Re: WG chair on buffer exhaustion (again)

On Friday, August 1, 2003, at 08:42 AM, Kanevsky, Arkady wrote:

> RC has has parts to its definition.
> One of them states that any failure to deliver message (process posted
> WQE) over connection breaks the connection.
>
> Are you saying that failure to deliver message causes retransmission or
> that failure to deliver message causes CQE with ~success but connection
> stays?

Any failure will result in terminating the stream (which for
MPA/TCP means breaking the connection). The proposal is to
essentially allow a "conditional placement" interface from
the LLP to DDP.

If a segment/chunk would not be placeable immediately, the
LLP would have the *option* of *delaying* placement. Exactly
what tricks it would use to delay placement are out of scope,
other than there is an expressed preference that forcing the
other side to do an LLP retransmit is NOT preferred and
should be used minimally and only as a last resort. If
the LLP wants to volunteer to hold the buffers for awhile,
that is just fine with DDP.

The LLP merely *wants* to minimize its own buffering.
But it MAY choose to provide that service. DDP, by
contrast, does not buffer.

So whatever delay  mechanism is used, LLP buffering
or LLP retransmission, there is simply no ULP visible
event to be placed on the CQE until placement *and*
delivery is successful or the LLP insists that DDP
accept placement *now* and DDP is forced to terminate
the stream for lack of a buffer.