this keepalive probe would deal with both issues (provided we do not
> > somehow swap out nbd-client).
>
> Do we need the server to do this keepalive stuff, though? I mean, it's
> really the client that cares if the connection goes down. The server
> ought to be happy to just sit there -- if it's not getting any requests,
> it shouldn't matter.

I agree, having the nbd-client do a keepalive is all that should be needed.

> >>> Implementing this backwards-compatibly is going to be the hardest part,
> >>> I guess. Perhaps opening one of the NBD devices to call a specific ioctl
> >>> to verify whether it supports this interface, and then setting a bit in
> >>> the field of 'reserved' bits in the handshake could work, but I'm not
> >>> sure how this would be best done.
> >> There'd probably be an ioctl to enable this ping packet, so you could
> >> just call the ioctl and ignore failure (for older kernels). Then there'd
> >> be a new packet flag for this ping packet (or maybe just use a
> >> zero-length read or write, assuming the nbd-server just ignores those
> >> and returns a response). That way it would almost completely backward
> >> compatible.
> >
> > Sorry, I'm not sure I understand that bit. Could you elaborate?
>
> I was just thinking out loud. Basically, an easy and backward compatible
> way to do this (client-side) ping is to periodically send a zero-length
> read or write, and as long as the server just ignores it (i.e., doesn't
> crash or do something otherwise stupid) and sends a response, everything
> will work and the protocol doesn't have to change.

How does one do a zero-length read or write?  I've been testing
various userspace methods to detect the nbd-client's connection to be
hung (TCP timeout).  Using dd is unreliable because the read request
is fulfilled from cache if you don't use O_DIRECT.

If we could do a zero-length write then this would very easily
excercise the nbd connection without any protocol change or O_DIRECT
reads.  Can you elaborate on what is needed for reliable zero-length
writes?  The 'man 2 write' states:
"If count is zero and the file descriptor refers to a regular file, 0
may be returned, or an error could be detected.  For a special file,
the results are not portable."

thanks,
Mike