On 06/29/2011 11:48 AM, Thomas Graf wrote:

On Wed, Jun 29, 2011 at 10:58:41AM -0400, Vladislav Yasevich wrote:

quoted

But what you are proposing violates the protocol.  Zero-window probes do
not count against max retransmissions, even when you are in shutdown pending
state.

You'll come out of this one of 2 ways:
  1) receiver wakes up and processes data.  This will allow for graceful termination.
  2) receiver dies.  Since receive window is full, we have data queued, and this will
     trigger an ABORT to be sent to the sender.

If by die you mean kill the process then this is exactly what I do to trigger
the issue. I simulatenously kill both processes. Not sure what you mean by
trigger an ABORT but I don't see that happen. I also don't see the rwnd reopen
after the socket is closed on the receiver side but that's a separate issue.

Right.  The lack of ABORT from the receive of data is a bug.  I was trying to point out
that instead of modified the sender of data to send the ABORT, you modify the receiver
to send the ABORT when it is being closed while having data queued.

quoted

What you patch is doing is taking a perfectly valid scenario and putting a time limit
on it in violation of the spec.

We also violate the spec by not doing so. The spec says that the number of
SHUTDOWN retransmissions has to be limited by Max.Retrans which we also
can't enforce because of the above.

But we don't even get to sending the SHUTDOWN, so from the wire protocol, we
do not violated it.  We have bad behavior in that when both sender and receiver
are dead, the association is hung.

The scenario is closed sockets on both sides, endpoints on both sides gone
already and retransmissions + heartbeat requests forever.

Any alternative suggestion how to fix this?

The receiver, on sctp_close() checks to see if it has queued data and if it does, the
ABORT is triggered.  This is the same behavior as tcp_close().

-vlad