Re: Avoiding deadlock in iSCSI

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

"Randall R. Stewart" wrote:

> Your reply leads me to believe a tight coupling exists between
> the iSCSI layer and the TCP stack. I.e. iSCSI is going in
> and "tweaking on a dynamic basis" the TCP implemenations
> rwnd. Other email that has crossed on this subject implies
> that "a TCP implementation is broken if it offers more
> buffers than it has"... This in some ways contradicts iSCSI
> controlling TCP rwnd and TCP controlling its own rwnd...
> 
> could you please clearify this for me? Is it TCP in control
> of rwnd or is there a tight coupling here?

I envision that iSCSI controls the rwnd by simply making a choice on
whether to read from that connection or not.  It of course tells
the TCP layer what its initial buffers are (ala setsockopt)
but if it doesn't read from a data connection that will cause
the window to shrink in response to data arriving causing
flow control.

In my extreme example where iSCSI maintains a zero window, this
would require iSCSI to have a hand into the TCP stack.  I am not
advocating this but just making a point on what could be
done at the extreme, but iSCSI should not require an
implentation to violate layering.

	-David