Re: iSCSI: unsolicited Vs immediate; restart delay

[Stephen Bailey <steph@cs.uchicago.edu>]

> There's no architected mechanism to figure out on its own!  Either it 
> has to keep the (session, connection, task) states around for a long time, 
> OR it can clean up the these states and trigger an unnecessary ULP 
> recovery on the initiator which is surprised at an unexpected restart failure
>  
> (keep in mind that initiator may not be able to restart a login precisely 
> after the "minimum time" specified, typically there's an aggregation 
> of multiple O/S timer requests into one master handler).  Providing an 
> upper limit is a cleaner, safer design for both initiator and target.

I agree with Mallikarjun.

This is a big problem already faced by increasingly general FCP
configurations.  The idea that RR_TOV (the resource recovery timer,
which is equivalent to the timer being discussed here) could be
fixedly specified (as opposed to communicated in the protocol) only
really works with a narrow set of assumptions about the application
configurations.

With many order of magnitude range in delays, as happens with
`internet' protocols, we really can't specify a fixed delay.
Furthermore, in this particular case, we can't really punt it to the
target to select.  The target has no idea.  The sad fact is that the
initiator may not have much of an idea either, but if we don't
provide the knob, a target choice is guaranteed to be wrong in some
cases.

Of course, you probably remember that I'm not a fan of recovery in any
case, so my solution would be to punt recovery altogether.  Clearly
that's not going to happen, so the next choice is to make it work :^)

Steph