[openib-general] Re: [PATCH][iWARP] Another CM Verbs Approach

[Caitlin Bestler]

I believe it would be safe to state that all hardware has one
or more methods for transitioning an established TCP connection
to RDMA mode.

What will vary is what restrictions there are, and how easily
that can be expanded.

When the RNIC also provides TCP offload services the current
firmware may only support transition from offloaded connections.
However, accepting an upload of the TCP state from the driver
is probably within the capability of most hardware. For one 
thing it is needed for failover and is quite useful for testing.

So right now it may be common for existing firmware/drivers to
only know how to extract the TCP state from a partner stack.

There is no reason why the Linux stack could not be a partner
stack, but kind of by definition no device vendor can just
decide to make the linux stack a partner. The linux stack
has some say in the matter. Nobody is going to eagerly write
firmware that relies on extracting state information based on
unpublished data structures that are subject to change without
notice. Inter-module dependencies need to be planned, not just
assumed by add-on software.
 

> -----Original Message-----
> From: openib-general-bounces at openib.org 
> [mailto:openib-general-bounces at openib.org] On Behalf Of Roland Dreier
> Sent: Friday, August 26, 2005 11:03 AM
> To: Tom Tucker
> Cc: openib-general at openib.org
> Subject: [openib-general] Re: [PATCH][iWARP] Another CM Verbs Approach
> 
>     Roland> I'd like to understand that last bit above about devices
>     Roland> that establish connections internally.  Does that mean
>     Roland> that there are RNICs that cannot support the iWARP verbs'
>     Roland> notion of migrating a connection?
> 
>     Tom> Yes, that's what it means and the reason is that Linux does
>     Tom> not (yet) provide a means to do so.
> 
> Now I'm confused.  Is the problem that the Linux network 
> stack does not have an interface for transferring TCP state, 
> or is it that some RNIC hardware does not support taking 
> control of an existing TCP connection?  The first problem we 
> can fix, but the second one we can't.
> 
>  - R.
> _______________________________________________
> openib-general mailing list
> openib-general at openib.org
> http://openib.org/mailman/listinfo/openib-general
> 
> To unsubscribe, please visit 
> http://openib.org/mailman/listinfo/openib-general
> 
>