RE: Avoiding deadlock in iSCSI

[David Robinson <David.Robinson@EBay.Sun.COM>]

> With separate TCP connections, there is no guarantee which connection will
> be sending with the lowest latency.  One would expect a command stream to be
> less, but it may be without inertia.  A command given to the NIC ahead of
> data may arrive after the data.  The arrival time of two streams sending
> different quantities would be difficult to predict.  As such, out of
> sequence delivery.  Would you recommend a skew buffer?

This is not an issue. If you preserve the ordering of data to
match the ordering of the commands things work.  You will need
to buffer data from the data connections until the commands are read
and processed.  A target that opens N connections without
N*max_window_size buffers available is poorly configured so buffering
should not be an issue.

	-David