<!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
  <meta http-equiv="Content-Type" content="text/html;charset=ISO-8859-1">
  <title></title>
</head>
<body>
<br>
<br>
'Christoph Hellwig' wrote:<br>
<blockquote type="cite" cite="mid20050802125727.GA2472@lst.de">
  <blockquote type="cite">
    <pre wrap="">Can you provide more details on exactly why you think this is a horrible
idea?  I agree it will be complex, but it _could_ be scoped such that the
complexity is reduced.  For instance, the "offload" function could fail
(with EBUSY or something) if there is _any_ data pending on the socket.
Thus removing any requirement to pass down pending unacked outgoing data, or
pending data that has been received but not yet "read" by the application.
The idea here is that the applications at the top "know" they are going into
RDMA mode and have effectively quiesced the connection before attempting to
move the connection into RDMA mode.  We could, in fact, _require_ the
connect be quiesced to keep things simpler.  I'm quickly sinking into gory
details, but I want to know if you have other reasons (other than the
complextity) for why this is a bad idea.
    </pre>
  </blockquote>
  <pre wrap=""><!---->
I think your writeup here is more than explanation enough.  The offload
can only work for few special cases, and even for those it's rather
complicated, especially if you take things as ipsec or complex tunneling
that get more and more common into account.  </pre>
</blockquote>
I think Steve's point was that it *can* be simplified as necessary to meet 
the demands/needs of the Linux community. It is certainly technically possible,
but agreeably complicated to offload an active socket.<br>
    <br>
 
<blockquote type="cite" cite="mid20050802125727.GA2472@lst.de">
  <pre wrap="">What do you archive by
implementing the offload except trying to make it look more integrated
to the user than it actually is?  Just offload rmda protocols to the
RDMA hardware and keep the IP stack out of that complexity.</pre>
</blockquote>
You get the benefit of things like SYN flood DOS attack avoidance built into
the host stack without replicating this functionality in the offloaded adapter.
There are other benefits of integration like failover, etc... IMHO, however,
the bulk of the benefits are for ULP offload like RDMA where the remote peer
may not be capable of HW RDMA acceleration. This kind of thing could be determined
in "streaming mode" using the host stack and then migrated to an adapter
for HW acceleration only if the remote peer is capable.<br>
<br>
<blockquote type="cite" cite="mid20050802125727.GA2472@lst.de">
  <pre wrap="">
_______________________________________________
openib-general mailing list
<a class="moz-txt-link-abbreviated" href="mailto:openib-general@openib.org">openib-general@openib.org</a>
<a class="moz-txt-link-freetext" href="http://openib.org/mailman/listinfo/openib-general">http://openib.org/mailman/listinfo/openib-general</a>

To unsubscribe, please visit <a class="moz-txt-link-freetext" href="http://openib.org/mailman/listinfo/openib-general">http://openib.org/mailman/listinfo/openib-general</a>
  </pre>
</blockquote>
<br>
</body>
</html>